Highlight

FlowNet: Learning Optical Flow with Convolutional Networks

Proposes FlowNet, convolutional networks that learn optical flow end-to-end as supervised learning, trained on the synthetic Flying Chairs dataset.

CNNs had excelled at recognition but not at optical flow estimation. This paper builds CNNs that solve optical flow as a supervised learning task, proposing and comparing two architectures: a generic one and another with a correlation layer that matches feature vectors across image locations. Because existing ground-truth datasets are too small to train a CNN, the authors generate a large synthetic Flying Chairs dataset. Networks trained on this unrealistic data generalize well to Sintel and KITTI, achieving competitive accuracy at 5 to 10 frames per second.

Based on: FlowNet: Learning Optical Flow with Convolutional Networks · IEEE International Conference on Computer Vision

Curated by Aramai Editorial

Read summary →
Highlight

A Survey of Large Language Models

A systematic survey of large language models across pre-training, post-training, utilization, and evaluation, plus open research challenges.

This survey systematically reviews advances in large language models across four dimensions: pre-training that builds core capabilities via large-scale self-supervised training and data curation; post-training through supervised fine-tuning and reinforcement learning for adaptation, alignment, and safety; utilization such as in-context learning, prompt engineering, and agentic reasoning; and evaluation via benchmarks for language, reasoning, and safety. It also flags open issues in theory, efficient scaling, alignment, and agentic capability.

Based on: A Survey of Large Language Models · Frontiers Comput. Sci.

Curated by Aramai Editorial

Read summary →
Highlight

SWISS-MODEL: modelling protein tertiary and quaternary structure using evolutionary information

Presents the latest SWISS-MODEL automated server for homology modelling of protein tertiary and quaternary structure using evolutionary information.

SWISS-MODEL is a fully automated web server that generates 3D protein models by homology when experimental structures are unavailable. This latest version adds a template library annotating quaternary structure, ligands, and cofactors to build complete oligomeric models, and uses model quality estimation to select templates and predict expected accuracy, with results continuously benchmarked by CAMEO. A new interface lets users search, cluster, and structurally compare templates, and guide selection to build models in different functional states.

Based on: SWISS-MODEL: modelling protein tertiary and quaternary structure using evolutionary information · Nucleic Acids Res.

Curated by Aramai Editorial

Read summary →
Highlight

Generative Agents: Interactive Simulacra of Human Behavior

Introduces generative agents: LLM-driven software agents that simulate believable human behavior in an interactive sandbox town.

Generative agents are computational agents that simulate believable human behavior—following routines, forming opinions, conversing, and reflecting. The architecture extends an LLM to store experiences in natural language, synthesize them into higher-level reflections, and retrieve them to plan actions. Instantiated in a Sims-inspired sandbox of 25 agents, they yield believable individual and emergent social behaviors, e.g., autonomously organizing a Valentine's Day party. Ablations confirm observation, planning, and reflection each contribute critically.

Based on: Generative Agents: Interactive Simulacra of Human Behavior · ACM Symposium on User Interface Software and Technology

Curated by Aramai Editorial

Read summary →
Highlight

Machine Learning: Algorithms, Real-World Applications and Research Directions

A comprehensive survey of machine learning algorithm types, their real-world application domains, and open challenges and research directions.

This paper gives a comprehensive overview of machine learning algorithms and their real-world applications for the Fourth Industrial Revolution, where abundant IoT, cybersecurity, mobile, business, social, and health data demand intelligent analysis. It explains the main learning types, supervised, unsupervised, semi-supervised, reinforcement, and deep learning, and their applicability across domains such as cybersecurity, smart cities, healthcare, e-commerce, and agriculture, while highlighting key challenges and future research directions.

Based on: Machine Learning: Algorithms, Real-World Applications and Research Directions · SN Computer Science

Curated by Aramai Editorial

Read summary →
Highlight

The COG database: an updated version includes eukaryotes

Reports an updated COG database of orthologous protein clusters, adding KOGs for seven eukaryotic genomes for comparative genomics and annotation.

This paper describes a major update to the Clusters of Orthologous Groups (COGs) system, which classifies genes from sequenced prokaryotic and unicellular eukaryotic genomes by orthology for comparative genomics and functional annotation. It adds orthology clusters for seven eukaryotic genomes, named KOGs. The COG set holds 138,458 proteins in 4873 COGs (~75% of proteins in 66 genomes); the KOG set has 4852 clusters covering 59,838 proteins (~54%), with about 20% forming a conserved core across all species.

Based on: The COG database: an updated version includes eukaryotes · BMC Bioinformatics

Curated by Aramai Editorial

Read summary →
Highlight

Toolformer: Language Models Can Teach Themselves to Use Tools

Introduces Toolformer, a language model that self-supervises when and how to call external tool APIs to improve zero-shot task performance.

