"I started Better.com because I went through the mortgage process with my wife... we were successful finance professionals, and we were like, 'This is so complicated,' like, 'How do I get pre-approved?'" recalls Garg. A decade later, Garg is leading his mortgage company with backing from some ...

How homeownership helps builds long-term wealth

Investing.com -- In a note Thursday, Jefferies upgraded Roku (NASDAQ:ROKU) to Hold, citing stronger-than-expected platform revenue growth, improving ad fill rates, and a clear path to profitability.

Jefferies: 'Better days ahead' for Roku

As a fitness editor and runner, I tested the Clifton 9 and Cloudmonster 2 based on cushioning, fit, stability, durability, and speed to find which is better.

Hoka Vs. On Cloud: Is The Clifton 9 Or Cloudmonster 2 Better?

Subscribe to our daily newsletter to keep in touch with the subjects shaping our future.

Insiders at Tesla Concerned That Musk's Madness Is Damaging Company, Saying It Would Be Better If He Just Resigned

You're reading a free article with opinions that may differ from The Motley Fool's Premium Investing Services. Become a Motley Fool member today to get instant access to our top analyst recommendations, in-depth research, investing resources, and more. Learn More

Better Artificial Intelligence (AI) Buy for 2025: AMD or Nvidia Stock? @themotleyfool #stocks $NVDA $AMD

Astra Security, a security platform with continuous vulnerability scanning and pentests, announced the closing of a growth capital round—led by Emergent Ventures, with participation from the Neon Fund, Better Capital, Blume Ventures, and PointOne Capital.

Astra Security: Funding Raised To Simplify Cybersecurity With Pentesting

We recently published an article titled Jim Cramer Discusses These 10 Stocks & An Outfit Better Than DeepSeek. In this article, we are going to take a look at where NVIDIA Corporation (NASDAQ:NVDA) stands against the other stocks. In his recent appearance on CNBC’s Squawk on the Street, Jim ...

Jim Cramer on NVIDIA Corporation (NVDA)’s Reduced Price Target: ‘The Downgrades Seemed Almost Out Of Sync’

Artificial intelligence (AI) is rapidly transforming our world, promising to revolutionize everything from human interaction and environmental stewardship to space exploration. This technological revolution, driving us toward what's known as Industry 5.0 (the next phase of industrialization, ...

Better Artificial Intelligence Stock: Nvidia vs. ASML

Better

better

Table of Links Abstract and 1. Introduction Abstract and 1. Introduction Related Work


Task, Datasets, Baseline


RQ 1: Leveraging the Neighbourhood at Inference
4.1. Methods
4.2. Experiments


RQ 2: Leveraging the Neighbourhood at Training
5.1. Methods
5.2. Experiments


RQ 3: Cross-Domain Generalizability


Conclusion


Limitations


Ethics Statement


Bibliographical References Related Work Related Work Related Work Task, Datasets, Baseline Task, Datasets, Baseline Task, Datasets, Baseline RQ 1: Leveraging the Neighbourhood at Inference
4.1. Methods
4.2. Experiments RQ 1: Leveraging the Neighbourhood at Inference 4.1. Methods 4.1. Methods 4.2. Experiments 4.2. Experiments RQ 2: Leveraging the Neighbourhood at Training
5.1. Methods
5.2. Experiments RQ 2: Leveraging the Neighbourhood at Training RQ 2: Leveraging the Neighbourhood at Training 5.1. Methods 5.1. Methods 5.2. Experiments 5.2. Experiments RQ 3: Cross-Domain Generalizability RQ 3: Cross-Domain Generalizability RQ 3: Cross-Domain Generalizability Conclusion Conclusion Conclusion Limitations Limitations Limitations Ethics Statement Ethics Statement Ethics Statement Bibliographical References Bibliographical References Bibliographical References 5. RQ 2: Leveraging the Neighbourhood at Training We leverage the knowledge from neighbour instances directly at the training process to improve the performance. We explore three methods: contrastive learning, single prototypical learning, and multi prototypical learning. These techniques draw inspiration from the same principles as their inference-time counterparts but serve as auxiliary loss constraints during training. Their primary aim is to improve the discriminative capability of embeddings by highlighting differences between instances with distinct rhetorical roles and similarities among instances sharing the same label. While the task-specific classification loss focuses on mapping contextualized representations to label outputs with supervision on individual instances, the methods in this section directly operate on embeddings in latent space. They exploit the interplay among instances to establish effective discriminative decision boundaries, serving as a form of regularization. 5.1. Methods 5.1.1. Contrastive Learning 5.1.1. Contrastive Learning Contrastive learning aims to bring an anchor point closer to related samples while pushing it away from unrelated samples in the embedding space. In a supervised setting, samples with the same/different labels are considered related/unrelated with respect to an anchor (Khosla et al., 2020). The loss is calculated as follows: where δ(ci , cj ) denotes 1 if ci and cj have same rhetorical label, else 0, N denotes batch size. Lengthy legal documents limits the number of documents that can be accommodated in a single batch and this raises concerns about having enough positive samples for the minority class instances within a batch for effective contrasting. To overcome this limitation, we utilize a memory bank (Wu et al., 2018), where we progressively reuse encoded representations from previous batches to compute the contrastive loss. In practice, we maintain a fixed-size representation queue for each rhetorical role. As new representations corresponding to specific labels are generated, they are enqueued into the respective queue with their gradients detached. If the queue size for a label exceeds the maximum limit, the oldest element is dequeued. When computing the contrastive loss, we use the same equation 7. However, in addition to the current batch instances, we employ all the representations stored in the memory bank for contrasting purposes, using them as both positives and negatives, based on the anchor point’s label. memory bank To incorporate the concept of context from surrounding sentences into contrastive learning, we introduce a novel discourse-aware contrastive loss. This is based on the idea that sentences in close proximity within a document, sharing the same label, should exhibit a stronger proximity in the embedding space compared to sentences with the same label but positioned farther apart in the document. To implement this concept, we introduce a penalty inversely proportional to the absolute difference in their positions. In particular, we impose a higher penalty on positive sentence pairs that are closer in the document, encouraging them to be closer in the embedding space than pairs originating from greater distances within the document. The discouse-aware loss is as follows: discourse-aware contrastive loss where β represents a penalty that considers positional information. When ci and cj come from different documents, such as cross-document positives/negatives from the memory bank or across the batch, we apply the lowest possible penalty, considering ci as the farthest sentence relative to in-document positives. We incorporate this additional contrastive loss alongside the classification loss during training. 5.1.2. Single Prototypical Learning 5.1.2. Single Prototypical Learning These both views shape the embedding space by aligning prototypes with their corresponding samples, forming distinct clusters of different labels, each centered around a specific prototype vector. 5.1.3. Multi Prototypical Learning 5.1.3. Multi Prototypical Learning Instead of using a single prototype for each label, this approach employs multiple prototypes for each label to capture the diverse variations within the sentences of the same label. To implement this, a set of M prototypes per label is randomly initialized and a diversity loss (Zhang et al., 2022) is integrated to penalize prototypes of the same label if they are too similar to each other. This ensures that prototypes of the same label are distributed across the embedding space, capturing the multifaceted nuances under each label. The Sample Centric View is also modified to ensure that each sample is in close proximity to at least one prototype among all the prototypes of the same class. Authors:
(1) Santosh T.Y.S.S, School of Computation, Information, and Technology; Technical University of Munich, Germany (santosh.tokala@tum.de);
(2) Hassan Sarwat, School of Computation, Information, and Technology; Technical University of Munich, Germany (hassan.sarwat@tum.de);
(3) Ahmed Abdou, School of Computation, Information, and Technology; Technical University of Munich, Germany (ahmed.abdou@tum.de);
(4) Matthias Grabmair, School of Computation, Information, and Technology; Technical University of Munich, Germany (matthias.grabmair@tum.de). Authors: Authors: (1) Santosh T.Y.S.S, School of Computation, Information, and Technology; Technical University of Munich, Germany (santosh.tokala@tum.de); (2) Hassan Sarwat, School of Computation, Information, and Technology; Technical University of Munich, Germany (hassan.sarwat@tum.de); (3) Ahmed Abdou, School of Computation, Information, and Technology; Technical University of Munich, Germany (ahmed.abdou@tum.de); (4) Matthias Grabmair, School of Computation, Information, and Technology; Technical University of Munich, Germany (matthias.grabmair@tum.de). This paper is available on arxiv under CC by 4.0 Deed (Attribution 4.0 International) license. This paper is available on arxiv under CC by 4.0 Deed (Attribution 4.0 International) license. available on arxiv available on arxiv

How AI Can Better Categorize Legal Documents by Learning from Similar Texts

Authors:
(1) Santosh T.Y.S.S, School of Computation, Information, and Technology; Technical University of Munich, Germany (santosh.tokala@tum.de);
(2) Hassan Sarwat, School of Computation, Information, and Technology; Technical University of Munich, Germany (hassan.sarwat@tum.de);
(3) Ahmed Abdou, School of Computation, Information, and Technology; Technical University of Munich, Germany (ahmed.abdou@tum.de);
(4) Matthias Grabmair, School of Computation, Information, and Technology; Technical University of Munich, Germany (matthias.grabmair@tum.de). Authors: Authors: (1) Santosh T.Y.S.S, School of Computation, Information, and Technology; Technical University of Munich, Germany (santosh.tokala@tum.de); (2) Hassan Sarwat, School of Computation, Information, and Technology; Technical University of Munich, Germany (hassan.sarwat@tum.de); (3) Ahmed Abdou, School of Computation, Information, and Technology; Technical University of Munich, Germany (ahmed.abdou@tum.de); (4) Matthias Grabmair, School of Computation, Information, and Technology; Technical University of Munich, Germany (matthias.grabmair@tum.de). Table of Links Abstract and 1. Introduction Abstract and 1. Introduction Related Work


Task, Datasets, Baseline


RQ 1: Leveraging the Neighbourhood at Inference
4.1. Methods
4.2. Experiments


RQ 2: Leveraging the Neighbourhood at Training
5.1. Methods
5.2. Experiments


RQ 3: Cross-Domain Generalizability


Conclusion


Limitations


Ethics Statement


Bibliographical References Related Work Related Work Related Work Task, Datasets, Baseline Task, Datasets, Baseline Task, Datasets, Baseline RQ 1: Leveraging the Neighbourhood at Inference
4.1. Methods
4.2. Experiments RQ 1: Leveraging the Neighbourhood at Inference 4.1. Methods 4.1. Methods 4.2. Experiments 4.2. Experiments RQ 2: Leveraging the Neighbourhood at Training
5.1. Methods
5.2. Experiments RQ 2: Leveraging the Neighbourhood at Training RQ 2: Leveraging the Neighbourhood at Training 5.1. Methods 5.1. Methods 5.2. Experiments 5.2. Experiments RQ 3: Cross-Domain Generalizability RQ 3: Cross-Domain Generalizability RQ 3: Cross-Domain Generalizability Conclusion Conclusion Conclusion Limitations Limitations Limitations Ethics Statement Ethics Statement Ethics Statement Bibliographical References Bibliographical References Bibliographical References Abstract Rhetorical Role Labeling (RRL) of legal judgments is essential for various tasks, such as case summarization, semantic search and argument mining. However, it presents challenges such as inferring sentence roles from context, interrelated roles, limited annotated data, and label imbalance. This study introduces novel techniques to enhance RRL performance by leveraging knowledge from semantically similar instances (neighbours). We explore inference-based and training-based approaches, achieving remarkable improvements in challenging macro-F1 scores. For inference-based methods, we explore interpolation techniques that bolster label predictions without re-training. While in training-based methods, we integrate prototypical learning with our novel discourse-aware contrastive method that work directly on embedding spaces. Additionally, we assess the cross-domain applicability of our methods, demonstrating their effectiveness in transferring knowledge across diverse legal domains. 1. Introduction In an era of rapid digitalization and exponential growth of legal case volumes, the demand for automated systems to assist legal professionals in tasks like extracting key case elements, summarizing cases, and retrieving relevant cases has surged (Zhong et al., 2020). At the core of these tasks lies Rhetorical Role Labeling (RRL), which involves assigning functional roles to the sentences in the document such as preamble, factual content, evidence, reasoning, etc. Legal documents, characterized by their extensive length, lengthy sentences with unusual word order, frequent cross-references, extensive citation usage, and intricate lexicon, often feature uncommon expressions from everyday language and borrowed terms from various languages to the extent that they are referred to as a sub-language of legalese (Chalkidis et al., 2022; Haigh, 2013). The task of RRL faces several distinctive challenges. Firstly, contextual dependencies, influenced by surrounding sentences and the case’s context, are pivotal in discerning rhetorical role of each sentence, distinguishing RRL as a sequential sentence classification task. Secondly, the intertwining nature of rhetorical roles further complicates the task. For instance, the rationale behind a judgment (Ratio of the decision) often overlaps with Precedents and Statutes, necessitating a nuanced understanding of these roles’ intricate distinctions (Bhattacharya et al., 2021). Thirdly, obtaining extensive annotated data for specialized domains like law is expensive, requiring expert annotators. Lastly, certain rhetorical roles are disproportionately represented in the dataset, leading to significant class imbalance (Malik et al., 2022; Bhattacharya et al., 2021). Traditional up/down sampling methods struggle to fully address this challenge due to the task’s nature, which involves sequences of sentences at the document level. Initially RRL task is formulated as sentence classification, treating each sentence in isolation (Ahmad et al., 2020; Walker et al., 2019). Researchers later adopted it as sequential sentence classification, addressing contextual dependencies between sentences (Bhattacharya et al., 2021; Ghosh and Wyner, 2019; Malik et al., 2022; Kalamkar et al., 2022). They introduced a two-level hierarchical model, encoding sentences independently at the lower level and contextualizing them with neighbouring sentences at the higher-level. While this approach effectively addressed the first challenge of RRL, other challenges remain unaddressed. Recently, Santosh et al. 2023 aimed to address data scarcity through data augmentation, but methods like word deletion, sentence swapping and backtranslation could introduce noise and disrupt coherence. However, this approach did not effectively address label imbalance and intricate role intertwining. In this work, we hypothesize that harnessing knowledge from semantically and contextually similar instances can provide valuable insights to grasp a broader context and reveal underlying rare patterns. This can enhance the understanding of complex label-semantics relationships, improve nuanced label assignments and equip the model to handle less common labels, thus addressing the distinctive challenges of RRL. We explore two approaches for harnessing this knowledge: one directly at inference time without additional parameters or re-training (Sec. 4), and the other during training by incorporating auxiliary loss constraints (Sec. 5). In the inference-based approaches, we interpolate the label distribution predicted by a model with the distribution derived from analogous instances in the training dataset, employing nearest neighbor-based, single, and multiple prototype-based methodologies. These methods enhance performance, particularly on more challenging macro-F1 scores, without requiring retraining. For training-based approaches, we integrate contrastive and prototypical learning which operate directly on the embedding space, leveraging neighborhood relationships. Additionally, we introduce a novel discourse-aware contrastive loss to address the contextual nature of the task. Our experimental results on four datasets from the Indian Jurisdiction validate our proposed methods. While it is common to develop models for specific courts or domains due to unique vocabulary, complex linguistic structures and specific writing styles, such specialization can hinder the adaptability of these models beyond their original context. In rhetorical role labeling, models might memorize context-specific vocabulary rather than understanding the underlying semantics, making cross-domain applications challenging (Savelka et al., 2021). In such cases, developing a model for a new context typically requires annotating a new dataset, which can be expensive. In our work, we assess the cross-domain generalizability of our methods and observe that they enhance model’s ability to transfer across different legal domains compared to a baseline model lacking these auxiliary techniques (Sec. 6). This paper is available on arxiv under CC by 4.0 Deed (Attribution 4.0 International) license. This paper is available on arxiv under CC by 4.0 Deed (Attribution 4.0 International) license. available on arxiv available on arxiv

