Highlight

Unsupervised Learning of Visual Features by Contrasting Cluster Assignments

Introduces SwAV, an online self-supervised method that clusters image views and swaps cluster-assignment predictions, avoiding pairwise comparisons.

Contrastive methods narrowed the gap between unsupervised and supervised pretraining but rely on costly pairwise feature comparisons. This paper proposes SwAV, an online algorithm that clusters data while enforcing consistency between cluster assignments for different augmented views, via a swapped prediction mechanism instead of direct comparison. It needs no memory bank or momentum network and scales to any batch size. A new multi-crop augmentation adds mixed-resolution views cheaply. SwAV reaches 75.3% top-1 on ImageNet with ResNet-50, surpassing supervised pretraining on transfer tasks.

Based on: Unsupervised Learning of Visual Features by Contrasting Cluster Assignments · Neural Information Processing Systems

Curated by Aramai Editorial

Read summary →
Highlight

ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design

Proposes ShuffleNet V2 and practical design guidelines for efficient CNNs by evaluating direct speed on target hardware rather than only FLOPs.

Neural architecture design is usually guided by an indirect metric, computation complexity (FLOPs), yet direct metrics like speed also depend on memory access cost and platform characteristics. This work argues for evaluating the direct metric on the target platform and, through controlled experiments, derives several practical guidelines for efficient network design. Based on these, it presents a new architecture, ShuffleNet V2. Comprehensive ablation experiments confirm the model is state-of-the-art in the tradeoff between speed and accuracy.

Based on: ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design · European Conference on Computer Vision

Curated by Aramai Editorial

Read summary →
Highlight

Computational Radiomics System to Decode the Radiographic Phenotype

Presents PyRadiomics, an open-source Python platform extracting a large panel of engineered radiomic features from medical images to standardize analyses.

Radiomics quantifies phenotypic traits on medical imaging using automated algorithms, enabling non-invasive biomarkers, but a lack of standardized algorithm definitions and image processing hampers reproducibility. To address this, the authors built PyRadiomics, a flexible open-source platform that extracts a large panel of engineered features from medical images. Implemented in Python, it runs standalone or via 3D-Slicer. The paper describes its architecture and demonstrates it on lung-lesion characterization, aiming to set a reference standard.

Based on: Computational Radiomics System to Decode the Radiographic Phenotype · Cancer Research

Curated by Aramai Editorial

Read summary →
Highlight

Fundamentals of Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) Network

A tutorial that formally derives RNN and LSTM equations from differential equations, justifies unrolling, and proposes a generalized Vanilla LSTM.

LSTM networks are widely covered, but most articles state inference formulas axiomatically, omit training formulas, and present RNN 'unrolling' without justification. This tutorial explains essential RNN and LSTM fundamentals in one document. Drawing on signal processing, it formally derives the canonical RNN from differential equations and proves a statement yielding the unrolling technique. It then transforms the RNN into a Vanilla LSTM through logical arguments, provides all governing equations, and introduces extensions producing the most general LSTM variant to date.

Based on: Fundamentals of Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) Network · Physica A: Statistical Mechanics and its Applications

Curated by Aramai Editorial

Read summary →
Highlight

Long-term recurrent convolutional networks for visual recognition and description

Introduces LRCN, an end-to-end recurrent convolutional architecture uniting CNNs and RNNs for video recognition, captioning, and narration.

This work asks whether recurrent, 'temporally deep' models help on sequential tasks such as video. It develops a novel, end-to-end trainable recurrent convolutional architecture and demonstrates it on video recognition, image description and retrieval, and video narration. Unlike models with a fixed spatio-temporal receptive field or simple temporal averaging, these 'doubly deep' models are compositional in space and time, map variable-length inputs to variable-length outputs, and train via backpropagation while jointly learning temporal dynamics and visual representations.

Based on: Long-term recurrent convolutional networks for visual recognition and description · Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →
Highlight

An experimental comparison of min-cut/max- flow algorithms for energy minimization in vision

Experimentally compares min-cut/max-flow algorithms for energy minimization in vision, introducing a new method often several times faster than others.

Min-cut/max-flow algorithms have become useful tools for energy minimization in low-level vision, but their practical efficiency had mostly been studied outside computer vision. This paper experimentally compares running times of several standard algorithms—including Goldberg-Tarjan push-relabel and Ford-Fulkerson augmenting-paths methods—plus a new algorithm the authors developed. Benchmarks span image restoration, stereo, and segmentation. In many cases the new algorithm runs several times faster, enabling near real-time performance.

Based on: An experimental comparison of min-cut/max- flow algorithms for energy minimization in vision · IEEE Transactions on Pattern Analysis and Machine Intelligence

Curated by Aramai Editorial

Read summary →
Highlight

Bag of Tricks for Efficient Text Classification

Presents fastText, a simple, fast text-classification baseline that rivals deep learning accuracy while training orders of magnitude faster on CPU.