Toolformer shows that language models can teach themselves to use external tools via simple APIs, uniting LM strengths with skills like arithmetic and factual lookup where smaller models excel. It learns which APIs to call, when, what arguments to pass, and how to fold results into token prediction, self-supervised from only a handful of demonstrations per API. Tools include a calculator, Q&A, two search engines, a translator, and a calendar. It substantially improves zero-shot performance across many tasks, often rivaling larger models, without hurting language modeling.

Based on: Toolformer: Language Models Can Teach Themselves to Use Tools · Neural Information Processing Systems

Curated by Aramai Editorial

Read summary →
Highlight

EfficientNetV2: Smaller Models and Faster Training

Introduces EfficientNetV2, convolutional networks from training-aware NAS and scaling that train faster and use parameters more efficiently.

EfficientNetV2 is a family of convolutional networks with faster training and better parameter efficiency than prior models, developed using training-aware neural architecture search and scaling over a space enriched with Fused-MBConv. Progressive image-size increase speeds training further, with adaptively adjusted regularization to offset the accuracy drop. The models are up to 6.8x smaller; pretrained on ImageNet21k, EfficientNetV2 reaches 87.3% top-1 on ImageNet, beating a recent ViT by 2.0% while training 5x-11x faster.

Based on: EfficientNetV2: Smaller Models and Faster Training · International Conference on Machine Learning

Curated by Aramai Editorial

Read summary →
Highlight

LabelMe: A Database and Web-Based Tool for Image Annotation

Presents LabelMe, a web-based image annotation tool and a large labeled dataset built with it for object detection and recognition research.

LabelMe is a web-based tool for building a large collection of images with ground-truth labels for object detection and recognition research. It makes annotation easy and enables instant sharing, and was used to collect a large dataset spanning many object categories, often with multiple instances per image. The authors quantify the dataset and compare it to existing state-of-the-art recognition and detection datasets. They also extend it via WordNet to enrich labels, discover object parts, recover depth ordering, and add labels with minimal supervision.

Based on: LabelMe: A Database and Web-Based Tool for Image Annotation · International Journal of Computer Vision

Curated by Aramai Editorial

Read summary →
Highlight

InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets

Introduces InfoGAN, a GAN that learns disentangled representations unsupervised by maximizing mutual information between latent codes and observations.

InfoGAN is an information-theoretic extension of the GAN that learns disentangled representations in a fully unsupervised manner. It maximizes the mutual information between a small subset of latent variables and the observation, deriving an efficiently optimizable lower bound and interpreting training as a variant of the Wake-Sleep algorithm. It disentangles writing style from digit shape on MNIST, pose from lighting on 3D images, and foreground from background digits on SVHN, and finds concepts like hairstyles and eyeglasses on CelebA, rivaling supervised methods.

Based on: InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets · Neural Information Processing Systems

Curated by Aramai Editorial

Read summary →
Highlight

ViLBERT: Pretraining Task-Agnostic Visiolinguistic Representations for Vision-and-Language Tasks

Presents ViLBERT, a two-stream BERT extension that learns task-agnostic joint vision-and-language representations via co-attentional transformers.

ViLBERT (Vision-and-Language BERT) learns task-agnostic joint representations of images and natural language. It extends BERT into a two-stream model that processes visual and textual inputs separately, letting them interact through co-attentional transformer layers. Pretrained on two proxy tasks over the Conceptual Captions dataset, it transfers with minor additions to visual question answering, visual commonsense reasoning, referring expressions, and caption-based image retrieval, reaching state of the art on all four.

Based on: ViLBERT: Pretraining Task-Agnostic Visiolinguistic Representations for Vision-and-Language Tasks · Neural Information Processing Systems

Curated by Aramai Editorial

Read summary →
Highlight

VoxelNet: End-to-End Learning for Point Cloud Based 3D Object Detection

Proposes VoxelNet, a single-stage end-to-end network that detects 3D objects from raw lidar point clouds without hand-crafted features.

VoxelNet is a generic, end-to-end trainable 3D detection network that removes hand-crafted point-cloud features by unifying feature extraction and bounding box prediction in one stage. It divides a point cloud into equally spaced 3D voxels and uses a new voxel feature encoding (VFE) layer to convert the points in each voxel into a unified representation that feeds a region proposal network. On the KITTI car benchmark it outperforms prior lidar-based methods by a large margin and gives encouraging results for pedestrians and cyclists using lidar alone.

Based on: VoxelNet: End-to-End Learning for Point Cloud Based 3D Object Detection · 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →