How JEPA Avoids Noisy Features: The Implicit Bias of Deep Linear Self Distillation Networks

How JEPA Avoids Noisy Features: The Implicit Bias of Deep Linear Self Distillation Networks

arXiv:2407.03475v1 Announce Type: new Abstract: Two competing paradigms exist for self-supervised learning of data representations. Joint Embedding Predictive Architecture (JEPA) is a class of architectures in which semantically similar inputs are encoded into representations that are predictive of each other. A recent successful approach that falls under the JEPA framework is self-distillation, where an online encoder is trained to predict the output of the target encoder, sometimes using a lightweight predictor network. This is contrasted with the Masked AutoEncoder (MAE) paradigm, where an encoder and decoder are trained to reconstruct missing parts of the input in the data space rather,…
Read More
Setting your remote development environment

Setting your remote development environment

Introduction At first, it works okay when you set your dev environment up and running. You might only need to run react dev server and next dev server. However, you'll see that things become slower when you run storybook dev server. What's worse, more components are added. Plus, you want to run unit/integration/e2e tests locally before making a pull request. Some developers might have a brand-new performant machine to handle all of them. Unfortunately, it wasn't me. To solve this issue, I looked into how to speed up my development environment. Buying a new local machine Most individuals choose this…
Read More
Visual Robustness Benchmark for Visual Question Answering (VQA)

Visual Robustness Benchmark for Visual Question Answering (VQA)

arXiv:2407.03386v1 Announce Type: new Abstract: Can Visual Question Answering (VQA) systems perform just as well when deployed in the real world? Or are they susceptible to realistic corruption effects e.g. image blur, which can be detrimental in sensitive applications, such as medical VQA? While linguistic or textual robustness has been thoroughly explored in the VQA literature, there has yet to be any significant work on the visual robustness of VQA models. We propose the first large-scale benchmark comprising 213,000 augmented images, challenging the visual robustness of multiple VQA models and assessing the strength of realistic visual corruptions. Additionally, we have…
Read More
UnSeenTimeQA: Time-Sensitive Question-Answering Beyond LLMs’ Memorization

UnSeenTimeQA: Time-Sensitive Question-Answering Beyond LLMs’ Memorization

arXiv:2407.03525v1 Announce Type: new Abstract: This paper introduces UnSeenTimeQA, a novel time-sensitive question-answering (TSQA) benchmark that diverges from traditional TSQA benchmarks by avoiding factual and web-searchable queries. We present a series of time-sensitive event scenarios decoupled from real-world factual information. It requires large language models (LLMs) to engage in genuine temporal reasoning, disassociating from the knowledge acquired during the pre-training phase. Our evaluation of six open-source LLMs (ranging from 2B to 70B in size) and three closed-source LLMs reveal that the questions from the UnSeenTimeQA present substantial challenges. This indicates the models' difficulties in handling complex temporal reasoning scenarios. Additionally,…
Read More
ChatGPT’s macOS app was storing chats in plain text – gHacks Tech News

ChatGPT’s macOS app was storing chats in plain text – gHacks Tech News

A software engineer has discovered that OpenAI's ChatGPT app for Mac was saving chats in plain text. Here is what happened. In case you missed it, the ChatGPT Mac app was released a week ago for all users. Pedro José Pereira Vieito published his findings on Threads, to reveal that the popular chatbot's desktop app was storing conversations that a user had with the chatbot, in plain text format on the local storage. ChatGPT for Mac was storing conversations in plain text The security enthusiast noted that macOS has been blocking other apps from snooping into user data, since macOS…
Read More
Boeing’s big mea culpa just dropped and it doesn’t look pretty

Boeing’s big mea culpa just dropped and it doesn’t look pretty

Boeing has agreed to plead guilty to defrauding the US, essentially admitting to accusations that it violated an earlier agreement to strengthen its safety measures in the wake of two fatal Boeing 737 Max crashes in 2018 and 2019."The parties have agreed that Boeing will plead guilty to the most serious readily provably offense," the Justice Department said in a court filing on Sunday night.Last week, Bloomberg reported that federal prosecutors had offered Boeing the choice of either accepting the plea deal or risk facing trial.Under the plea deal, Boeing will have to pay a fine of $243.6 million. This…
Read More
Data Machina #260

Data Machina #260

Vision-Language Models Booming. VLMs are experiencing a boom. Large foundation models like OpenAI GPT-4o, Anthropic Claude 3.5 Sonnet, and Google Gemini Pro 1.5 keep showing amazing vision-language capabilities and still dominate the benchmarks. But in a race to democratise VLMs at an efficient cost of operation -while maintaining performance- there is a new type of emerging, small, versatile, specialised VLMs that are becoming very powerful. And that is great!Start here: The best intro to VLMs, 2024. Probably - by far- the best introduction to VLMs. A mega paper in the format of a pdf book, published by Meta AI, NYU,…
Read More
A Role of Environmental Complexity on Representation Learning in Deep Reinforcement Learning Agents

A Role of Environmental Complexity on Representation Learning in Deep Reinforcement Learning Agents

arXiv:2407.03436v1 Announce Type: new Abstract: The environments where individuals live can present diverse navigation challenges, resulting in varying navigation abilities and strategies. Inspired by differing urban layouts and the Dual Solutions Paradigm test used for human navigators, we developed a simulated navigation environment to train deep reinforcement learning agents in a shortcut usage task. We modulated the frequency of exposure to a shortcut and navigation cue, leading to the development of artificial agents with differing abilities. We examined the encoded representations in artificial neural networks driving these agents, revealing intricate dynamics in representation learning, and correlated them with shortcut use…
Read More
Power-hungry AI is driving a surge in tech giant carbon emissions. Nobody knows what to do about it

Power-hungry AI is driving a surge in tech giant carbon emissions. Nobody knows what to do about it

Since the release of ChatGPT in November 2022, the world has seen an incredible surge in investment, development and use of artificial intelligence (AI) applications. According to one estimate, the amount of computational power used for AI is doubling roughly every 100 days. The social and economic impacts of this boom have provoked reactions around the world. European regulators recently pushed Meta to pause plans to train AI models on users’ Facebook and Instagram data. The Bank of International Settlements, which coordinates the world’s central banks, has warned AI adoption may change the way inflation works. The environmental impacts have…
Read More
No widgets found. Go to Widget page and add the widget in Offcanvas Sidebar Widget Area.