Highlight

Prioritized Experience Replay

Proposes prioritized experience replay, which samples important transitions more often to make deep reinforcement learning more efficient.

Experience replay lets online reinforcement learning agents reuse past experiences, but uniform sampling replays transitions at their original frequency regardless of importance. This paper introduces a framework that prioritizes experience so that significant transitions are replayed more often, improving learning efficiency. Applied to Deep Q-Networks (DQN), prioritized replay sets a new state of the art, outperforming DQN with uniform replay on 41 of 49 Atari games.

Based on: Prioritized Experience Replay · International Conference on Learning Representations

Curated by Aramai Editorial

Read summary →
Highlight

HellaSwag: Can a Machine Really Finish Your Sentence?

Introduces HellaSwag, an adversarially-filtered commonsense NLI benchmark that is easy for humans but hard for state-of-the-art models.

HellaSwag is a challenge dataset for commonsense natural language inference, where a model must pick the most plausible continuation of an event description. It is built with Adversarial Filtering, in which discriminators iteratively select machine-generated wrong answers, scaling examples into a 'Goldilocks' zone that is trivial for humans (>95%) yet hard for top models (<48%). The work sheds light on pretrained models and argues that benchmarks should co-evolve adversarially with the state of the art.

Based on: HellaSwag: Can a Machine Really Finish Your Sentence? · Annual Meeting of the Association for Computational Linguistics

Curated by Aramai Editorial

Read summary →
Highlight

STRING v9.1: protein-protein interaction networks, with increased coverage and integration

Describes STRING v9.1, a database integrating known and predicted protein-protein interactions across 1100+ organisms with improved text mining.

Complete knowledge of protein interactions would be a major step toward understanding cellular mechanisms, yet such data are scattered across resources at varying detail. The STRING database gives a global view of all interaction data—including lower-quality data and predictions—by scoring and integrating known and predicted associations into networks covering over 1100 organisms. Version 9.1 adds full-text text mining, a redesigned cross-organism interaction-transfer algorithm, and functional-enrichment statistics for user networks.

Based on: STRING v9.1: protein-protein interaction networks, with increased coverage and integration · Nucleic Acids Res.

Curated by Aramai Editorial

Read summary →
Highlight

Qwen2.5 Technical Report

Technical report on Qwen2.5, a series of LLMs scaled to 18T pretraining tokens with extensive post-training, released in open-weight and MoE variants.

Qwen2.5 is a comprehensive series of large language models improved in pre-training and post-training. Pre-training data was scaled from 7 to 18 trillion tokens to strengthen knowledge and reasoning, while post-training used supervised finetuning on 1M+ samples plus multi-stage reinforcement learning to improve instruction following, long-text generation, and structured-data analysis. It ships in many sizes: open-weight base, instruction-tuned, and quantized models, plus proprietary MoE variants Turbo and Plus. Qwen2.5-72B-Instruct rivals the ~5x larger Llama-3-405B.

Based on: Qwen2.5 Technical Report · arXiv.org

Curated by Aramai Editorial

Read summary →
Highlight

SUN database: Large-scale scene recognition from abbey to zoo

Introduces the SUN database of 899 scene categories and 130,519 images, and benchmarks scene-recognition algorithms against human performance.

Scene categorization is a fundamental computer vision problem, but progress has been limited by databases covering only a few scene categories—the largest prior set had just 15 classes. The authors propose the Scene UNderstanding (SUN) database of 899 categories and 130,519 images. Using 397 well-sampled categories, they evaluate many state-of-the-art scene-recognition algorithms and set new performance bounds. They also measure human scene-classification performance for comparison and study a finer-grained representation to detect scenes embedded within larger scenes.

Based on: SUN database: Large-scale scene recognition from abbey to zoo · 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →
Highlight

DeepLabCut: markerless pose estimation of user-defined body parts with deep learning

Presents DeepLabCut, a markerless pose-estimation toolbox using transfer learning to track user-defined body parts with minimal labeled frames.

Quantifying behavior is crucial in neuroscience, but extracting specific behaviors from video is time consuming, and physical markers for tracking are intrusive and require predetermined placement. The authors present DeepLabCut, a markerless pose-estimation method based on transfer learning with deep neural networks that works well with minimal training data. They show its versatility by tracking diverse body parts across multiple species and behaviors. Even with only about 200 labeled frames, it reaches test-frame tracking comparable to human accuracy.

Based on: DeepLabCut: markerless pose estimation of user-defined body parts with deep learning · Nature Neuroscience

Curated by Aramai Editorial

Read summary →
Highlight

Tree of Thoughts: Deliberate Problem Solving with Large Language Models

Introduces Tree of Thoughts, a framework letting LLMs explore and self-evaluate multiple reasoning paths with lookahead and backtracking.

LLMs are widely used for problem solving but remain confined to token-level, left-to-right inference, limiting tasks needing exploration or strategic lookahead. The authors introduce Tree of Thoughts (ToT), generalizing Chain-of-Thought prompting by letting models explore coherent text units ('thoughts') as intermediate steps. ToT considers multiple reasoning paths, self-evaluates choices, and looks ahead or backtracks for global decisions. On Game of 24, Creative Writing, and Mini Crosswords it sharply improves results—raising GPT-4's Game of 24 success from 4% to 74%.

Based on: Tree of Thoughts: Deliberate Problem Solving with Large Language Models · Neural Information Processing Systems

Curated by Aramai Editorial

Read summary →
Highlight

Transformer-XL: Attentive Language Models beyond a Fixed-Length Context

Proposes Transformer-XL, adding segment-level recurrence and a new positional encoding so language models learn dependency beyond a fixed-length context.

Transformers learn long-term dependency but are limited by a fixed-length context in language modeling. The authors propose Transformer-XL, combining segment-level recurrence with a novel positional encoding to learn dependencies beyond a fixed length without disrupting temporal coherence, also resolving context fragmentation. It captures dependency 80% longer than RNNs and 450% longer than vanilla Transformers, performs better on short and long sequences, is up to 1,800x faster at evaluation, and sets new state-of-the-art results on five benchmarks.

Based on: Transformer-XL: Attentive Language Models beyond a Fixed-Length Context · Annual Meeting of the Association for Computational Linguistics

Curated by Aramai Editorial

Read summary →
Highlight

Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference

Introduces a quantization scheme enabling integer-only neural network inference, co-designed with a training procedure that preserves accuracy.

Motivated by the cost of deep models on mobile devices, the authors propose a quantization scheme that runs inference using integer-only arithmetic, more efficient than floating-point on common integer-only hardware. They co-design a training procedure that preserves end-to-end accuracy after quantization, improving the accuracy-versus-latency tradeoff. Gains are significant even for MobileNets, an already efficient family, and are demonstrated on ImageNet classification and COCO detection on popular CPUs.

Based on: Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference · 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →
Highlight

High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs

Presents a conditional GAN that synthesizes 2048x1024 photo-realistic images from semantic label maps and supports interactive semantic manipulation.

The authors synthesize high-resolution, photo-realistic images from semantic label maps using conditional GANs. To overcome prior low-resolution, unrealistic results, they propose a new adversarial loss plus multi-scale generator and discriminator architectures, producing 2048x1024 images. Instance segmentation enables adding, removing, or changing objects, and another feature generates diverse outputs for the same input for interactive appearance editing. Human studies show it significantly outperforms existing methods in quality and resolution.

Based on: High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs · 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →
Highlight

A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks

Proposes a simple softmax-probability baseline for detecting misclassified and out-of-distribution examples across vision, NLP, and speech tasks.

The paper addresses two related problems: detecting whether an example is misclassified or out-of-distribution. It presents a simple baseline using the maximum softmax probability, since correctly classified inputs tend to have higher maximum probabilities than erroneous or out-of-distribution ones. The authors define evaluation tasks across computer vision, NLP, and speech recognition, showing the baseline works throughout. They also show it can sometimes be surpassed, highlighting room for future work on these underexplored tasks.

Based on: A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks · International Conference on Learning Representations

Curated by Aramai Editorial

Read summary →
Highlight

EEGNet: a compact convolutional neural network for EEG-based brain–computer interfaces

Introduces EEGNet, a compact CNN using depthwise and separable convolutions to classify EEG signals across multiple brain-computer interface paradigms.

BCIs use EEG signals for control, but feature extractors and classifiers are usually tailored to one paradigm, limiting generality. The authors introduce EEGNet, a compact CNN using depthwise and separable convolutions to encapsulate EEG feature-extraction concepts. They evaluate it within- and cross-subject against state-of-the-art methods across four paradigms (P300, ERN, MRCP, SMR), and add three ways to visualize the learned features.

Based on: EEGNet: a compact convolutional neural network for EEG-based brain–computer interfaces · Journal of Neural Engineering

Curated by Aramai Editorial

Read summary →