New AI Methods Help Machines Understand Legal Text Better

Turning AI Into Better Thinkers With Pointer-Based Memory

Authors:
(1) Hung Le, Applied AI Institute, Deakin University, Geelong, Australia;
(2) Dung Nguyen, Applied AI Institute, Deakin University, Geelong, Australia;
(3) Kien Do, Applied AI Institute, Deakin University, Geelong, Australia;
(4) Svetha Venkatesh, Applied AI Institute, Deakin University, Geelong, Australia;
(5) Truyen Tran, Applied AI Institute, Deakin University, Geelong, Australia. Table of Links Abstract & Introduction Methods Methods Part 2 Experimental Results Experimental Results Part 2 Related Works, Discussion, & References Appendix A, B, & C Appendix D 3.2 Dyck Language Recognition 3.3 Compositional Learning SCAN In this task , one needs to map an input sentence into an output sequence of commands [Lake and Baroni, 2018]. The sequences are compositional, consisting of reusable parts. For example, in one case, “jump twice” should be mapped to “JUMP JUMP” and in another, “walk twice” becomes “WALK WALK”. We focus on the “length split” datasets where the training sequences are shorter than the test ones with 11 length modes L = 22, 24, .., 40 [Newman et al., 2020]. We adopt the benchmark, training procedure and baselines prepared by Csord´as et al. [2021], which achieves strong results under standard s2s learning. Here, our aim is not to break SOTA, which can be achieve by hybrid-symbolic architectures [Chen et al., 2020, Shaw et al., 2021]. Instead, we focus on improving Transformer generalization in this task, hence the baselines are chosen as several variants of Transformers (TRM) targeted to sequence extrapolation, including those using Relative Positional Encoding (RPE [Dai et al., 2019]) and Universal Transformer (U. TRM [Dehghani et al., 2018]), which is an advanced Transformer variant that recurrently processes each token, and can dynamically adjust the number of processing steps. Following Csord´as et al. [2021], each baseline is trained 5 times for 50K steps and the resulting model after training is used for evaluation (no validation). Here, we use Transformer as the Encoder, which is the same as the TRM, and stack the Controller to another Transform decoder (see details in Appendix D.3). Hence, the only difference is the decoding where PANM leverages pointer manipulation. Table 2 shows that PANM outperforms other baselines in the hardest settings when the training length is up-to 22, 24, and 25. For 22 and 24 cases, general models like PANM cannot show perfect generalization because some testing compositions is entirely discarded from the train set. In easier settings, PANM shares the perfect median accuracy with the sophisticated U. TRM + RPE although it does not use RPE. Remarkably, despite sharing the same encoder, TRM performs much worse than PANM and even fails to learn in easy modes (33, 36, 40), indicating the importance of pointer handling in this testbed. One problem for other baselines is the EOS decision (when to generate ending token), which requires length tracking [Newman et al., 2020]. As they do not have content free sequence iteration mechanisms, it is extremely hard to trace the length without overfitting to the training data. On the other hand, PANM can hypothetically generate pointer incrementally and capture the difference between the last and the first pointers, i.e. the input length, and infer the output sequence length based on that information. Mathematical Problems We test our model on mathematics [Saxton et al., 2018] where the input/output are questions and answers about math and each token is a character. For example, What is − 5 − 110911? → −110916 (add or sub) and What is the hundreds digit of 31253? → 2 (place value). The task requires not only math reasoning, but also natural language understanding. We follow the training from Csord´as et al. [2021] to conduct experiments on 2 subsets: add or sub (a.s) and place value (p.v), and compare our method with Transformer-based baselines. Here, we focus on the extrapolating test set involving larger numbers, more numbers, more compositions, and thus longer input sequences than the training. We use TRM + RPE as the Encoder and the Controller is added to a normal TRM decoder. As shown in Table 2, on place value, PANM does not suffer from performance crash as TRM + RPE (0% test accuracy, as admitted in the paper [Csord´as et al., 2021] even though it uses the same encoder). PANM achieves similar results as U. TRM+ RPE on add or sub while outperforming it by 11% on place value. We also examine PANM with the original Transformer and report additional results in Appendix D.3. 3.4 Other NLP Tasks Question Answering Our objective is to explore the PANM’s generalization beyond obviously compositional data by applying it in a more practical setting of question answering. For this purpose, we utilize two datasets, namely bAbI [Weston et al., 2015] and SQUAD 1.1 [Rajpurkar et al., 2016] where the input sequence is a context paragraph and a question, and the output is the answer. To add complexity to the task, we ensure the length of test sequence is greater than that of the training by sorting the context paragraph by length and splitting the sorted data into 0.8/0.2 and 0.5/0.5 ratio. Details of the data/task are in Appendix D.4. In bAbI, we configure the PANM similarly to the one described in § 3.3 using Transformer backbone, and test the models after 100-epoch training. The models predict the answer tokens given the context and question tokens. As shown in Table 3 and Appendix Fig. 5 (right), PANM helps Transformer generalize better, consistently improving around 6% and 5% using 0.8/0.2 and 0.5/0.5 splits, respectively. Notably, PANM’s testing loss is not diverged quickly as Transformer’s, indicating PANM’s capability of reducing overfitting. In SQUAD, we use BERT as the backbone to predict the start and the end of the answer as in Kenton and Toutanova [2019]. PANM-assisted model outperforms the baselines by 1% and 2% exact match accuracy, respectively (Table 4). The improvement is significant as BERT is a big foundation model already pretrained with big data and robust against novel test data. Machine Translation Here, we want to verify the PANM in machine translation and show that PANM can work with different number layers of Transformer. The results are presented in Fig. 3 (b) where we report the model perplexity on Multi30K (en-de) dataset. The 30K-sample dataset is sorted by input length and split into training and testing s.t. testing sequences are longer, similar to QA task. The results demonstrate PANM can consistently improve the generalization performance of Transformer across different split ratios and the number of encoder/decoder layers. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Hung Le, Applied AI Institute, Deakin University, Geelong, Australia; (2) Dung Nguyen, Applied AI Institute, Deakin University, Geelong, Australia; (3) Kien Do, Applied AI Institute, Deakin University, Geelong, Australia; (4) Svetha Venkatesh, Applied AI Institute, Deakin University, Geelong, Australia; (5) Truyen Tran, Applied AI Institute, Deakin University, Geelong, Australia. Authors: Authors: (1) Hung Le, Applied AI Institute, Deakin University, Geelong, Australia; (2) Dung Nguyen, Applied AI Institute, Deakin University, Geelong, Australia; (3) Kien Do, Applied AI Institute, Deakin University, Geelong, Australia; (4) Svetha Venkatesh, Applied AI Institute, Deakin University, Geelong, Australia; (5) Truyen Tran, Applied AI Institute, Deakin University, Geelong, Australia. Table of Links Abstract & Introduction Abstract & Introduction Methods Methods Methods Part 2 Methods Part 2 Experimental Results Experimental Results Experimental Results Part 2 Experimental Results Part 2 Related Works, Discussion, & References Related Works, Discussion, & References Appendix A, B, & C Appendix A, B, & C Appendix D Appendix D 3.2 Dyck Language Recognition 3.3 Compositional Learning SCAN In this task , one needs to map an input sentence into an output sequence of commands [Lake and Baroni, 2018]. The sequences are compositional, consisting of reusable parts. For example, in one case, “jump twice” should be mapped to “JUMP JUMP” and in another, “walk twice” becomes “WALK WALK”. We focus on the “length split” datasets where the training sequences are shorter than the test ones with 11 length modes L = 22, 24, .., 40 [Newman et al., 2020]. We adopt the benchmark, training procedure and baselines prepared by Csord´as et al. [2021], which achieves strong results under standard s2s learning. Here, our aim is not to break SOTA, which can be achieve by hybrid-symbolic architectures [Chen et al., 2020, Shaw et al., 2021]. Instead, we focus on improving Transformer generalization in this task, hence the baselines are chosen as several variants of Transformers (TRM) targeted to sequence extrapolation, including those using Relative Positional Encoding (RPE [Dai et al., 2019]) and Universal Transformer (U. TRM [Dehghani et al., 2018]), which is an advanced Transformer variant that recurrently processes each token, and can dynamically adjust the number of processing steps. Following Csord´as et al. [2021], each baseline is trained 5 times for 50K steps and the resulting model after training is used for evaluation (no validation). Here, we use Transformer as the Encoder, which is the same as the TRM, and stack the Controller to another Transform decoder (see details in Appendix D.3). Hence, the only difference is the decoding where PANM leverages pointer manipulation. SCAN Table 2 shows that PANM outperforms other baselines in the hardest settings when the training length is up-to 22, 24, and 25. For 22 and 24 cases, general models like PANM cannot show perfect generalization because some testing compositions is entirely discarded from the train set. In easier settings, PANM shares the perfect median accuracy with the sophisticated U. TRM + RPE although it does not use RPE. Remarkably, despite sharing the same encoder, TRM performs much worse than PANM and even fails to learn in easy modes (33, 36, 40), indicating the importance of pointer handling in this testbed. One problem for other baselines is the EOS decision (when to generate ending token), which requires length tracking [Newman et al., 2020]. As they do not have content free sequence iteration mechanisms, it is extremely hard to trace the length without overfitting to the training data. On the other hand, PANM can hypothetically generate pointer incrementally and capture the difference between the last and the first pointers, i.e. the input length, and infer the output sequence length based on that information. Mathematical Problems We test our model on mathematics [Saxton et al., 2018] where the input/output are questions and answers about math and each token is a character. For example, What is − 5 − 110911? → −110916 (add or sub) and What is the hundreds digit of 31253? → 2 (place value). The task requires not only math reasoning, but also natural language understanding. We follow the training from Csord´as et al. [2021] to conduct experiments on 2 subsets: add or sub (a.s) and place value (p.v), and compare our method with Transformer-based baselines. Here, we focus on the extrapolating test set involving larger numbers, more numbers, more compositions, and thus longer input sequences than the training. We use TRM + RPE as the Encoder and the Controller is added to a normal TRM decoder. As shown in Table 2, on place value, PANM does not suffer from performance crash as TRM + RPE (0% test accuracy, as admitted in the paper [Csord´as et al., 2021] even though it uses the same encoder). PANM achieves similar results as U. TRM+ RPE on add or sub while outperforming it by 11% on place value. We also examine PANM with the original Transformer and report additional results in Appendix D.3. Mathematical Problems 3.4 Other NLP Tasks Question Answering Our objective is to explore the PANM’s generalization beyond obviously compositional data by applying it in a more practical setting of question answering. For this purpose, we utilize two datasets, namely bAbI [Weston et al., 2015] and SQUAD 1.1 [Rajpurkar et al., 2016] where the input sequence is a context paragraph and a question, and the output is the answer. To add complexity to the task, we ensure the length of test sequence is greater than that of the training by sorting the context paragraph by length and splitting the sorted data into 0.8/0.2 and 0.5/0.5 ratio. Details of the data/task are in Appendix D.4. In bAbI , we configure the PANM similarly to the one described in § 3.3 using Transformer backbone, and test the models after 100-epoch training. The models predict the answer tokens given the context and question tokens. As shown in Table 3 and Appendix Fig. 5 (right), PANM helps Transformer generalize better, consistently improving around 6% and 5% using 0.8/0.2 and 0.5/0.5 splits, respectively. Notably, PANM’s testing loss is not diverged quickly as Transformer’s, indicating PANM’s capability of reducing overfitting. In SQUAD , we use BERT as the backbone to predict the start and the end of the answer as in Kenton and Toutanova [2019]. PANM-assisted model outperforms the baselines by 1% and 2% exact match accuracy, respectively (Table 4). The improvement is significant as BERT is a big foundation model already pretrained with big data and robust against novel test data. Question Answering bAbI SQUAD Machine Translation Here, we want to verify the PANM in machine translation and show that PANM can work with different number layers of Transformer. The results are presented in Fig. 3 (b) where we report the model perplexity on Multi30K (en-de) dataset. The 30K-sample dataset is sorted by input length and split into training and testing s.t. testing sequences are longer, similar to QA task. The results demonstrate PANM can consistently improve the generalization performance of Transformer across different split ratios and the number of encoder/decoder layers. Machine Translation This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

Researchers Built a Pointer System That Helps AI Generalize Better

There are a couple of reasons why I started writing. First, it’s so easy to get it published today that I can’t believe that I hit publish and it’s available to read. Sometimes, I write it for myself so that I can come back to it when I want. Great job, team! Second, I don’t want to watch the show from the sidelines. What I mean by that is that in this content creator economy, I feel I am not contributing, and given that I have never done it in the past, it’s really hard to get started now. Everyone around you is creating content. Some people are so good at it that they take your work and repackage it well, which is easily digestible. There is an art to that. I feel there are two parts to the story: 1) You need to have original thoughts and ideas, and 2) you need to know how to present it to the world so that they can digest it and learn from it. I think I do a pretty good job at #1. I have written more than 100 design docs, wikis, etc., which cater to a specific group of people. We have various AWS services based on the ideas presented in those docs to VPs and Directors here at AWS. So, I would say I am decent at coming up with ideas and original content. #2 is where I feel I haven’t done a good job. It could be because of a skill gap or my lack of motivation to make strides here. I honestly don’t know where I lack (maybe this post is coming out that way; I’ll see in a year or so where I land). I call this a known-unknown for me. Known, as I know I lack in this aspect; unknown, as I don’t know why. While I understand it’s a “me” problem, I’m almost certain that a lot more people struggle with this. I have several 1:1s and discussions where people express the desire to become more visible and be able to present their work. My message to them is: Do this, Try, and Fail. It’s okay to fall flat on your face! You may get a few bruises, but next time, you will be ready and might be able to add a cushion to the blow. You don’t have to go all out and start writing in forums and all, but you can take baby steps. Here are some baby steps that you can start with and slowly build your muscle and get better at it. Meeting Notes One of the things I did early in my career was to always volunteer to take notes in meetings. I paired with a few senior engineers and told them I would be happy to take notes for them if they invited me to meetings. Initially, they invited me to the ones where I was needed, but then, due to the convenience of note-taking, they invited me to other meetings if I was free to tag along. I mostly said yes and tagged along. I not only got better at note-taking, but I also learned so much more. I was more informed than most of my other teammates and certainly knew what was coming next or what may disrupt our roadmap/sprint. This was such a great advantage I had. Of course, I had to work extra hours to compensate for this, but it was a no-brainer for me at the time. I was in the mood to drink from the firehose. It was definitely overwhelming at times! Anyways, enough about my little hack, back to note-taking. When I first started, I jotted down bullet points and sent them to the organizer before sending them to a wider audience. I got immediate feedback and improved. By the 5th or 6th meeting, I had built a template that I followed. Here was that sample template that I wrote and got approved with a senior engineer at the time: MOM (Minutes of the meeting)Date: XX/YY/ZZZZAttendees: ABC@, xyz@….Summary: We discussed ABC. There were two options. Option A and Option B. We discussed and decided to go ahead with option B for XY reason. Please refer to XYZ doc for more information.Action Items: 1. Specific Action Item#1. ETA: AA/BB2. Specific Action Item#2. ETA: XX/YY3. …..Next Steps: Finish AIs and then jump on to the implementation. These are just some random text. Trying to highlight what I was doing as SDE 1 fresh out of college. MOM (Minutes of the meeting)Date: XX/YY/ZZZZ MOM Date: Attendees: ABC@, xyz@…. Attendees: Summary: We discussed ABC. There were two options. Option A and Option B. We discussed and decided to go ahead with option B for XY reason. Please refer to XYZ doc for more information. Summary: Action Items: 1. Specific Action Item#1. ETA: AA/BB2. Specific Action Item#2. ETA: XX/YY3. ….. Next Steps: Finish AIs and then jump on to the implementation. These are just some random text. Trying to highlight what I was doing as SDE 1 fresh out of college. As you can see, this was nothing fancy, and I definitely had help from senior engineers at the time. This was a good start, giving some structure to meetings and how to keep making progress on things. It was a good writing exercise for me, as I got instant feedback on my notes! ADRs (Architectural Decision Records) After my success with note-taking, I wanted to take it to the next level where I not only took the notes but also started to capture the actual discussion, as the notes were not sufficient for capturing the nuances in disputed situations. Even though we agreed on something, we kept coming back to why we decided that and again required us to go through the whole design doc. So, one of the seniors suggested doing something about it, and I did some research and found about ADRs and thought it would be a good idea. I experimented with it and used it a couple of times, but it took a lot more time than just note-taking. Also, it needed a lot more context and nuance capturing. I suggested the doc owners own this. Here is the template I used then: Title: For e.g. Selecting java as a language to write libraries for this specific implementation.
Status: (Accepted/Rejected/Proposed)
Record Date
Date when this got finalized.
Context
What is the issue that we’re seeing that is motivating this decision or change?
Decision
What is the change that we’re proposing and/or doing?
Consequences
What becomes easier or more difficult to do because of this change?
Future Action Item Title: For e.g. Selecting java as a language to write libraries for this specific implementation. Title: Status: (Accepted/Rejected/Proposed) Status: Record Date Record Date Date when this got finalized. Context Context What is the issue that we’re seeing that is motivating this decision or change? Decision Decision What is the change that we’re proposing and/or doing? Consequences Consequences What becomes easier or more difficult to do because of this change? Future Action Item Future Action Item We followed it for some time, and then, because of the overhead it added, we decided to discontinue it. Even though it was discontinued, it gave me a couple of good opportunities to exercise my writing muscles. One Pagers As you might have guessed, I was not only just writing these notes but also growing in my career. I was continuously delivering my assigned projects and exceeding expectations, and I got promoted within 2 years’ time. I wrote various design docs and got them reviewed by my colleagues and senior engineers. One thing I always felt was that sometimes you don’t need a full-blown design doc to seek guidance from senior folks on the team. I myself faced this dilemma on what is the right level of detail to have before seeking feedback or ideas. We definitely did a lot of whiteboarding, but we never followed any quick one-pager document style review. To address this, I created a one-pager template for my team (which I evolved over the years and still actively use). Here is the template that I created and wrote several docs on: Background
Tell us about the background. For e.g. I am working on project xyz which will help improve the deployment time from N hours to M hours. As a part of this project, I am working on migration of active prod workload.
What is the problem?While doing the active workload, I have to pause customer operations….
Why is this a problem? As this will be visible to customers in terms of 5xx errors and ….How would you solve this problem? I was thinking of building a staging area…ConclusionTODO: Fill after the meeting. Background Background Tell us about the background. For e.g. I am working on project xyz which will help improve the deployment time from N hours to M hours. As a part of this project, I am working on migration of active prod workload. What is the problem? What is the problem? While doing the active workload, I have to pause customer operations…. Why is this a problem? As this will be visible to customers in terms of 5xx errors and …. How would you solve this problem? I was thinking of building a staging area… Conclusion Conclusion TODO: Fill after the meeting. It not only helped me, but the team also liked it, and our discussions became more frequent and more structured. These are the things that I did earlier in my career. All this work that I did, I never wrote about it. If you’ve read this article till this point, now you may understand why I claimed that I’m not struggling with #1 (having original thoughts and ideas) but #2 (knowing how to present it to the world) is where I have to shift my focus, which is equally important to share the knowledge and make yourself heard. Conclusion Every one should write. Writing is a fantastic way to express your authentic self and share your unique perspective with the world. It doesn’t have to be perfect right from the start — the most important thing is to simply begin. There are opportunities all around you to start writing. Start small, experiment, and don’t be afraid to improve over time. So take that first step, and let your voice be heard! Disclaimer: These are my own views. They have no relation to where I work. Disclaimer: Disclaimer: These are my own views. They have no relation to where I work.

Can Writing Make You Better at Your Job? This Software Developer Thinks So

One of the first things I did upon moving into a new apartment last November was get a better deal on my broadband internet. To discover what other internet service providers (ISPs) offered service to my building, I used a handy resource that the Federal Communications Commission had just released to the public a few days earlier—a new version of the National Broadband Map. It’s an interactive tool that displays which companies offer connections at every location in the country and the maximum speed for those connections. the National Broadband Map When I entered my address into the map, I discovered that Astound, a provider I didn’t know about, could offer me the same speeds I was getting from Spectrum, which was already the internet provider for the apartment. A visit to Astound’s website revealed that the company could also give me a cheaper price, which ended up saving me around $27 a month. Astound website The National Broadband Map has gotten major updates twice each year for more than a decade with info about which companies provide internet service to a given area. However, until last year, the map only had details for a general area—not for individual addresses. This lack of granularity presented a problem because, as an investigation into the digital divide by The Markup showed, the maximum speed offered by just one provider can vary dramatically from block to block, or even house to house. investigation The new version of the map tries to identify every location across the country capable of receiving broadband internet service and what level of internet service, if any, a company can provide at that location. But the map, as it currently exists, is far from perfect. Locations are missing, and the availability of information, as submitted by providers, often doesn’t match the reality on the ground. However, the process of using the map to help you shop around has the upside of the potential to reveal, and allow you to personally help fix, inaccuracies on the map. Members of the public can identify errors and submit corrections themselves. Since the map is a starting point for allocating more than $42 billion in federal broadband funding in America, by identifying inaccuracies on the map, you can help play a small part in potentially guiding money to where it is most urgently needed. allocating Here’s how to search for a better deal on broadband. 1. Find your home on the map Visit the National Broadband Map at https://broadbandmap.fcc.gov and enter your address. Ideally, your correct address should immediately come up as an option with all of the availability data displayed for your perusal. But the map can be, at times, a little temperamental. It may take a bit of resubmitting your address and clicking around to get the data you’re looking for to display. https://broadbandmap.fcc.gov If you can’t get the map to locate your address, skip to step four. 2. Look at the providers claiming to offer you service Are any offering high speeds apart from the one you currently use? (Make sure the “Fixed broadband” button in the upper-right-hand corner is highlighted.) The type of each connection offered is listed in the “Technology” column and, broadly speaking, “Fiber to the Premises” is considered the best type of connection in terms of speed and reliability, followed by cable. Connections listed under copper are usually slow and outdated. Satellite and fixed wireless connections can be extremely useful if there are no other options available, but they are generally considered inferior to fiber and cable. 3. Go to the providers’ websites to see their rates Providers are identified by their corporate name, which is not necessarily the same one they present to the public when offering internet service. Charter Communications’ internet brand is Spectrum. Astound is listed under Radiate Holdings, LP. You may have to do a bit of Google sleuthing to figure out who is who. There’s no direct link on the map, so you need to search on Google (or your search engine of choice) to get the URL. For example, I searched “Radiate Holdings broadband” and found Astound’s website. Once you’re on an ISP’s website, enter your address into the tool that lets you search to see if the company serves your location—most ISPs have this in a pretty obvious place. You might find a faster speed, a lower price, or both. 4. If what’s listed on the ISP’s website does not match what’s on the FCC’s map Hit the map’s “Availability Challenge” button and select the provider on whose site you noticed a contradiction. Decide if you want to file an official challenge, which forces the provider to check for errors, or submit an FYI to the FCC. If you file an official challenge, FCC staff will perform an initial review, and if it passes muster, the internet service provider has 60 days to either concede the challenge and accept your revision or make its own argument to the agency backing up why what it originally submitted to the map was correct. If the provider doesn’t concede the challenge, the company may get in touch with you directly about it and point out where you may have been mistaken. In any case, the ultimate decision about what to do regarding any challenge is left up to the FCC. If you file an official challenge If you’re submitting an FYI, your information will be used to inform the agency’s general efforts to improve the map but won’t kick off a challenge process. If you’re submitting an FYI Fill out your information and, in the “reason for challenge” section, select either the ISP “does not offer the speed(s) reported” or the ISP “does not offer the technology reported” at your location. Describe your experience and, if you can, upload a screenshot to prove what you saw. You may not have gotten a better deal, but you just did your part to improve the FCC’s map! (Note: You have to either own or live at a location in order to submit an availability challenge there.) For more info on submitting an availability challenge, check out the FCC’s page on it here, and for info about what happens after submission, check out the FCC’s page explaining that here. here here 5. If your address does not show up on the map at all, or shows up weirdly or incorrectly For example, people who live in apartment buildings with multiple units might see their building listed as a single-family home. If you see something wonky, or can’t find your location, look around on the map to find the building that most likely matches your home. Then, click the “Location Challenge” button, and then describe the issue. For more info, check out the FCC’s pages on submitting a location challenge and what happens after you submit. submitting a location challenge what happens after Again, no better deal for you, but you’re still helping improve the map! 6. Tell us what you found If you followed the above steps, we’d love to know what happened. Did you find a better deal on your internet or submit a challenge to the FCC? We’re working with Consumer Reports to see what people find and where we should follow up. Use this form to tell us what you found.
We’re also working with Outlier Media’s Detroit Documenters program to see what’s going on in Detroit. We’re working with Consumer Reports to see what people find and where we should follow up. Use this form to tell us what you found. Use this form to tell us what you found We’re also working with Outlier Media’s Detroit Documenters program to see what’s going on in Detroit. Outlier Media Detroit Documenters Interested in following our reporting on internet disparities? Subscribe to The Markup’s newsletter. Subscribe to The Markup’s newsletter Credits: Aaron Sankin Credits: Aaron Sankin Aaron Sankin Also published here Also published here here Photo by Chase Clark on Unsplash Photo by Chase Clark on Unsplash Chase Clark Unsplash

