Highlight

Zero-Shot Text-to-Image Generation

Models text and image tokens as one autoregressive transformer stream for zero-shot text-to-image generation competitive with domain-specific models.

Text-to-image generation has traditionally relied on better modeling assumptions for a fixed dataset, sometimes via complex architectures, auxiliary losses, or side information like object part labels or segmentation masks during training. This paper describes a simple alternative: a transformer that autoregressively models text and image tokens together as a single stream of data. Given sufficient data and scale, the approach is competitive with prior domain-specific models when evaluated in a zero-shot setting.

Based on: Zero-Shot Text-to-Image Generation · International Conference on Machine Learning

Curated by Aramai Editorial

Read summary →
Highlight

DeepFace: Closing the Gap to Human-Level Performance in Face Verification

DeepFace uses 3D alignment and a nine-layer, 120M-parameter deep network to reach 97.35% on LFW face verification, nearly matching human performance.

Face recognition typically runs a pipeline of detect, align, represent, and classify. DeepFace revisits alignment and representation, using explicit 3D face modeling for a piecewise affine transformation and a nine-layer deep network. The network has over 120 million parameters and uses locally connected layers without weight sharing rather than standard convolutions, trained on four million images of over 4,000 identities. It reaches 97.35% accuracy on LFW, cutting the prior state-of-the-art error by over 27% and approaching human performance.

Based on: DeepFace: Closing the Gap to Human-Level Performance in Face Verification · 2014 IEEE Conference on Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →
Highlight

Show and tell: A neural image caption generator

A neural image caption generator pairing computer vision with a deep recurrent language model, trained to maximize likelihood, setting new BLEU scores.

Automatically describing an image connects computer vision and natural language processing. This paper presents a generative model with a deep recurrent architecture that draws on advances in computer vision and machine translation to produce natural sentences describing an image, trained to maximize the likelihood of the target description. It is accurate and fluent, learning language solely from image descriptions. Its BLEU-1 on Pascal reaches 59 versus a prior 25 (human ~69), with gains on Flickr30k (56 to 66) and SBU (19 to 28) and a BLEU-4 of 27.7 on COCO.

Based on: Show and tell: A neural image caption generator · Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →
Highlight

Visual Genome: Connecting Language and Vision Using Crowdsourced Dense Image Annotations

Presents Visual Genome, a dataset of 108K+ images with object, attribute, and relationship annotations canonicalized to WordNet for visual reasoning.

Computers do well on perceptual tasks like image classification but poorly on cognitive tasks such as description and question answering, which require reasoning about object relationships, not just recognizing them. The authors present Visual Genome to enable modeling of these interactions. It contains over 108K images, each averaging 35 objects, 26 attributes, and 21 relationships, with objects, attributes, relationships, and noun phrases canonicalized to WordNet synsets, forming a dense, large dataset of descriptions, objects, attributes, relationships, and QA pairs.

Based on: Visual Genome: Connecting Language and Vision Using Crowdsourced Dense Image Annotations · International Journal of Computer Vision

Curated by Aramai Editorial

Read summary →
Highlight

Supervised Contrastive Learning

Proposes a supervised contrastive loss that leverages label information to outperform cross-entropy on image classification and improve robustness.

Cross-entropy is the dominant loss for supervised image classification, but this paper proposes a training method that outperforms it across architectures and augmentations. It adapts the self-supervised batch contrastive loss to supervised training, pulling same-class embeddings together while pushing different classes apart, using large batches and normalized embeddings. On ResNet-50 and ResNet-200 it beats cross-entropy by over 1%, setting a new state of the art of 78.8% with AutoAugment, and also improves robustness and hyperparameter stability.

Based on: Supervised Contrastive Learning · Neural Information Processing Systems

Curated by Aramai Editorial

Read summary →
Highlight

Massive MIMO for next generation wireless systems

Overview of massive MIMO, which uses a large excess of base-station antennas over terminals with TDD to boost wireless throughput and energy efficiency.

Multi-user MIMO improves on point-to-point MIMO by using cheap single-antenna terminals, not requiring rich scattering, and simplifying resource allocation, but in its original form it does not scale. Massive MIMO deploys far more service antennas than active terminals with time-division duplex operation. The extra antennas focus energy into smaller regions, greatly improving throughput and energy efficiency while enabling low-power components, lower latency, a simpler MAC layer, and robustness to jamming. The article surveys the concept and open challenges.

