Highlight

Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation

Presents GNMT, Google's deep LSTM neural machine translation system with attention, wordpieces, and low-precision inference for production use.

Neural machine translation (NMT) enables end-to-end translation but is costly to train and run and handles rare words poorly. GNMT uses a deep LSTM with 8 encoder and 8 decoder layers plus attention and residual connections, wiring the decoder's bottom layer to the encoder's top layer for parallelism. It uses low-precision inference for speed, sub-word 'wordpieces' for rare words, and length-normalized beam search with a coverage penalty. On WMT'14 benchmarks it rivals the state of the art, and human evaluation shows 60% fewer errors than Google's phrase-based system.

Based on: Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation · arXiv.org

Curated by Aramai Editorial

Read summary →
Highlight

Accurate Image Super-Resolution Using Very Deep Convolutional Networks

Presents a highly accurate single-image super-resolution method using a very deep 20-layer VGG-inspired convolutional network with residual learning.

The authors present a highly accurate single-image super-resolution method built on a very deep convolutional network inspired by VGG-net. Increasing depth to 20 weight layers significantly improves accuracy, as cascading small filters exploits context over large image regions efficiently. To address slow convergence, they learn only residuals and use extremely high learning rates (10^4 times higher than SRCNN) enabled by adjustable gradient clipping. The method outperforms existing approaches in accuracy, with easily noticeable visual improvements.

Based on: Accurate Image Super-Resolution Using Very Deep Convolutional Networks · Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →
Highlight

Enhanced Deep Residual Networks for Single Image Super-Resolution

Introduces EDSR and multi-scale MDSR deep residual networks for single-image super-resolution, winners of the NTIRE2017 SR Challenge.

Building on deep convolutional networks and residual learning, the paper develops EDSR, an enhanced deep super-resolution network that surpasses state-of-the-art methods. Its gains come largely from removing unnecessary modules in conventional residual networks, then expanding model size while stabilizing training. The authors also propose MDSR, a multi-scale system and training method that reconstructs high-resolution images at several upscaling factors within one model. Both outperform prior methods on benchmarks and won the NTIRE2017 Super-Resolution Challenge.

Based on: Enhanced Deep Residual Networks for Single Image Super-Resolution · 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)

Curated by Aramai Editorial

Read summary →
Highlight

Enrichr: interactive and collaborative HTML5 gene list enrichment analysis tool

Presents Enrichr, a web-based gene list enrichment analysis tool with new gene-set libraries, a new ranking approach, and interactive D3 visualizations.

Genome-wide profiling produces lists of differentially expressed genes that must be analyzed for collective function, usually by computing enrichment against gene-set libraries from prior knowledge. The authors present Enrichr, an integrative web-based and mobile app with new gene-set libraries, an alternative method to rank enriched terms, and interactive visualizations built with the D3 JavaScript library. It can be embedded into other gene-list analysis tools. Applied to nine cancer cell lines vs matched normal tissues, it revealed common PRC2/H3K27me3 enrichment.

Based on: Enrichr: interactive and collaborative HTML5 gene list enrichment analysis tool · BMC Bioinformatics

Curated by Aramai Editorial

Read summary →
Highlight

BEDTools: a flexible suite of utilities for comparing genomic features

Introduces BEDTools, an efficient C++ suite for comparing, manipulating, and annotating genomic features in BED, GFF, and BAM formats.

Testing correlations between sets of genomic features is fundamental, but web-based overlap searches struggle with today's massive sequencing datasets. BEDTools is a suite for comparing, manipulating, and annotating genomic features in BED and GFF formats, and can compare BAM alignments against BED and GFF features. The tools are efficient, handle large sequencing datasets with public and custom annotation tracks, and chain with each other or standard UNIX commands to build pipelines. BEDTools is written in C++ with freely available source and documentation.

Based on: BEDTools: a flexible suite of utilities for comparing genomic features · Bioinform.

Curated by Aramai Editorial

Read summary →
Highlight

EfficientDet: Scalable and Efficient Object Detection

Introduces EfficientDet, a family of object detectors using a weighted bi-directional feature pyramid network (BiFPN) and compound scaling for efficiency.

This paper studies architecture choices for efficient object detection and proposes two optimizations: a weighted bi-directional feature pyramid network (BiFPN) for fast multi-scale feature fusion, and a compound scaling method that jointly scales resolution, depth, and width across all sub-networks. Built on EfficientNet backbones, the resulting EfficientDet detectors are far more efficient than prior art. EfficientDet-D7 reaches 52.2 AP on COCO test-dev with 52M parameters and 325B FLOPs, being 4x-9x smaller and using 13x-42x fewer FLOPs than previous detectors.

Based on: EfficientDet: Scalable and Efficient Object Detection · Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →
Highlight

AERONET-a federated instrument network and data archive for aerosol Characterization

Describes AERONET, a NASA-initiated federated network of automatic sun/sky radiometers and open data archive for aerosol characterization.