Looking for a Better Deal on Broadband? What You Should Be Doing to Find It

Why AI Needs to Lose (a Little) to Recognize Your Face Better

لماذا يحتاج الذكاء الاصطناعي إلى فقدان (قليلاً) للتعرف على وجهك بشكل أفضل

Warum KI (ein wenig) verlieren muss, um Ihr Gesicht besser zu erkennen

Por qué la IA necesita perder (un poco) para reconocer mejor tu rostro

Ինչու է AI-ն պետք է կորցնի (մի քիչ) ձեր դեմքը ավելի լավ ճանաչելու համար

Perché l'intelligenza artificiale deve perdere (un po') per riconoscere meglio il tuo volto

מדוע בינה מלאכותית צריכה להפסיד (קצת) כדי לזהות את הפנים שלך טוב יותר

AI が顔をより正確に認識するには、なぜ (少し) 負ける必要があるのか

რატომ სჭირდება ხელოვნური ინტელექტის დაკარგვა (ცოტა) თქვენი სახის უკეთ ამოცნობისთვის

Pourquoi AI Esengeli Ebungisa (Mwa) Pona Reconnaître Elongi Na Yo Mieux

Зашто АИ треба да изгуби (мало) да би боље препознао ваше лице

Kwa nini AI Inahitaji Kupoteza (Kidogo) Ili Kutambua Uso Wako Bora

Nima uchun AI sizning yuzingizni yaxshiroq tanib olish uchun (bir oz) yo'qotishi kerak

I hope this will be my final piece on Tonga's connectivity saga. In the wake of my previous analysis critiquing Tonga's submarine cable infrastructure, I present a tangible, forward thinking solution that challenges the status quo of our national connectivity strategy. The Game Changing Alternative: Starlink Community Gateway While governments debate infrastructure and contractors count their potential profits, a revolutionary approach stands ready to transform Tonga's internet landscape. The Starlink Community Gateway isn't just an alternative, it's a comprehensive solution that addresses our connectivity challenges head-on. The Numbers Speak Volumes Let's break down the stark financial reality: Current proposed cable project: 80 million pa'anga
Starlink Community Gateway solution: Approximately 32.1 million pa'anga
Savings: Nearly 60% of the proposed budget
Coverage: All major island groups (Tongatapu, Ha'apai, Vava'u) Technical Superiority The Starlink solution offers capabilities that make our current cable infrastructure look antiquated: Fiber-like speeds up to 10 Gbps (symmetric upload/download)
Over 99% uptime reliability
Flexible distribution technologies including:
Last-mile fiber
Fixed wireless
Mobile wireless connectivity The Geopolitical Elephant in the Room We're acutely aware that New Zealand and Australia will be reluctant to fund this proposed solution, and it isn't purely technical. It could fundamentally be about sovereignty and geopolitical influence. Their preferred cable infrastructure represents more than just connectivity, it sure looks like a strategic justification for a submarine cable investment that has gone redundant because no one else in the Pacific has been utilizing the Hawaiki cable other than American Samoa. Breaking Down the Investment The Starlink Community Gateway requires: Initial Setup: USD$1.25 million per gateway


Monthly Service: USD$75,000 per Gbps


Total 15-Year Investment: USD$13,500,000 = 32.1 million pa'anga Compared to the 80-million pa'anga cable project, this solution: Costs significantly less
Provides more reliable connectivity
Offers greater technological flexibility
Reduces dependency on vulnerable submarine infrastructure Detailed Cost Analysis For a 1 Gbps scenario: Initial setup cost: USD $1,250,000.00
Monthly cost at Year 1: USD$900,000
Monthly cost at Year 15: USD$116,560.09
Total cost after 15 years: USD $13,500,000
Total cost for redundancy across all islands in Tongan Pa'anga (15 years): TOP $32,104,620 This is still way cheaper than the $80 million submarine cable option and solves all our connectivity issues. Resilience in the Face of Environmental Challenges Tonga's geographical reality i.e. a seismically active maritime environment that demands a resilient solution. Satellite technology eliminates the risks associated with submarine cables: No physical infrastructure vulnerable to earthquakes and ships dragging their anchors
Instant rerouting capabilities
Rapid deployment and maintenance A Message to Decision Makers To our ministers and policy makers: This isn't just about technology. It's about investing in our people's future. A 60% more affordable, technologically superior solution exists. The question is whether we have the courage to embrace it, but unfortunately we do not having the funding to pursue it and our development partners do not see any return on it for their investment. The People's Technology Starlink represents more than an internet solution, it's a lifeline for our communities. It symbolizes our ability to leapfrog traditional infrastructure limitations and position ourselves at the forefront of technological innovation. The Innovation Dividend: Flexibility vs. Lock-In The most profound advantage of the Starlink Community Gateway is its inherent flexibility, a stark contrast to the rigid, 15-year lock-in of the proposed submarine cable project. This isn't just about technology, it's about future proofing our national connectivity strategy. The Exponential Innovation Curve With the submarine cable project, Tonga is committing to a static technological solution that will become obsolete before the contract expires. In contrast, the Starlink approach offers: Pay-as-You-Go Model: Unlike the fixed infrastructure investment, Starlink allows for: Incremental upgrades


Flexible scaling


Immediate adaptation to technological advances Cost Trajectory: The economics of satellite technology follow a dramatically different path: Initial investment remains stable
Technology costs decrease logarithmically
Performance continues to improve exponentially A Glimpse into the Future Consider the technological landscape over 15 years: 2025: Current Starlink technology at $75,000 per Gbps
2030: Anticipated 40-50% reduction in per Gbps costs
2035: Potential for 70-80% cost decrease with emerging technologies
2040: Unimaginable improvements in satellite connectivity The submarine cable represents a snapshot in time i.e. frozen for 15 years. The Starlink Community Gateway is a living, breathing solution that evolves with technological innovation. Conclusion: A Bold Path Forward Tonga stands at a critical junction. We can continue down the path of expensive, vulnerable infrastructure, or we can choose a bold, innovative approach that truly serves our people. The Starlink Community Gateway isn't just an alternative. It's our gateway to a more connected, resilient future. Having said all of that, there is one profound truth when it comes to telecommunications: Nothing comes free in this world, there will always be costs associated with it. I hope this will be my final piece on Tonga's connectivity saga. In the wake of my previous analysis critiquing Tonga's submarine cable infrastructure, I present a tangible, forward thinking solution that challenges the status quo of our national connectivity strategy. previous The Game Changing Alternative: Starlink Community Gateway While governments debate infrastructure and contractors count their potential profits, a revolutionary approach stands ready to transform Tonga's internet landscape. The Starlink Community Gateway isn't just an alternative, it's a comprehensive solution that addresses our connectivity challenges head-on. The Numbers Speak Volumes Let's break down the stark financial reality: Current proposed cable project: 80 million pa'anga Starlink Community Gateway solution: Approximately 32.1 million pa'anga Savings: Nearly 60% of the proposed budget Coverage: All major island groups (Tongatapu, Ha'apai, Vava'u) Current proposed cable project: 80 million pa'anga Starlink Community Gateway solution: Approximately 32.1 million pa'anga Savings: Nearly 60% of the proposed budget Coverage: All major island groups (Tongatapu, Ha'apai, Vava'u) Technical Superiority The Starlink solution offers capabilities that make our current cable infrastructure look antiquated: Fiber-like speeds up to 10 Gbps (symmetric upload/download) Over 99% uptime reliability Flexible distribution technologies including: Last-mile fiber Fixed wireless Mobile wireless connectivity Fiber-like speeds up to 10 Gbps (symmetric upload/download) Over 99% uptime reliability Flexible distribution technologies including: Last-mile fiber Fixed wireless Mobile wireless connectivity The Geopolitical Elephant in the Room We're acutely aware that New Zealand and Australia will be reluctant to fund this proposed solution, and it isn't purely technical. It could fundamentally be about sovereignty and geopolitical influence. Their preferred cable infrastructure represents more than just connectivity, it sure looks like a strategic justification for a submarine cable investment that has gone redundant because no one else in the Pacific has been utilizing the Hawaiki cable other than American Samoa. Breaking Down the Investment The Starlink Community Gateway requires: Initial Setup: USD$1.25 million per gateway Monthly Service: USD$75,000 per Gbps Total 15-Year Investment: USD$13,500,000 = 32.1 million pa'anga Initial Setup: USD$1.25 million per gateway Initial Setup: USD$1.25 million per gateway Monthly Service: USD$75,000 per Gbps Monthly Service: USD$75,000 per Gbps Total 15-Year Investment: USD$13,500,000 = 32.1 million pa'anga Total 15-Year Investment: USD$13,500,000 = 32.1 million pa'anga Compared to the 80-million pa'anga cable project, this solution: Costs significantly less Provides more reliable connectivity Offers greater technological flexibility Reduces dependency on vulnerable submarine infrastructure Costs significantly less Provides more reliable connectivity Offers greater technological flexibility Reduces dependency on vulnerable submarine infrastructure Detailed Cost Analysis For a 1 Gbps scenario: Initial setup cost: USD $1,250,000.00 Monthly cost at Year 1: USD$900,000 Monthly cost at Year 15: USD$116,560.09 Total cost after 15 years: USD $13,500,000 Total cost for redundancy across all islands in Tongan Pa'anga (15 years): TOP $32,104,620 Initial setup cost: USD $1,250,000.00 Monthly cost at Year 1: USD$900,000 Monthly cost at Year 15: USD$116,560.09 Total cost after 15 years: USD $13,500,000 Total cost for redundancy across all islands in Tongan Pa'anga (15 years): TOP $32,104,620 This is still way cheaper than the $80 million submarine cable option and solves all our connectivity issues. Resilience in the Face of Environmental Challenges Tonga's geographical reality i.e. a seismically active maritime environment that demands a resilient solution. Satellite technology eliminates the risks associated with submarine cables: No physical infrastructure vulnerable to earthquakes and ships dragging their anchors Instant rerouting capabilities Rapid deployment and maintenance No physical infrastructure vulnerable to earthquakes and ships dragging their anchors Instant rerouting capabilities Rapid deployment and maintenance A Message to Decision Makers To our ministers and policy makers: This isn't just about technology. It's about investing in our people's future. A 60% more affordable, technologically superior solution exists. The question is whether we have the courage to embrace it, but unfortunately we do not having the funding to pursue it and our development partners do not see any return on it for their investment. The People's Technology Starlink represents more than an internet solution, it's a lifeline for our communities. It symbolizes our ability to leapfrog traditional infrastructure limitations and position ourselves at the forefront of technological innovation. The Innovation Dividend: Flexibility vs. Lock-In The most profound advantage of the Starlink Community Gateway is its inherent flexibility, a stark contrast to the rigid, 15-year lock-in of the proposed submarine cable project. This isn't just about technology, it's about future proofing our national connectivity strategy. The Exponential Innovation Curve With the submarine cable project, Tonga is committing to a static technological solution that will become obsolete before the contract expires. In contrast, the Starlink approach offers: Pay-as-You-Go Model : Unlike the fixed infrastructure investment, Starlink allows for: Pay-as-You-Go Model Incremental upgrades Flexible scaling Immediate adaptation to technological advances Incremental upgrades Incremental upgrades Flexible scaling Flexible scaling Immediate adaptation to technological advances Immediate adaptation to technological advances Cost Trajectory : The economics of satellite technology follow a dramatically different path: Cost Trajectory Initial investment remains stable Technology costs decrease logarithmically Performance continues to improve exponentially Initial investment remains stable Technology costs decrease logarithmically Performance continues to improve exponentially A Glimpse into the Future Consider the technological landscape over 15 years: 2025: Current Starlink technology at $75,000 per Gbps 2030: Anticipated 40-50% reduction in per Gbps costs 2035: Potential for 70-80% cost decrease with emerging technologies 2040: Unimaginable improvements in satellite connectivity 2025: Current Starlink technology at $75,000 per Gbps 2030: Anticipated 40-50% reduction in per Gbps costs 2035: Potential for 70-80% cost decrease with emerging technologies 2040: Unimaginable improvements in satellite connectivity The submarine cable represents a snapshot in time i.e. frozen for 15 years. The Starlink Community Gateway is a living, breathing solution that evolves with technological innovation. Conclusion: A Bold Path Forward Tonga stands at a critical junction. We can continue down the path of expensive, vulnerable infrastructure, or we can choose a bold, innovative approach that truly serves our people. The Starlink Community Gateway isn't just an alternative. It's our gateway to a more connected, resilient future. Having said all of that, there is one profound truth when it comes to telecommunications: Nothing comes free in this world, there will always be costs associated with it.

Elon Musk's Starlink May Be The Better Option for Internet in the Pacific Islands

Table of links ABSTRACT 1 INTRODUCTION 2 BACKGROUND: OMNIDIRECTIONAL 3D OBJECT DETECTION 3 PRELIMINARY EXPERIMENT 3.1 Experiment Setup 3.2 Observations 3.3 Summary and Challenges 4 OVERVIEW OF PANOPTICUS 5 MULTI-BRANCH OMNIDIRECTIONAL 3D OBJECT DETECTION 5.1 Model Design 6 SPATIAL-ADAPTIVE EXECUTION 6.1 Performance Prediction 5.2 Model Adaptation 6.2 Execution Scheduling 7 IMPLEMENTATION 8 EVALUATION 8.1 Testbed and Dataset 8.2 Experiment Setup 8.3 Performance 8.4 Robustness 8.5 Component Analysis 8.6 Overhead 9 RELATED WORK 10 DISCUSSION AND FUTURE WORK 11 CONCLUSION AND REFERENCES 5 MULTI-BRANCH OMNIDIRECTIONAL 3D OBJECT DETECTION We describe a multi-branch design of our omnidirectional 3D detection model. We also explain how the model is modified to fit the given device and latency target. 5.1 Model De-sign The goal of our multi-branch model is to allow each region in a surrounding space can be processed through various capabilities of the BEV-based 3D detectors detailed in Table 1. Instead of simply using all types of models loaded on the memory, we modularized the models and combined the core modules, eliminating redundant ones. The model follows the baseline pipeline of the BEV-based 3D detection illustrated in Figure 2. However, as shown in Figure 7, an image can be processed by one of several modules at each inference stage. Each branch within the model represents a unique execution path for an image, comprising a sequence of module decisions at all stages. We detailed the branch design following the order of the baseline pipeline. The initial stage of BEV feature generation is to extract 2D feature maps from all multi-view images. At this stage, each image can be processed through one of four backbone. networks, each of which is differentiated by its capacity and input resolution, as detailed in Table 1. From the extracted 2D feature maps, the model infers the 3D depth information in each camera coordinate. To enable the adjustment of quality and latency of the depth prediction, our model is equipped with two types of DepthNets. One is a simple 1- layer convolutional network trained using only object depth loss, providing sparse supervision signals. The other is a deep convolutional network trained on dense depth maps derived from 3D point clouds. As illustrated in Figure 8, the densely supervised DepthNet generates an accurate and detailed depth map from an image. The depth prediction quality is further increased by incorporating a large-scale backbone network with high-resolution images, providing a clear distinction between foreground objects and the background. BEV feature maps are generated by projecting the extracted 2D feature map into a unified grid space using the predicted depth and camera parameters. Fusing the generated BEV features over time enhances the robustness of perception by providing contiguous observations of objects. As shown in Figure 7, we developed a module that can fuse the BEV features extracted from consecutive images of each camera view 𝑗. Due to the possible movements of the camera (or a robot carrying the camera), directly fusing the BEV features may result in misalignment of spatial features over time, greatly reducing the accuracy gain. Therefore, the module aligns the last frame’s BEV feature map 𝐹 𝑗 𝑡−1 into the current frame’s camera coordinate system. The aligned BEV feature map 𝐹 ′𝑗 𝑡−1→𝑡 is acquired as follows: where T is the operation of transforming the spatial locations of features in the BEV grid using the relative camera pose, i.e., the camera motion 𝜃𝑡−1→𝑡 . To estimate 𝜃𝑡−1→𝑡 , we employ a neural network that uses inertial sensor data and consecutive images. At runtime, Panopticus buffers the previous image 𝐼 𝑗 𝑡−1 and 𝐹 𝑗 𝑡−1 of each camera view 𝑗. Since the prediction results of 𝜃𝑡−1→𝑡 are consistent across camera views, the model predicts it once per frame, allowing the sharing of predicted 𝜃𝑡−1→𝑡 with the alignment operations for other views. After aggregating 𝐹 𝑗 𝑡 and the aligned 𝐹 ′𝑗 𝑡−1→𝑡 across all camera views, the model concatenates the integrated BEV feature maps, which are then fed into the BEV head. The resulting 3D bounding boxes are post-processed using nonmax suppression (NMS) to remove duplicates. Panopticus utilizes a 3D object tracker to keep track of the latest states of detected objects. We employed a 3D Kalman filter that can efficiently forecast and update the status of tracked objects. Our model features a branch that outputs the predicted states of tracked objects, skipping new detection for the target camera view. The future state x ′ 𝑡 of a tracked object at the incoming time 𝑡 is predicted as follows: where 𝐴𝑡−1 and x𝑡−1 are the state transition model and the object state vector at time 𝑡−1, respectively. x𝑡−1 is parameterized by the object’s 3D location (𝑥, 𝑦, 𝑧), velocity (𝑣𝑥, 𝑣𝑦, 𝑣𝑧 ) and size (𝑤, ℎ,𝑙). Future state estimation of the target object using 𝐴𝑡−1 is simply applying the velocity predicted by the BEV head to its location, which is processed instantaneously. Overall, the image from each camera view can be processed by one of 17 branches: 16 detection branches and 1 lightweight tracker’s branch utilizing objects’ predicted states. 5.2 Model Adaptation Panopticus supports offline model adaptation to meet the memory and latency constraints on a given device. To achieve this, the memory consumption and processing latency of all modules in the model are profiled on the target device. First, the model is adjusted to adhere to the memory constraints. For example, for the Jetson Orin Nano [8] with 6.3 GB of limited memory, large-sized modules such as the R152 backbone are detached to prevent overloading the memory capacity. Second, some inference branches require significant computation; in fact, inferring a single image may surpass the target latency. To meet latency requirements, branches with latency profiles that exceed the target limit are also removed. The model, with its number of branches 𝑀 changed, is then deployed on the target device. Authors:

