Highlight

Scalable Diffusion Models with Transformers

Introduces Diffusion Transformers (DiT): latent diffusion models that swap the U-Net backbone for a transformer, setting state-of-the-art ImageNet FID.

This paper introduces Diffusion Transformers (DiTs), diffusion models built on the transformer architecture. Training latent diffusion models of images, the authors replace the usual U-Net backbone with a transformer operating on latent patches, and study scalability via forward-pass complexity in Gflops. DiTs with higher Gflops—via greater depth, width, or more tokens—consistently achieve lower FID. The largest, DiT-XL/2, beats all prior diffusion models on class-conditional ImageNet 512×512 and 256×256, reaching a state-of-the-art FID of 2.27 on the latter.

Based on: Scalable Diffusion Models with Transformers · IEEE International Conference on Computer Vision

Curated by Aramai Editorial

Read summary →
Highlight

The entity-relationship model: toward a unified view of data

Proposes the entity-relationship model, a semantic data model with a diagrammatic technique unifying views of data for database design.

This paper proposes the entity-relationship model, a data model that incorporates the semantic information of the real world by describing data in terms of entities and the relationships among them. A special diagrammatic technique is introduced to depict these entities and relationships, and the paper gives a worked example of database design and description using both the model and the diagrams. The author discusses the model's implications for data integrity, information retrieval, and data manipulation, working toward a unified view of data.

Based on: The entity-relationship model: toward a unified view of data · Very Large Data Bases Conference

Curated by Aramai Editorial

Read summary →
Highlight

ShapeNet: An Information-Rich 3D Model Repository

Presents ShapeNet, a large-scale, richly annotated repository of 3D CAD models organized under the WordNet taxonomy as a benchmark for graphics and vision.

ShapeNet is a large-scale, richly annotated repository of 3D shapes represented as CAD models spanning many semantic categories and organized under the WordNet taxonomy. Each model carries annotations such as rigid alignments, parts, bilateral symmetry planes, physical sizes, and keywords, exposed via a public web interface for visualization and data-driven geometric analysis. It is meant as a large-scale quantitative benchmark for graphics and vision research. At the report's time, ShapeNet indexed over 3,000,000 models, 220,000 of them classified into 3,135 categories.

Based on: ShapeNet: An Information-Rich 3D Model Repository · arXiv.org

Curated by Aramai Editorial

Read summary →
Highlight

The STRING database in 2017: quality-controlled protein–protein association networks, made broadly accessible

The 2017 STRING update integrates known and predicted protein–protein associations across organisms, with a redesigned interface and Cytoscape access.

STRING integrates known and predicted protein–protein associations for many organisms, spanning direct physical interactions and indirect but specific functional ones. Beyond curated experiments, pathways, and complexes, it derives predictions from co-expression, shared genomic signals, literature text-mining, and cross-organism transfer via orthology. Version 10.5 emphasizes dissemination: a redesigned web frontend reduces reliance on outdated browsers, the database can be queried from within Cytoscape, and user inputs get automated enrichment analysis.

Based on: The STRING database in 2017: quality-controlled protein–protein association networks, made broadly accessible · Nucleic Acids Res.

Curated by Aramai Editorial

Read summary →
Highlight

VQA: Visual Question Answering

Proposes the task of free-form, open-ended Visual Question Answering and releases a large dataset of images, questions, and answers with baselines.

This paper proposes free-form, open-ended Visual Question Answering (VQA): given an image and a natural language question, a system must produce an accurate natural language answer. Because questions can target background details and context, VQA demands finer image understanding and more complex reasoning than generic captioning. The task supports automatic evaluation, since answers are often short or multiple-choice. The authors release a dataset of ~0.25M images, 0.76M questions, and 10M answers, and compare baselines with human performance.

Based on: VQA: Visual Question Answering · International Journal of Computer Vision

Curated by Aramai Editorial

Read summary →
Highlight

SILVA: a comprehensive online resource for quality checked and aligned ribosomal RNA sequence data compatible with ARB

SILVA, an online resource of quality-checked, aligned rRNA sequence databases for Bacteria, Archaea, and Eukarya, compatible with ARB software.

rRNA gene sequencing is the standard method for phylogenetic reconstruction and quantifying microbial diversity, but rapidly growing public data has strained curated databases. SILVA offers a central web resource of up-to-date, quality-controlled, aligned rRNA sequences from all three domains, each checked for anomalies and enriched with contextual and taxonomic information. Two ARB-compatible datasets are provided: near-full-length Ref sets for phylogenetics and comprehensive Parc sets for biodiversity analysis. Release 91 held 547,521 sequences.