AERONET is a NASA-initiated, internationally expanded network for remote sensing of atmospheric aerosols. Weather-resistant automatic sun and sky radiometers measure aerosol optical properties and precipitable water at remote sites, relaying data through geostationary satellites for near real-time processing over most of the Earth. NASA built a UNIX-based system giving internet access to a centralized, open-access database with standardized measurements, supporting radiation budget studies, radiative transfer modeling, and validation of satellite aerosol retrievals.

Based on: AERONET-a federated instrument network and data archive for aerosol Characterization · Semantic Scholar

Curated by Aramai Editorial

Read summary →
Highlight

Self-Consistency Improves Chain of Thought Reasoning in Language Models

Introduces self-consistency, a decoding strategy that samples diverse reasoning paths and picks the most consistent answer for chain-of-thought prompting.

Chain-of-thought prompting with large language models works well on complex reasoning. This paper proposes self-consistency, a decoding strategy replacing naive greedy decoding: it samples a diverse set of reasoning paths, then selects the most consistent final answer by marginalizing over them, on the intuition that hard problems admit multiple valid routes to one correct answer. Across arithmetic and commonsense benchmarks it substantially improves accuracy, including GSM8K (+17.9%), SVAMP (+11.0%), AQuA (+12.2%), StrategyQA (+6.4%) and ARC-challenge (+3.9%).

Based on: Self-Consistency Improves Chain of Thought Reasoning in Language Models · International Conference on Learning Representations

Curated by Aramai Editorial

Read summary →
Highlight

Spatio-temporal Graph Convolutional Neural Network: A Deep Learning Framework for Traffic Forecasting

Proposes STGCN, a fully convolutional graph deep learning framework for fast, accurate traffic forecasting on road networks.

STGCN, or Spatio-Temporal Graph Convolutional Networks, is proposed for traffic forecasting, where highly nonlinear flow makes mid- and long-term prediction hard and spatial-temporal dependencies are often neglected. Instead of regular convolutional and recurrent units, it formulates the problem on graphs with complete convolutional structures, enabling faster training with fewer parameters. Experiments show STGCN captures comprehensive spatio-temporal correlations by modeling multi-scale traffic networks and consistently beats state-of-the-art baselines on real-world datasets.

Based on: Spatio-temporal Graph Convolutional Neural Network: A Deep Learning Framework for Traffic Forecasting · International Joint Conference on Artificial Intelligence

Curated by Aramai Editorial

Read summary →
Highlight

DeCAF: A Deep Convolutional Activation Feature for Generic Visual Recognition

Introduces DeCAF, showing features from a supervised deep CNN transfer to novel recognition tasks and beat prior state-of-the-art.

DeCAF evaluates whether features from a deep convolutional network trained fully supervised on a large fixed set of object recognition tasks transfer to novel generic tasks, even when tasks differ significantly and labeled data is scarce. It investigates and visualizes the semantic clustering of deep features across scene recognition, domain adaptation, and fine-grained recognition, comparing features from different network layers. The transferred features beat prior state of the art on several vision challenges, and DeCAF is released open-source with its network parameters.

Based on: DeCAF: A Deep Convolutional Activation Feature for Generic Visual Recognition · International Conference on Machine Learning

Curated by Aramai Editorial

Read summary →
Highlight

Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions

Introduces PVT, a convolution-free pyramid Transformer backbone for dense prediction tasks like detection and segmentation.

The Pyramid Vision Transformer (PVT) is a convolution-free backbone for dense prediction, unlike ViT which targets image classification. A progressive shrinking pyramid lets PVT produce high-resolution outputs while reducing computation on large feature maps, and it inherits advantages of both CNNs and Transformers as a unified backbone that directly replaces CNN backbones. Across object detection, instance and semantic segmentation it boosts performance; PVT+RetinaNet reaches 40.4 AP on COCO, surpassing ResNet50+RetinaNet's 36.3 AP by 4.1 points.

Based on: Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions · IEEE International Conference on Computer Vision

Curated by Aramai Editorial

Read summary →
Highlight

CNN Features Off-the-Shelf: An Astounding Baseline for Recognition

Shows off-the-shelf CNN features from OverFeat, with a linear SVM, beat tuned state-of-the-art systems across many recognition tasks.

Generic descriptors extracted from convolutional networks are remarkably powerful. Using features from the publicly available OverFeat network, trained for object classification on ILSVRC13, as a generic image representation, the paper addresses tasks moving further from the original: object and scene recognition, fine-grained recognition, attribute detection, and image retrieval. A simple linear SVM (or L2 distance for retrieval) on a 4096-dimensional feature consistently beats highly tuned state-of-the-art systems, suggesting deep CNN features should be the default choice.

Based on: CNN Features Off-the-Shelf: An Astounding Baseline for Recognition · 2014 IEEE Conference on Computer Vision and Pattern Recognition Workshops

Curated by Aramai Editorial

Read summary →