Jeho Lee
Chanyoung Jung
Jiwon Kim
Hojung Cha This paper is available on arxiv under CC by 4.0 Deed (Attribution 4.0 International) license. Table of links Table of links ABSTRACT ABSTRACT ABSTRACT 1 INTRODUCTION 1 INTRODUCTION 1 INTRODUCTION 2 BACKGROUND: OMNIDIRECTIONAL 3D OBJECT DETECTION 2 BACKGROUND: OMNIDIRECTIONAL 3D OBJECT DETECTION 2 BACKGROUND: OMNIDIRECTIONAL 3D OBJECT DETECTION 3 PRELIMINARY EXPERIMENT 3 PRELIMINARY EXPERIMENT 3 PRELIMINARY EXPERIMENT 3.1 Experiment Setup 3.2 Observations 3.3 Summary and Challenges 4 OVERVIEW OF PANOPTICUS 4 OVERVIEW OF PANOPTICUS 4 OVERVIEW OF PANOPTICUS 5 MULTI-BRANCH OMNIDIRECTIONAL 3D OBJECT DETECTION 5 MULTI-BRANCH OMNIDIRECTIONAL 3D OBJECT DETECTION 5 MULTI-BRANCH OMNIDIRECTIONAL 3D OBJECT DETECTION 5.1 Model Design 6 SPATIAL-ADAPTIVE EXECUTION 6 SPATIAL-ADAPTIVE EXECUTION 6 SPATIAL-ADAPTIVE EXECUTION 6.1 Performance Prediction 5.2 Model Adaptation 6.2 Execution Scheduling 7 IMPLEMENTATION 7 IMPLEMENTATION 7 IMPLEMENTATION 8 EVALUATION 8 EVALUATION 8 EVALUATION 8.1 Testbed and Dataset 8.2 Experiment Setup 8.3 Performance 8.4 Robustness 8.5 Component Analysis 8.6 Overhead 9 RELATED WORK 9 RELATED WORK 9 RELATED WORK 10 DISCUSSION AND FUTURE WORK 10 DISCUSSION AND FUTURE WORK 10 DISCUSSION AND FUTURE WORK 11 CONCLUSION AND REFERENCES 11 CONCLUSION AND REFERENCES 11 CONCLUSION AND REFERENCES 5 MULTI-BRANCH OMNIDIRECTIONAL 3D OBJECT DETECTION 5 MULTI-BRANCH OMNIDIRECTIONAL 3D OBJECT DETECTION We describe a multi-branch design of our omnidirectional 3D detection model. We also explain how the model is modified to fit the given device and latency target. 5.1 Model De-sign 5.1 Model De-sign The goal of our multi-branch model is to allow each region in a surrounding space can be processed through various capabilities of the BEV-based 3D detectors detailed in Table 1. Instead of simply using all types of models loaded on the memory, we modularized the models and combined the core modules, eliminating redundant ones. The model follows the baseline pipeline of the BEV-based 3D detection illustrated in Figure 2. However, as shown in Figure 7, an image can be processed by one of several modules at each inference stage. Each branch within the model represents a unique execution path for an image, comprising a sequence of module decisions at all stages. We detailed the branch design following the order of the baseline pipeline. The initial stage of BEV feature generation is to extract 2D feature maps from all multi-view images. At this stage, each image can be processed through one of four backbone. networks, each of which is differentiated by its capacity and input resolution, as detailed in Table 1. From the extracted 2D feature maps, the model infers the 3D depth information in each camera coordinate. To enable the adjustment of quality and latency of the depth prediction, our model is equipped with two types of DepthNets. One is a simple 1- layer convolutional network trained using only object depth loss, providing sparse supervision signals. The other is a deep convolutional network trained on dense depth maps derived from 3D point clouds. As illustrated in Figure 8, the densely supervised DepthNet generates an accurate and detailed depth map from an image. The depth prediction quality is further increased by incorporating a large-scale backbone network with high-resolution images, providing a clear distinction between foreground objects and the background. BEV feature maps are generated by projecting the extracted 2D feature map into a unified grid space using the predicted depth and camera parameters. Fusing the generated BEV features over time enhances the robustness of perception by providing contiguous observations of objects. As shown in Figure 7, we developed a module that can fuse the BEV features extracted from consecutive images of each camera view 𝑗. Due to the possible movements of the camera (or a robot carrying the camera), directly fusing the BEV features may result in misalignment of spatial features over time, greatly reducing the accuracy gain. Therefore, the module aligns the last frame’s BEV feature map 𝐹 𝑗 𝑡−1 into the current frame’s camera coordinate system. The aligned BEV feature map 𝐹 ′𝑗 𝑡−1→𝑡 is acquired as follows: where T is the operation of transforming the spatial locations of features in the BEV grid using the relative camera pose, i.e., the camera motion 𝜃𝑡−1→𝑡 . To estimate 𝜃𝑡−1→𝑡 , we employ a neural network that uses inertial sensor data and consecutive images. At runtime, Panopticus buffers the previous image 𝐼 𝑗 𝑡−1 and 𝐹 𝑗 𝑡−1 of each camera view 𝑗. Since the prediction results of 𝜃𝑡−1→𝑡 are consistent across camera views, the model predicts it once per frame, allowing the sharing of predicted 𝜃𝑡−1→𝑡 with the alignment operations for other views. After aggregating 𝐹 𝑗 𝑡 and the aligned 𝐹 ′𝑗 𝑡−1→𝑡 across all camera views, the model concatenates the integrated BEV feature maps, which are then fed into the BEV head. The resulting 3D bounding boxes are post-processed using nonmax suppression (NMS) to remove duplicates. Panopticus utilizes a 3D object tracker to keep track of the latest states of detected objects. We employed a 3D Kalman filter that can efficiently forecast and update the status of tracked objects. Our model features a branch that outputs the predicted states of tracked objects, skipping new detection for the target camera view. The future state x ′ 𝑡 of a tracked object at the incoming time 𝑡 is predicted as follows: where 𝐴𝑡−1 and x𝑡−1 are the state transition model and the object state vector at time 𝑡−1, respectively. x𝑡−1 is parameterized by the object’s 3D location (𝑥, 𝑦, 𝑧), velocity (𝑣𝑥, 𝑣𝑦, 𝑣𝑧 ) and size (𝑤, ℎ,𝑙). Future state estimation of the target object using 𝐴𝑡−1 is simply applying the velocity predicted by the BEV head to its location, which is processed instantaneously. Overall, the image from each camera view can be processed by one of 17 branches: 16 detection branches and 1 lightweight tracker’s branch utilizing objects’ predicted states. 5.2 Model Adaptation 5.2 Model Adaptation Panopticus supports offline model adaptation to meet the memory and latency constraints on a given device. To achieve this, the memory consumption and processing latency of all modules in the model are profiled on the target device. First, the model is adjusted to adhere to the memory constraints. For example, for the Jetson Orin Nano [8] with 6.3 GB of limited memory, large-sized modules such as the R152 backbone are detached to prevent overloading the memory capacity. Second, some inference branches require significant computation; in fact, inferring a single image may surpass the target latency. To meet latency requirements, branches with latency profiles that exceed the target limit are also removed. The model, with its number of branches 𝑀 changed, is then deployed on the target device. Authors: Jeho Lee
Chanyoung Jung
Jiwon Kim
Hojung Cha Authors: Jeho Lee Chanyoung Jung Jiwon Kim Hojung Cha Jeho Lee Jeho Lee Chanyoung Jung Chanyoung Jung Jiwon Kim Jiwon Kim Hojung Cha Hojung Cha This paper is available on arxiv under CC by 4.0 Deed (Attribution 4.0 International) license. This paper is available on arxiv under CC by 4.0 Deed (Attribution 4.0 International) license. available on arxiv available on arxiv

Edge Devices Are Getting Better at Seeing the World in 3D

About one third of people in the world experience sleep problems. Being able to sleep well can make a big difference in the quality of work you can accomplish during the day and relationships you have with others. It can also improve your overall physical health. For that reason, sleeping more soundly and better is worth working on. In my experience, many people don’t think that simple fixes can improve the quality and amount of sleep. I know I did not think so either. Ultimately after a multi year journey trying to improve my sleep, I found that the simple things really did matter a lot. Let me tell you about them. Let’s go back to the basics. According to the CDC (https://www.cdc.gov/sleep/about/index.html), better sleep practices begin with ideas like this: Going to bed and getting up at the same time every day.
Keeping your bedroom quiet, relaxing, and at a cool temperature.
Turning off electronic devices at least 30 minutes before bedtime.
Avoiding large meals and alcohol before bedtime.
Avoiding caffeine in the afternoon or evening.
Exercising regularly and maintaining a healthy diet. This article from Harvard Health talked about creating a sleep sanctuary (https://www.health.harvard.edu/staying-healthy/sleep-hygiene-simple-practices-for-better-rest) Reduce or mask noise. Heavy curtains and rugs can help absorb sound. A sleep machine that provides white noise, or a recording of soothing sounds such as falling rain, can mask outside noise.
Minimize light. Consider light-blocking curtains or shades if outside lights shine into your bedroom.
Replace your mattress and pillows if they're worn or uncomfortable.
Most people sleep better in a room that's slightly cool. Keep room temperature around 65°F to 68° F at night.
Only use the bedroom for sleep and intimacy.
Have children and pets sleep elsewhere.
Keep electronic devices and anything work-related in another room. Exercise improves sleep also: (https://www.hopkinsmedicine.org/health/wellness-and-prevention/exercising-for-better-sleep) Exercising 30 minutes each day may help with sleep quality
Exercising can help release endorphins and improve mood and sleep I tried many of these things. Although I thought I was doing the right things in order to sleep better, I was not able to see any patterns until I kept a sleep diary. A sleep diary is a written record of how much you sleep and the quality of your sleep along with any stressors that might have affected your sleep. Things that one might include in a sleep diary include when you: Go to bed.
Wake up during the night.
Wake up in the morning.
Take naps.
Exercise.
Drink alcohol or caffeinated drinks.
Take medications. An example of my sleep diary When I did a sleep diary even just for 1 week, I started being more mindful about following better sleep hygiene practices. Just the simple action of noticing my actions more helped me adjust them to sleep better. It was a case of what gets measured, improves. What caused the biggest improvements for me? Managing my light exposure improved my sleep the most. After doing the sleep diary and noticing things in my day to day life, I found out that I was very sensitive to light from electronics and just any ambient light in general. I tried to create a bedroom sleep sanctuary, but as a remote worker, I was not able to remove electronics entirely from my bedroom. I was frustrated when I was not able to completely control my bedroom environment to be more comfortable. While I might have given up there, I eventually learned to: Use an eye mask to cover my eyes when getting in bed and going to sleep. I found there are a variety of eye masks, and I found one that contains lavender herbs. The scent of lavender can help one sleep.
I chose a sunrise alarm clock that wakes you up with gradual light instead of focusing on a typical sound-making alarm clock.
I also created a sleep routine around shutting down and not using my technology about 30 minutes to an hour before bedtime.
If I woke up in the middle of the night and was not able to get back to sleep, I did not use my technology, but instead looked at a book or wrote. Sometimes that helped me get back to sleep. A second method of improving sleep involved reducing my dependence on social media and dopamine releasing apps on my phone. Although I was not able to totally eliminate social media in my life, I was able to minimize its effects on my sleep by: Removing social media apps from my phone to prevent scrolling at night.
Deleting some social media apps that had a more negative effect.
Using a website blocker on my phone and computer so that I would be less tempted to go to social media sites or other sites that are triggering for me. Finally, exercise made a big difference for my sleep too: Being a remote worker, I had struggled to move around enough during my workday. When I incorporated stretching for 5 or 10 mins between work commitments, I started to feel a lot better.


I used a stretching youtube video to help me and guide me
https://youtu.be/Aj6jyIEmZzs?si=6iasl3fwr1Uq6Ufy


I also made an effort to get 30 minutes of exercise daily if possible. Some of these adjustments to my daily routine helped me sleep better. Although I don't always get quality sleep, it has improved some and I feel that makes all this effort worthwhile. I think keeping a sleep diary is a beneficial practice that I may continue doing so that I can see my stressors and be more mindful of what is affecting my life and sleep. ** About one third of people in the world experience sleep problems. Being able to sleep well can make a big difference in the quality of work you can accomplish during the day and relationships you have with others. It can also improve your overall physical health. For that reason, sleeping more soundly and better is worth working on. In my experience, many people don’t think that simple fixes can improve the quality and amount of sleep. I know I did not think so either. Ultimately after a multi year journey trying to improve my sleep, I found that the simple things really did matter a lot. Let me tell you about them. Let’s go back to the basics. About one third of people in the world experience sleep problems. Being able to sleep well can make a big difference in the quality of work you can accomplish during the day and relationships you have with others. It can also improve your overall physical health. For that reason, sleeping more soundly and better is worth working on. In my experience, many people don’t think that simple fixes can improve the quality and amount of sleep. I know I did not think so either. Ultimately after a multi year journey trying to improve my sleep, I found that the simple things really did matter a lot. Let me tell you about them. Let’s go back to the basics. According to the CDC ( https://www.cdc.gov/sleep/about/index.html ), better sleep practices begin with ideas like this: https://www.cdc.gov/sleep/about/index.html https://www.cdc.gov/sleep/about/index.html Going to bed and getting up at the same time every day. Keeping your bedroom quiet, relaxing, and at a cool temperature. Turning off electronic devices at least 30 minutes before bedtime. Avoiding large meals and alcohol before bedtime. Avoiding caffeine in the afternoon or evening. Exercising regularly and maintaining a healthy diet. Going to bed and getting up at the same time every day. Keeping your bedroom quiet, relaxing, and at a cool temperature. Turning off electronic devices at least 30 minutes before bedtime. Avoiding large meals and alcohol before bedtime. Avoiding caffeine in the afternoon or evening. Exercising regularly and maintaining a healthy diet. This article from Harvard Health talked about creating a sleep sanctuary ( https://www.health.harvard.edu/staying-healthy/sleep-hygiene-simple-practices-for-better-rest ) https://www.health.harvard.edu/staying-healthy/sleep-hygiene-simple-practices-for-better-rest https://www.health.harvard.edu/staying-healthy/sleep-hygiene-simple-practices-for-better-rest Reduce or mask noise. Heavy curtains and rugs can help absorb sound. A sleep machine that provides white noise, or a recording of soothing sounds such as falling rain, can mask outside noise. Minimize light. Consider light-blocking curtains or shades if outside lights shine into your bedroom. Replace your mattress and pillows if they're worn or uncomfortable. Most people sleep better in a room that's slightly cool. Keep room temperature around 65°F to 68° F at night. Only use the bedroom for sleep and intimacy. Have children and pets sleep elsewhere. Keep electronic devices and anything work-related in another room. Reduce or mask noise. Heavy curtains and rugs can help absorb sound. A sleep machine that provides white noise, or a recording of soothing sounds such as falling rain, can mask outside noise. Minimize light. Consider light-blocking curtains or shades if outside lights shine into your bedroom. Replace your mattress and pillows if they're worn or uncomfortable. Most people sleep better in a room that's slightly cool. Keep room temperature around 65°F to 68° F at night. Only use the bedroom for sleep and intimacy. Have children and pets sleep elsewhere. Keep electronic devices and anything work-related in another room. Exercise improves sleep also: ( https://www.hopkinsmedicine.org/health/wellness-and-prevention/exercising-for-better-sleep ) https://www.hopkinsmedicine.org/health/wellness-and-prevention/exercising-for-better-sleep https://www.hopkinsmedicine.org/health/wellness-and-prevention/exercising-for-better-sleep Exercising 30 minutes each day may help with sleep quality Exercising can help release endorphins and improve mood and sleep Exercising 30 minutes each day may help with sleep quality Exercising can help release endorphins and improve mood and sleep I tried many of these things. Although I thought I was doing the right things in order to sleep better, I was not able to see any patterns until I kept a sleep diary. A sleep diary is a written record of how much you sleep and the quality of your sleep along with any stressors that might have affected your sleep. Things that one might include in a sleep diary include when you: Go to bed. Wake up during the night. Wake up in the morning. Take naps. Exercise. Drink alcohol or caffeinated drinks. Take medications. Go to bed. Wake up during the night. Wake up in the morning. Take naps. Exercise. Drink alcohol or caffeinated drinks. Take medications. An example of my sleep diary When I did a sleep diary even just for 1 week, I started being more mindful about following better sleep hygiene practices. Just the simple action of noticing my actions more helped me adjust them to sleep better. It was a case of what gets measured, improves. What caused the biggest improvements for me? What caused the biggest improvements for me? Managing my light exposure improved my sleep the most. After doing the sleep diary and noticing things in my day to day life, I found out that I was very sensitive to light from electronics and just any ambient light in general. I tried to create a bedroom sleep sanctuary, but as a remote worker, I was not able to remove electronics entirely from my bedroom. I was frustrated when I was not able to completely control my bedroom environment to be more comfortable. While I might have given up there, I eventually learned to: Use an eye mask to cover my eyes when getting in bed and going to sleep. I found there are a variety of eye masks, and I found one that contains lavender herbs. The scent of lavender can help one sleep. I chose a sunrise alarm clock that wakes you up with gradual light instead of focusing on a typical sound-making alarm clock. I also created a sleep routine around shutting down and not using my technology about 30 minutes to an hour before bedtime. If I woke up in the middle of the night and was not able to get back to sleep, I did not use my technology, but instead looked at a book or wrote. Sometimes that helped me get back to sleep. Use an eye mask to cover my eyes when getting in bed and going to sleep. I found there are a variety of eye masks, and I found one that contains lavender herbs. The scent of lavender can help one sleep. I chose a sunrise alarm clock that wakes you up with gradual light instead of focusing on a typical sound-making alarm clock. I also created a sleep routine around shutting down and not using my technology about 30 minutes to an hour before bedtime. If I woke up in the middle of the night and was not able to get back to sleep, I did not use my technology, but instead looked at a book or wrote. Sometimes that helped me get back to sleep. A second method of improving sleep involved reducing my dependence on social media and dopamine releasing apps on my phone. Although I was not able to totally eliminate social media in my life, I was able to minimize its effects on my sleep by: Removing social media apps from my phone to prevent scrolling at night. Deleting some social media apps that had a more negative effect. Using a website blocker on my phone and computer so that I would be less tempted to go to social media sites or other sites that are triggering for me. Removing social media apps from my phone to prevent scrolling at night. Deleting some social media apps that had a more negative effect. Using a website blocker on my phone and computer so that I would be less tempted to go to social media sites or other sites that are triggering for me. Finally, exercise made a big difference for my sleep too: Being a remote worker, I had struggled to move around enough during my workday. When I incorporated stretching for 5 or 10 mins between work commitments, I started to feel a lot better. I used a stretching youtube video to help me and guide me
https://youtu.be/Aj6jyIEmZzs?si=6iasl3fwr1Uq6Ufy I also made an effort to get 30 minutes of exercise daily if possible. Being a remote worker, I had struggled to move around enough during my workday. When I incorporated stretching for 5 or 10 mins between work commitments, I started to feel a lot better. Being a remote worker, I had struggled to move around enough during my workday. When I incorporated stretching for 5 or 10 mins between work commitments, I started to feel a lot better. I used a stretching youtube video to help me and guide me https://youtu.be/Aj6jyIEmZzs?si=6iasl3fwr1Uq6Ufy I used a stretching youtube video to help me and guide me https://youtu.be/Aj6jyIEmZzs?si=6iasl3fwr1Uq6Ufy https://youtu.be/Aj6jyIEmZzs?si=6iasl3fwr1Uq6Ufy https://youtu.be/Aj6jyIEmZzs?si=6iasl3fwr1Uq6Ufy I also made an effort to get 30 minutes of exercise daily if possible. I also made an effort to get 30 minutes of exercise daily if possible. Some of these adjustments to my daily routine helped me sleep better. Although I don't always get quality sleep, it has improved some and I feel that makes all this effort worthwhile. I think keeping a sleep diary is a beneficial practice that I may continue doing so that I can see my stressors and be more mindful of what is affecting my life and sleep. **

Stories, News & Mentions about BETTER

Categories

Trending Topics

Stories, News & Mentions about BETTER

Better’s Stories on HackerNoon

Better's News & Mentions

Categories

Trending Topics