HighlightCurated by Aramai Editorial

ImageNet Large Scale Visual Recognition Challenge

Describes the ImageNet Large Scale Visual Recognition Challenge benchmark and the advances in object recognition it enabled.

This paper describes the ImageNet Large Scale Visual Recognition Challenge, an annual benchmark since 2010 for object category classification and detection across hundreds of categories and millions of images, with over fifty participating institutions. It covers dataset creation, key breakthroughs in categorical object recognition, and analysis of the state of the field, comparing computer vision accuracy to human accuracy, concluding with lessons learned and proposed future directions.

Based on: ImageNet Large Scale Visual Recognition Challenge · International Journal of Computer Vision

HighlightCurated by Aramai Editorial

Cytoscape: a software environment for integrated models of biomolecular interaction networks.

Introduces Cytoscape, open-source software for visualizing and analyzing biomolecular interaction networks with expression data.

Cytoscape is open-source software integrating biomolecular interaction networks with high-throughput expression data and other molecular states into a unified framework. Its core supports network layout, querying, visual integration with expression profiles and phenotypes, and links to functional annotation databases, and is extensible via plug-ins. Case studies include pathway searches tied to expression changes, DNA-damage-recovery protein complexes, an inferred Halobacterium network, and an interface to kinetic gene regulatory models.

Based on: Cytoscape: a software environment for integrated models of biomolecular interaction networks. · Genome Research

HighlightCurated by Aramai Editorial

Argonaute—a database for gene regulation by mammalian microRNAs

Presents Argonaute, a curated database of mammalian microRNAs, their origins, expression, and predicted target genes.

Argonaute is a curated database compiling information on mammalian microRNAs (miRNAs), small non-coding RNAs that regulate gene expression and are implicated in differentiation, proliferation, apoptosis, and cancer. It records each miRNA's host gene, tissue expression, proposed function, target genes with Gene Ontology annotation, miRNA families, and processing proteins. A web interface allows querying single or multiple miRNAs, linking targets to sequence databases and MEDLINE. It covered 839 miRNAs from human, mouse, and rat.

Based on: Argonaute—a database for gene regulation by mammalian microRNAs · Nucleic Acids Res.

HighlightCurated by Aramai Editorial

You Only Look Once: Unified, Real-Time Object Detection

Introduces YOLO, a single neural network that frames object detection as a regression problem for real-time bounding box and class prediction.

YOLO reframes object detection as a single regression problem, predicting bounding boxes and class probabilities directly from full images in one pass through a unified neural network trained end-to-end. The base model runs at 45 frames per second, while a smaller Fast YOLO variant reaches 155 fps with double the mAP of other real-time detectors. YOLO makes more localization errors than prior systems like DPM and R-CNN but fewer false positives on background, and it generalizes better to non-natural image domains such as artwork.

Based on: You Only Look Once: Unified, Real-Time Object Detection · Computer Vision and Pattern Recognition

HighlightCurated by Aramai Editorial

Neural Machine Translation by Jointly Learning to Align and Translate

Introduces a soft-attention mechanism letting an encoder-decoder network jointly align and translate without fixed-length bottleneck.

Neural machine translation models typically use an encoder-decoder architecture where an encoder compresses a source sentence into a fixed-length vector from which a decoder generates the translation. The authors conjecture this fixed-length vector is a bottleneck, and propose extending the model to automatically (soft-)search source-sentence parts relevant to each target word, without explicit segmentation. This matches existing state-of-the-art phrase-based system performance on English-to-French, and the learned soft alignments agree well with intuition.

Based on: Neural Machine Translation by Jointly Learning to Align and Translate · International Conference on Learning Representations

HighlightCurated by Aramai Editorial

RoBERTa: A Robustly Optimized BERT Pretraining Approach

A replication study showing BERT was undertrained, and that a tuned pretraining recipe matches or beats later models.

Language model pretraining yields strong gains but careful comparison across approaches is difficult, since training is expensive, done on private datasets of varying sizes, and sensitive to hyperparameter choices. This paper presents a replication study of BERT pretraining that measures the impact of key hyperparameters and training data size, finding BERT was significantly undertrained. A better-tuned BERT can match or exceed every model published after it, achieving state-of-the-art results on GLUE, RACE, and SQuAD; the authors release their models and code.

Based on: RoBERTa: A Robustly Optimized BERT Pretraining Approach · arXiv.org

HighlightCurated by Aramai Editorial

Generative Adversarial Networks

A comprehensive guide to Generative Adversarial Networks covering architecture, loss functions, training, applications, and challenges.

This paper is a comprehensive guide to Generative Adversarial Networks (GANs), deep learning techniques known for generating realistic images, videos, and other data. It reviews GAN history, details the generator-discriminator architecture and design variations, and surveys loss functions from the minimax objective to Wasserstein distance with gradient penalty. It covers training techniques, applications, evaluation metrics, and open challenges including mode collapse, instability, and ethics, closing with future directions.

Based on: Generative Adversarial Networks · International Conference on Computing Communication and Networking Technologies

HighlightCurated by Aramai Editorial

Rethinking the Inception Architecture for Computer Vision

Explores factorized convolutions and regularization to scale up convolutional networks efficiently for computer vision.

Since 2014, very deep convolutional networks have driven gains across vision benchmarks, but larger models are not always feasible for mobile or big-data use cases. This work explores scaling networks up while using added computation efficiently, via suitably factorized convolutions and aggressive regularization. On ILSVRC 2012 classification, it reaches 21.2% top-1 and 5.6% top-5 error single-frame with 5 billion multiply-adds and under 25 million parameters, and a 4-model ensemble with multi-crop reaches 3.5% top-5 error on validation.

Based on: Rethinking the Inception Architecture for Computer Vision · Computer Vision and Pattern Recognition

HighlightCurated by Aramai Editorial

Mask R-CNN

Extends Faster R-CNN with a parallel mask-prediction branch for simple, general object instance segmentation.

Mask R-CNN is a simple, flexible framework for object instance segmentation that detects objects while generating a high-quality mask for each instance. It extends Faster R-CNN by adding a branch predicting an object mask in parallel with the bounding-box branch, staying simple to train with small overhead, running at 5 fps. It generalizes to other tasks like human pose estimation and tops all three COCO challenge tracks, beating all existing single-model entries including the COCO 2016 winners.

Based on: Mask R-CNN · arXiv

HighlightCurated by Aramai Editorial

Denoising Diffusion Probabilistic Models

Presents diffusion probabilistic models for high quality image synthesis, connecting them to denoising score matching.

The paper demonstrates high quality image synthesis using diffusion probabilistic models, latent variable models inspired by nonequilibrium thermodynamics. Best results come from training on a weighted variational bound derived from a novel connection to denoising score matching with Langevin dynamics, and the models support a progressive lossy decompression scheme generalizing autoregressive decoding. On unconditional CIFAR10 they obtain an Inception score of 9.46 and a state-of-the-art FID of 3.17, and on 256x256 LSUN they achieve sample quality similar to ProgressiveGAN.

Based on: Denoising Diffusion Probabilistic Models · Neural Information Processing Systems

HighlightCurated by Aramai Editorial

Swin Transformer: Hierarchical Vision Transformer using Shifted Windows

Presents Swin Transformer, a hierarchical vision Transformer using shifted windows as a general-purpose vision backbone.

Adapting Transformers from language to vision is hard because visual entities vary greatly in scale and images have far higher pixel resolution than text has words. Swin Transformer addresses this with a hierarchical architecture using shifted windows, limiting self-attention to non-overlapping local windows while allowing cross-window connections, giving linear complexity in image size. As a general-purpose backbone it reaches 87.3% top-1 on ImageNet-1K, 58.7 box AP and 51.1 mask AP on COCO, and 53.5 mIoU on ADE20K, surpassing prior state of the art by wide margins.

Based on: Swin Transformer: Hierarchical Vision Transformer using Shifted Windows · IEEE International Conference on Computer Vision

HighlightCurated by Aramai Editorial

Bleu: a Method for Automatic Evaluation of Machine Translation

Proposes BLEU, a quick, inexpensive, language-independent automatic method for evaluating machine translation quality.

Human evaluation of machine translation is thorough but slow, costly, and its labor cannot be reused. The authors propose an automatic evaluation method that is fast, inexpensive, and language-independent, correlating highly with human judgments while adding little marginal cost per run. It is presented as an automated substitute for skilled human judges, useful whenever quick or frequent evaluation is needed.

Based on: Bleu: a Method for Automatic Evaluation of Machine Translation · Annual Meeting of the Association for Computational Linguistics