Based on: SILVA: a comprehensive online resource for quality checked and aligned ribosomal RNA sequence data compatible with ARB · Nucleic Acids Research

Curated by Aramai Editorial

Read summary →
Highlight

Robust Speech Recognition via Large-Scale Weak Supervision

Shows that predicting 680,000 hours of internet transcripts yields speech models that generalize zero-shot and approach human accuracy and robustness.

The authors study speech processing systems trained simply to predict large amounts of internet audio transcripts. Scaled to 680,000 hours of multilingual and multitask supervision, the resulting models generalize well to standard benchmarks and are often competitive with prior fully supervised results in a zero-shot transfer setting, with no fine-tuning required. Compared to humans, the models approach their accuracy and robustness. The authors release the models and inference code as a foundation for further work on robust speech processing.

Based on: Robust Speech Recognition via Large-Scale Weak Supervision · International Conference on Machine Learning

Curated by Aramai Editorial

Read summary →
Highlight

Dynamic Graph CNN for Learning on Point Clouds

Introduces EdgeConv, a differentiable module operating on dynamically computed graphs to learn local and global features on point clouds.

Point clouds are a flexible geometric representation and the raw output of most 3D acquisition devices, but they lack topological information. The authors propose EdgeConv, a module for CNN-based classification and segmentation on point clouds. EdgeConv operates on graphs dynamically computed at each layer, is differentiable, and plugs into existing architectures. It captures local neighborhoods, can be stacked to learn global shape, and its feature-space affinity captures long-range semantics. Evaluated on ModelNet40, ShapeNetPart, and S3DIS.

Based on: Dynamic Graph CNN for Learning on Point Clouds · ACM Transactions on Graphics

Curated by Aramai Editorial

Read summary →
Highlight

ReAct: Synergizing Reasoning and Acting in Language Models

Introduces ReAct, prompting LLMs to interleave reasoning traces and task actions so they plan, handle exceptions, and query external sources.

LLMs are strong at reasoning and acting, but the two are usually studied separately. ReAct prompts a model to interleave reasoning traces and task actions, so reasoning helps track and update plans and handle exceptions while actions gather information from external sources like knowledge bases or environments. On HotpotQA and Fever, using a simple Wikipedia API reduces the hallucination and error propagation of chain-of-thought. On ALFWorld and WebShop, ReAct beats imitation and RL baselines by 34% and 10% absolute success with 1-2 in-context examples.

Based on: ReAct: Synergizing Reasoning and Acting in Language Models · International Conference on Learning Representations

Curated by Aramai Editorial

Read summary →
Highlight

User's guide to PHREEQC (Version 2)-a computer program for speciation, batch-reaction, one-dimensional transport, and inverse geochemical calculations

User's guide to PHREEQC v2, a C program for aqueous speciation, batch-reaction, 1D transport, and inverse geochemical modeling.

PHREEQC version 2 is a C program for low-temperature aqueous geochemical calculations, based on an ion-association aqueous model. It performs speciation and saturation-index calculations; batch-reaction and one-dimensional transport with reversible and irreversible reactions; and inverse modeling that finds mineral and gas mole transfers explaining compositional differences between waters. Version 2 adds dispersion and stagnant zones, user-defined kinetic rates, solid solutions, and isotope mole balances. The report presents the governing equations, input, and examples.

Based on: User's guide to PHREEQC (Version 2)-a computer program for speciation, batch-reaction, one-dimensional transport, and inverse geochemical calculations · Semantic Scholar

Curated by Aramai Editorial

Read summary →
Highlight

On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜

An examination of ever-larger language models, weighing their risks and recommending cost-aware, well-documented, stakeholder-driven alternatives.

The paper steps back from three years of ever-larger English language models such as BERT, GPT-2/3, and Switch-C, which pushed benchmark state of the art through architecture and sheer size. It asks how big is too big and what risks the technology poses, plus paths to mitigate them. The authors recommend weighing environmental and financial costs first, curating and documenting datasets rather than ingesting everything on the web, running pre-development checks of fit with research goals and stakeholder values, and pursuing directions beyond ever-larger models.

Based on: On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜 · Conference on Fairness, Accountability and Transparency

Curated by Aramai Editorial

Read summary →
Highlight

Path Aggregation Network for Instance Segmentation

Proposes PANet, which boosts information flow in proposal-based instance segmentation via bottom-up path augmentation and adaptive feature pooling.

Path Aggregation Network (PANet) improves information flow in proposal-based instance segmentation. Bottom-up path augmentation enriches the feature hierarchy with accurate localization signals from lower layers, shortening the path between low and top features, while adaptive feature pooling links each proposal to all feature levels. A complementary branch captures different views per proposal to improve masks. These additions add little overhead yet took 1st in the COCO 2017 instance segmentation and 2nd in detection, and are state of the art on MVD and Cityscapes.

Based on: Path Aggregation Network for Instance Segmentation · 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →