Highlight

DARTS: Differentiable Architecture Search

Introduces DARTS, a differentiable neural architecture search that relaxes the search space to be continuous, enabling efficient gradient-based search.

DARTS tackles the scalability of neural architecture search by casting the task in a differentiable form. Instead of evolution or reinforcement learning over a discrete, non-differentiable space, it uses a continuous relaxation of the architecture representation so search runs efficiently via gradient descent. Experiments on CIFAR-10, ImageNet, Penn Treebank, and WikiText-2 show it finds high-performance convolutional architectures for image classification and recurrent ones for language modeling, orders of magnitude faster than non-differentiable methods.

Based on: DARTS: Differentiable Architecture Search · International Conference on Learning Representations

Curated by Aramai Editorial

Read summary →
Highlight

SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis

Presents SDXL, a latent diffusion text-to-image model with a 3x larger UNet, dual text encoders, and a refinement model for higher-fidelity synthesis.

SDXL is a latent diffusion model for text-to-image synthesis that scales up Stable Diffusion with a UNet backbone three times larger, driven by more attention blocks, a larger cross-attention context, and a second text encoder. It adds novel conditioning schemes, training on multiple aspect ratios, and a refinement model that boosts visual fidelity via a post-hoc image-to-image step. SDXL substantially outperforms prior Stable Diffusion versions and rivals black-box state-of-the-art generators. Code and model weights are released openly.

Based on: SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis · International Conference on Learning Representations

Curated by Aramai Editorial

Read summary →
Highlight

HEALPix: A Framework for High-Resolution Discretization and Fast Analysis of Data Distributed on the Sphere

HEALPix: a hierarchical equal-area, isolatitude pixelization of the sphere with algorithms for fast analysis of large astronomical data.

HEALPix, the Hierarchical Equal Area isoLatitude Pixelization, is a versatile structure for pixelizing data on the sphere. It ships with algorithms and visualization software for fast scientific work on discretized spherical maps from large astronomical datasets. Originally built for cosmic microwave background experiments such as BOOMERANG and WMAP, it can extend to future missions like Planck and Herschel. The paper details constraints for efficient discretization with hierarchical indexation and fast analysis/synthesis of functions on the sphere.

Based on: HEALPix: A Framework for High-Resolution Discretization and Fast Analysis of Data Distributed on the Sphere · arXiv

Curated by Aramai Editorial

Read summary →
Highlight

Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge

Introduces ARC, a grade-school science question set, corpus, and baselines requiring far more knowledge and reasoning than SQuAD or SNLI.

ARC assembles a question set, text corpus, and baselines to advance research in question answering. Its natural grade-school science questions split into a Challenge Set and an Easy Set; the Challenge Set holds only questions that both a retrieval-based and a word co-occurrence method answer wrong. With 7,787 questions, it is the largest public set of its kind. Leading neural models from SQuAD and SNLI fail to beat a random baseline, showing the task's difficulty. The authors also release a 14M-sentence science corpus and baseline implementations.

Based on: Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge · arXiv.org

Curated by Aramai Editorial

Read summary →
Highlight

MatchIt: Nonparametric Preprocessing for Parametric Causal Inference

Presents MatchIt, an R package that preprocesses data with nonparametric matching to make parametric causal inference more robust.

MatchIt implements the suggestions of Ho, Imai, King and Stuart (2007) for improving parametric statistical models by preprocessing data with nonparametric matching methods. Its wide range of matching methods greatly reduces the dependence of causal inferences on hard-to-justify but common modeling assumptions. It fits existing research practices: after preprocessing, researchers use whatever parametric model they would have used, but obtain inferences that are more robust and less sensitive to those assumptions. MatchIt is an R program that works seamlessly with Zelig.

Based on: MatchIt: Nonparametric Preprocessing for Parametric Causal Inference · Semantic Scholar

Curated by Aramai Editorial

Read summary →
Highlight

Places: A 10 Million Image Database for Scene Recognition

Presents Places, a 10-million-image scene database, with baseline Places-CNNs for scene recognition that outperform prior approaches.

The Places Database is a repository of 10 million scene photographs labeled with scene semantic categories spanning a large, diverse list of environments encountered in the world. Using state-of-the-art CNNs, the authors provide scene classification baselines, the Places-CNNs, which significantly outperform previous approaches. Visualization of the trained CNNs shows that object detectors emerge as an intermediate representation of scene classification. Its high coverage and diversity make Places and the Places-CNNs a resource to guide scene recognition progress.

Based on: Places: A 10 Million Image Database for Scene Recognition · IEEE Transactions on Pattern Analysis and Machine Intelligence

Curated by Aramai Editorial

Read summary →
Highlight

GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models

Explores text-conditional diffusion models for image synthesis and editing, comparing CLIP guidance with classifier-free guidance.

Diffusion models generate high-quality synthetic images, especially with guidance that trades diversity for fidelity. They study text-conditional image synthesis and compare two guidance strategies: CLIP guidance and classifier-free guidance. Human evaluators prefer classifier-free guidance for photorealism and caption similarity. Samples from their 3.5B-parameter model are favored over DALL-E even when it uses costly CLIP reranking. Models can also be fine-tuned for inpainting, enabling text-driven editing; code and weights for a smaller filtered model are released.

Based on: GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models · International Conference on Machine Learning

Curated by Aramai Editorial

Read summary →
Highlight

Hierarchical Attention Networks for Document Classification

Proposes a hierarchical attention network with word- and sentence-level attention that mirrors document structure for text classification.

The authors propose a hierarchical attention network for document classification with two distinctive characteristics: a hierarchical structure that mirrors the structure of documents, and attention mechanisms applied at both the word and sentence levels, letting it attend differentially to important content when building the document representation. On six large-scale text classification tasks it outperforms previous methods by a substantial margin, and visualization of the attention layers illustrates that the model selects informative words and sentences.

Based on: Hierarchical Attention Networks for Document Classification · North American Chapter of the Association for Computational Linguistics

Curated by Aramai Editorial

Read summary →
Highlight

Natural Questions: A Benchmark for Question Answering Research

Presents Natural Questions, a QA benchmark of real Google queries with Wikipedia long and short answer annotations, plus metrics and baselines.

Natural Questions is a question answering corpus built from real anonymized, aggregated queries issued to Google search. For each question, an annotator sees a top-5 Wikipedia page and marks a long answer (a paragraph) and a short answer (entities), or null if none is present. The public release has 307,373 single-annotation training examples, 7,830 5-way development and 7,842 5-way test examples. Experiments validate data quality, a 25-way study on 302 examples reveals human variability, and the paper adds robust metrics with high human upper bounds and baselines.

Based on: Natural Questions: A Benchmark for Question Answering Research · Transactions of the Association for Computational Linguistics

Curated by Aramai Editorial

Read summary →
Highlight

PubChem Substance and Compound databases

Overviews PubChem's Substance and Compound databases: data sources, organization, standardization, search interfaces, and programmatic access.

PubChem is a public repository of chemical substances and their bioactivities, launched in 2004 under the NIH Molecular Libraries Roadmap. It comprises three inter-linked databases: Substance, holding contributor-deposited chemical information; Compound, storing unique structures extracted from Substance; and BioAssay, holding activity data from assay experiments. It reviews the Substance and Compound databases, covering data sources and contents, organization, submission, standardization, search interfaces, and programmatic access, plus PubChem3D and PubChemRDF.

Based on: PubChem Substance and Compound databases · Nucleic Acids Res.

Curated by Aramai Editorial

Read summary →
Highlight

iCaRL: Incremental Classifier and Representation Learning

Introduces iCaRL, a class-incremental training strategy that jointly learns classifiers and representations as new classes arrive over time.

The authors introduce iCaRL, a training strategy for class-incremental learning where a system learns more and more classes over time from a data stream. Only data for a small number of classes need be present at once, and new classes are added progressively. Unlike earlier methods limited to fixed representations, iCaRL learns strong classifiers and a data representation simultaneously, making it compatible with deep architectures. On CIFAR-100 and ImageNet ILSVRC 2012 it learns many classes incrementally over a long period where other strategies quickly fail.

Based on: iCaRL: Incremental Classifier and Representation Learning · Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →
Highlight

KEGG for integration and interpretation of large-scale molecular data sets

Reports KEGG Mapper and knowledge-base extensions that integrate and interpret large-scale genomic, chemical, and functional molecular data.

KEGG is a database resource that integrates genomic, chemical and systemic functional information, linking gene catalogs from sequenced genomes to higher-level functions of the cell, organism and ecosystem. Systemic functions are captured as KEGG pathway maps, BRITE hierarchies and modules, with cross-species annotation via the KEGG Orthology system. This update reports KEGG Mapper, tools for PATHWAY, BRITE and MODULE mapping, and a variant that extends the knowledge base with disease genes and drug targets to support translational bioinformatics.

Based on: KEGG for integration and interpretation of large-scale molecular data sets · Nucleic Acids Res.

Curated by Aramai Editorial

Read summary →