Based on: Massive MIMO for next generation wireless systems · IEEE Communications Magazine

Curated by Aramai Editorial

Read summary →
Highlight

Categorical Reparameterization with Gumbel-Softmax

Introduces Gumbel-Softmax, a differentiable reparameterization enabling gradient-based training through categorical latent variables in neural networks.

Categorical variables naturally represent discrete structure, but stochastic neural networks rarely use categorical latents because one cannot backpropagate through samples. This work presents an efficient gradient estimator that replaces the non-differentiable categorical sample with a differentiable sample from a novel Gumbel-Softmax distribution, which anneals smoothly toward a categorical distribution. It beats state-of-the-art gradient estimators on structured output prediction and unsupervised generative modeling with categorical latents, and speeds up semi-supervised classification.

Based on: Categorical Reparameterization with Gumbel-Softmax · International Conference on Learning Representations

Curated by Aramai Editorial

Read summary →
Highlight

Tabu Search

A book presenting tabu search, a meta-heuristic for optimization problems, illustrated with applications from scheduling and VLSI design to logistics.

This book presents tabu search, a meta-heuristic that broadens the ability to solve difficult problems across resource planning, telecommunications, VLSI design, scheduling, logistics, pattern classification, and manufacturing, among many domains. The major ideas are explained through examples tied to real applications, with numerous illustrations and diagrams clarifying key principles. The aim is to convey hands-on knowledge and insight rather than only computational recipes or abstract themes, while also addressing advanced issues.

Based on: Tabu Search · Semantic Scholar

Curated by Aramai Editorial

Read summary →
Highlight

The STRING database in 2021: customizable protein–protein networks, and functional characterization of user-uploaded gene/measurement sets

The 2021 STRING update adds customizable protein–protein networks and functional characterization of user-uploaded gene or measurement sets.

STRING integrates known and predicted associations between proteins, spanning physical interactions and functional associations. It scores evidence from literature text-mining, interaction-experiment and curated pathway databases, predictions from co-expression and conserved genomic context, and cross-organism transfer. Version 11.5 will cover over 14,000 organisms, and adds a reworked text-mining system, a new scoring mode for physical interactions, interface features for customizing and sharing networks, and querying with genome-wide data plus automated enrichment and bias detection.

Based on: The STRING database in 2021: customizable protein–protein networks, and functional characterization of user-uploaded gene/measurement sets · Nucleic Acids Res.

Curated by Aramai Editorial

Read summary →
Highlight

Classifier-Free Diffusion Guidance

Trades diffusion sample quality against diversity without a separate classifier by combining joint conditional and unconditional score estimates.

Classifier guidance adjusts the trade-off between mode coverage and sample fidelity in conditional diffusion models after training, mixing a diffusion model's score estimate with the gradient of a separately trained image classifier. This paper shows the same guidance works without any classifier: a single generative model jointly trains conditional and unconditional diffusion models, then combines their score estimates. This classifier-free guidance yields a quality-diversity trade-off comparable to classifier guidance.

Based on: Classifier-Free Diffusion Guidance · arXiv.org

Curated by Aramai Editorial

Read summary →
Highlight

Inference and analysis of cell-cell communication using CellChat

CellChat infers and analyzes cell–cell communication networks from single-cell RNA-seq data using a curated ligand–receptor interaction database.

Understanding cell communication requires accurate signaling links and systems-level analysis. The authors build a database of ligand, receptor, and cofactor interactions capturing known heteromeric complexes, then develop CellChat to quantitatively infer intercellular communication networks from single-cell RNA-seq data. Via network analysis, pattern recognition, and manifold learning, it predicts signaling inputs and outputs, classifies pathways, and separates conserved from context-specific signaling. Applied to mouse and human skin, it extracts complex patterns and offers a web Explorer.

Based on: Inference and analysis of cell-cell communication using CellChat · Nature Communications

Curated by Aramai Editorial

Read summary →
Highlight

Deformable Convolutional Networks

Introduces deformable convolution and deformable RoI pooling, adding learnable spatial offsets to boost CNNs' geometric transformation modeling.

CNNs are limited in modeling geometric transformations because of the fixed geometric structures in their modules. This work adds two modules, deformable convolution and deformable RoI pooling, that augment spatial sampling locations with additional offsets learned from the target task without extra supervision. They drop into existing CNNs and train end-to-end via back-propagation. Experiments show learning dense spatial transformation helps object detection and semantic segmentation.

Based on: Deformable Convolutional Networks · IEEE International Conference on Computer Vision

Curated by Aramai Editorial

Read summary →