This paper explores a simple and efficient baseline for text classification. The authors' fast classifier, fastText, is often on par with deep learning classifiers in accuracy while being many orders of magnitude faster to train and evaluate. It can be trained on more than one billion words in under ten minutes using a standard multicore CPU, and can classify half a million sentences among 312K classes in less than a minute.

Based on: Bag of Tricks for Efficient Text Classification · Conference of the European Chapter of the Association for Computational Linguistics

Curated by Aramai Editorial

Read summary →
Highlight

The MR-Base platform supports systematic causal inference across the human phenome

Introduces MR-Base, a platform pairing a curated GWAS database with software that automates 2-sample Mendelian randomization for causal inference.

GWAS results can infer causal relationships between phenotypes via 2-sample Mendelian randomization (2SMR) without individual-level data, but rapidly evolving methods and poorly curated data hinder implementation. MR-Base integrates a curated database of complete GWAS results with an API, web app, and R packages that automate 2SMR, plus sensitivity analyses for horizontal pleiotropy. The database holds 11 billion SNP-trait associations from 1673 GWAS and is regularly updated, letting millions of potential causal relationships be evaluated in phenome-wide studies.

Based on: The MR-Base platform supports systematic causal inference across the human phenome · eLife

Curated by Aramai Editorial

Read summary →
Highlight

Review of deep learning: concepts, CNN architectures, challenges, applications, future directions

A comprehensive survey of deep learning: concepts, CNN architectures from AlexNet to HRNet, challenges, applications, and future directions.

This review offers a holistic survey of deep learning (DL), the dominant machine learning paradigm, which matches or exceeds human performance on many complex tasks across domains like cybersecurity, NLP, bioinformatics, and robotics. Noting that prior reviews each covered only one aspect, it presents DL's importance, its techniques and network types, and focuses on convolutional neural networks (CNNs), tracing architectures from AlexNet to the High-Resolution network (HRNet). It also covers challenges and solutions, applications, computational tools (FPGA, GPU, CPU), and benchmark datasets.

Based on: Review of deep learning: concepts, CNN architectures, challenges, applications, future directions · Journal of Big Data

Curated by Aramai Editorial

Read summary →
Highlight

Least Squares Generative Adversarial Networks

Introduces Least Squares GANs, which use a least squares discriminator loss to improve image quality and stabilize GAN training.

Regular GANs treat the discriminator as a classifier with a sigmoid cross-entropy loss, which the authors show can cause vanishing gradients. This paper proposes Least Squares GANs (LSGANs), which adopt a least squares loss for the discriminator. Minimizing the LSGAN objective corresponds to minimizing the Pearson chi-square divergence. LSGANs generate higher quality images and train more stably. Experiments on LSUN and CIFAR-10 show better image quality than regular GANs, and comparisons illustrate improved stability.

Based on: Least Squares Generative Adversarial Networks · IEEE International Conference on Computer Vision

Curated by Aramai Editorial

Read summary →
Highlight

Isogeometric analysis : CAD, finite elements, NURBS, exact geometry and mesh refinement

Proposes isogeometric analysis, using NURBS basis functions to unify exact CAD geometry with finite element analysis and mesh refinement.

The paper proposes isogeometric analysis, in which basis functions from NURBS (Non-Uniform Rational B-Splines) construct an exact geometric model. For analysis, the basis is refined or order-elevated without changing the geometry, giving analogues of finite element h- and p-refinement plus a new, more efficient higher-order k-refinement. Exact geometry is maintained without further CAD communication. Numerical examples show optimal convergence for linear elasticity, convergence to thin elastic shells, and monotone behavior for advection-diffusion with sharp layers.

Based on: Isogeometric analysis : CAD, finite elements, NURBS, exact geometry and mesh refinement · Semantic Scholar

Curated by Aramai Editorial

Read summary →
Highlight

Spectral Networks and Locally Connected Networks on Graphs

Generalizes CNNs to graphs via hierarchical clustering and the graph Laplacian spectrum, learning convolutions independent of input size.

Convolutional neural networks are highly efficient for image and audio recognition because they exploit the local translational invariance of signals. This paper explores generalizing CNNs to signals defined on general domains lacking a translation group. It proposes two constructions: one based on hierarchical clustering of the domain, another on the spectrum of the graph Laplacian. Experiments show that for low-dimensional graphs, convolutional layers can be learned with parameters independent of input size, yielding efficient deep architectures.

Based on: Spectral Networks and Locally Connected Networks on Graphs · International Conference on Learning Representations

Curated by Aramai Editorial

Read summary →