Highlight

Residual Dense Network for Image Super-Resolution

Proposes the Residual Dense Network, which fully exploits hierarchical features from all conv layers for image super-resolution.

Very deep CNNs have succeeded at image super-resolution but most fail to fully use hierarchical features from low-resolution inputs, limiting performance. The paper proposes the Residual Dense Network (RDN), exploiting features from all convolutional layers. Its residual dense block extracts abundant local features via densely connected layers with direct connections from preceding blocks, forming a contiguous memory mechanism. Local and global feature fusion then adaptively combine local and holistic features, and RDN performs favorably against state-of-the-art methods.

Based on: Residual Dense Network for Image Super-Resolution · 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →
Highlight

Optimization Methods for Large-Scale Machine Learning

Reviews numerical optimization for machine learning, centering on stochastic gradient methods and directions for next-generation algorithms.

This review examines numerical optimization algorithms for machine learning. Through case studies on text classification and deep neural network training, it discusses how optimization problems arise and why they are challenging. A central theme is that large-scale machine learning is a setting where the stochastic gradient (SG) method has been central while conventional gradient-based techniques falter. The authors present a theory of a versatile SG algorithm and highlight next-generation directions, including noise-reduction and second-order methods.

Based on: Optimization Methods for Large-Scale Machine Learning · SIAM Review

Curated by Aramai Editorial

Read summary →
Highlight

A Comparison

A case study comparing three text analysis approaches for classifying patient status from clinic letters to inform scheduling.

This case study compares text analysis approaches for classifying a patient's current status to inform scheduling for a large UK healthcare provider. The aim is to systematically capture patient outcomes after clinic attendance, closing records at discharge and scheduling follow-ups within safe time-scales. Analysing patient letters lets discharge or follow-up information update records automatically. Three approaches are compared: lexicon-based phrase identification, word-frequency analysis, and supervised text mining, evaluated by precision and stakeholder acceptability.

Based on: A Comparison · Texas medical journal

Curated by Aramai Editorial

Read summary →
Highlight

YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information

Introduces Programmable Gradient Information and the GELAN architecture to combat information loss in deep networks for object detection.

The paper addresses information loss as data passes layer-by-layer through deep networks, framed via the information bottleneck and reversible functions. It proposes Programmable Gradient Information (PGI), providing complete input information and reliable gradients for weight updates, plus GELAN, a lightweight architecture based on gradient path planning. On MS COCO detection, GELAN with conventional convolution beats SOTA depth-wise designs on parameter utilization, and PGI lets train-from-scratch models outperform models pre-trained on large datasets.

Based on: YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information · European Conference on Computer Vision

Curated by Aramai Editorial

Read summary →
Highlight

Scene Parsing through ADE20K Dataset

Introduces the ADE20K dataset and benchmark for scene parsing, plus a Cascade Segmentation Module network.

Scene parsing, recognizing and segmenting objects and stuff in an image, is a key vision problem lacking datasets with wide coverage and dense annotations. The paper introduces and analyzes ADE20K, which densely annotates scenes, objects, object parts, and sometimes parts of parts. A benchmark of 150 object and stuff classes is built on it, and segmentation baselines are evaluated. A novel Cascade Segmentation Module parses a scene into stuff, objects, and parts in cascade, improving over baselines, and enables applications like image content removal and scene synthesis.

Based on: Scene Parsing through ADE20K Dataset · Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →
Highlight

Statistical Phrase-Based Translation

Proposes a phrase-based statistical translation model and decoder, and analyzes why phrase-based methods outperform word-based ones.

The paper proposes a phrase-based translation model and decoding algorithm that lets several previously proposed phrase-based models be evaluated within one framework. Extensive experiments help explain why phrase-based models beat word-based ones. Across all examined language pairs, high performance comes from simple means: heuristic learning of phrase translations from word-based alignments and lexical weighting. Surprisingly, phrases longer than three words add little, while restricting to syntactically motivated phrases degrades performance.

Based on: Statistical Phrase-Based Translation · North American Chapter of the Association for Computational Linguistics

Curated by Aramai Editorial

Read summary →
Highlight

Discriminative topological features reveal biological network mechanisms

Presents a classification method to systematically determine which network generation model best describes a given biological network.

Many network models reproduce coarse features like degree distributions and clustering coefficients with indistinguishable fidelity despite different mechanisms, so such features cannot identify the best model. The authors present a method to assess which generation algorithm most accurately describes a given biological network by mapping all graphs into a high-dimensional word space used as input for classifiers. Training on networks from 17 models, they find duplication-mutation schemes best describe the E. coli genetic, S. cerevisiae protein-interaction, and C. elegans neuronal networks.

Based on: Discriminative topological features reveal biological network mechanisms · BMC Bioinformatics

Curated by Aramai Editorial

Read summary →
Highlight

Unsupervised Feature Learning via Non-parametric Instance Discrimination

Introduces instance discrimination, an unsupervised method treating each image as its own class to learn discriminative visual features.

The paper asks whether good feature representations can be learned by making features discriminative of individual instances rather than classes. It frames this as a non-parametric instance-level classification problem, using noise-contrastive estimation to handle the huge number of instance classes. Under unsupervised settings, the method surpasses prior state-of-the-art on ImageNet by a large margin and improves with more data and better architectures. The learned features also transfer competitively to semi-supervised learning and object detection, while staying highly compact.

Based on: Unsupervised Feature Learning via Non-parametric Instance Discrimination · 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →
Highlight

The Virtual Community: Homesteading on the Electronic Frontier

A book review of Howard Rheingold's The Virtual Community, assessing its account of the social aspects of computer networks.

This is a review of Howard Rheingolds 1993 book The Virtual Community: Homesteading on the Electronic Frontier. Rheingold, a self-described uncredentialed social scientist rooted in 1960s activism, organizes the book around reviews of computer networks: a Bay Area case study, a history of the Internet, groupmind systems such as Usenet, MUDs, and IRC, networks in Japan, France, and England, and political-organizing applications. The reviewer values the enduring historical and theoretical sections, but notes it predates the World Wide Web.

Based on: The Virtual Community: Homesteading on the Electronic Frontier · Semantic Scholar

Curated by Aramai Editorial

Read summary →
Highlight

Self-Attentive Sequential Recommendation

Introduces SASRec, a self-attention sequential recommender that captures long-term semantics from relatively few recent actions.

Sequential recommenders capture the context of users recent activities, typically via Markov Chains (MCs), which excel in sparse data, or RNNs, which suit denser data with longer-term semantics. SASRec, a self-attention model, balances these: it captures long-term semantics like an RNN but, using attention, bases predictions on relatively few actions like an MC. At each step it identifies which past items are relevant to predict the next item. Experiments show SASRec outperforms state-of-the-art MC/CNN/RNN models on sparse and dense datasets and is far more efficient.

Based on: Self-Attentive Sequential Recommendation · Industrial Conference on Data Mining

Curated by Aramai Editorial

Read summary →
Highlight

Normalization and variance stabilization of single-cell RNA-seq data using regularized negative binomial regression

Presents sctransform, using regularized negative binomial regression to normalize and stabilize variance in single-cell RNA-seq data.

Single-cell RNA-seq data show large cell-to-cell variation from technical factors such as detected molecule counts, which can confound biological heterogeneity. The authors present a framework for normalization and variance stabilization using Pearson residuals from regularized negative binomial regression, with sequencing depth as a covariate in a generalized linear model. Because an unconstrained model can overfit, they pool information across genes of similar abundance for stable estimates. The method preserves biology and is available in the R package sctransform.

Based on: Normalization and variance stabilization of single-cell RNA-seq data using regularized negative binomial regression · Genome Biology

Curated by Aramai Editorial

Read summary →
Highlight

Data Analysis

A book covering methods for finding relevant data dimensions and clustering, and links between data mining and data analysis.

The first part of this book covers methods for finding relevant dimensions of data, producing synthetic descriptions that often lead to graphical representations, following a general presentation of discriminating analysis. The second part turns to clustering methods, an approach often complementary to the first, used to synthesize and analyze data. The book concludes by examining the links that exist between data mining and data analysis.

Based on: Data Analysis · Data Science Journal

Curated by Aramai Editorial

Read summary →