Highlight

Detecting Faces in Images: A Survey

Surveys and categorizes single-image face detection algorithms, discussing benchmarks, evaluation metrics, limitations, and future research directions.

Face images are central to intelligent human-computer interaction, but many methods assume faces are already localized. Given a single image, face detection aims to identify all regions containing a face regardless of 3D position, orientation, or lighting, a hard problem because faces are non-rigid and vary widely in size, shape, color, and texture. This survey categorizes and evaluates the many single-image face detection techniques and discusses data collection, evaluation metrics, and benchmarking, concluding with the algorithms' limitations and promising future directions.

Based on: Detecting Faces in Images: A Survey · IEEE Transactions on Pattern Analysis and Machine Intelligence

Curated by Aramai Editorial

Read summary →
Highlight

Practical Black-Box Attacks against Machine Learning

Demonstrates the first practical black-box adversarial attack on a remote DNN using only its output labels, via a locally trained substitute model.

DNNs and other ML models are vulnerable to adversarial examples: inputs subtly modified to cause misclassification while appearing unchanged to humans. Prior attacks required access to the model internals or training data, but this work presents the first practical attack needing only the labels the target assigns to chosen inputs. The adversary trains a local substitute on synthetic inputs labeled by the target, then crafts adversarial examples that transfer to it. Against models hosted by MetaMind, Amazon, and Google, misclassification reached 84.24%, 96.19%, and 88.94%.

Based on: Practical Black-Box Attacks against Machine Learning · ACM Asia Conference on Computer and Communications Security

Curated by Aramai Editorial

Read summary →
Highlight

Restormer: Efficient Transformer for High-Resolution Image Restoration

Proposes Restormer, an efficient Transformer for high-resolution image restoration that captures long-range pixel interactions without quadratic cost.

CNNs learn generalizable image priors well but have limited receptive fields and cannot adapt to input content, while Transformers fix these issues yet scale quadratically with resolution, making them impractical for high-resolution restoration. Restormer is an efficient Transformer that redesigns the multi-head attention and feed-forward blocks to capture long-range pixel interactions while remaining applicable to large images. It reaches state-of-the-art results on deraining, motion deblurring, defocus deblurring, and Gaussian and real image denoising.

Based on: Restormer: Efficient Transformer for High-Resolution Image Restoration · Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →
Highlight

HuggingFace's Transformers: State-of-the-art Natural Language Processing

Presents Transformers, an open-source library offering unified access to state-of-the-art Transformer architectures and curated pretrained models.

Recent NLP progress has been driven by advances in Transformer architectures, which enable higher-capacity models, and by pretraining, which lets that capacity be used across many tasks. Transformers is an open-source library that brings these advances to the broader machine learning community, providing state-of-the-art Transformer architectures under a unified API together with a curated collection of community-contributed pretrained models. It is designed to be extensible for researchers, simple for practitioners, and fast and robust in industrial deployment.

Based on: HuggingFace's Transformers: State-of-the-art Natural Language Processing · arXiv.org

Curated by Aramai Editorial

Read summary →
Highlight

An Introduction to Convolutional Neural Networks

A brief introductory overview of convolutional neural networks and recent techniques for image-driven pattern recognition, aimed at readers new to CNNs.

This document offers a brief introduction to Convolutional Neural Networks (CNNs), a biologically inspired class of artificial neural network that has far exceeded prior AI methods on common machine learning tasks. CNNs are used mainly for difficult image-driven pattern recognition, offering a precise yet simple architecture that eases getting started with ANNs. The paper discusses recently published work and newly developed techniques for building these image recognition models, assuming familiarity with ANN and machine learning fundamentals.

Based on: An Introduction to Convolutional Neural Networks · arXiv.org

Curated by Aramai Editorial

Read summary →
Highlight

Masked-attention Mask Transformer for Universal Image Segmentation

Introduces Mask2Former, a single architecture using masked attention to address panoptic, instance, and semantic segmentation tasks.

Image segmentation tasks (panoptic, instance, semantic) differ only in semantics, yet are typically tackled with specialized architectures. Mask2Former is a single architecture able to address any of these tasks. Its key component is masked attention, which extracts localized features by restricting cross-attention to predicted mask regions. Beyond cutting research effort by at least three times, it outperforms the best specialized models, setting new state-of-the-art results: 57.8 PQ on COCO panoptic, 50.1 AP on COCO instance, and 57.7 mIoU on ADE20K semantic segmentation.

Based on: Masked-attention Mask Transformer for Universal Image Segmentation · Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →
Highlight

A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play

Presents AlphaZero, a single reinforcement learning algorithm that masters chess, shogi, and Go from self-play with no domain knowledge beyond rules.

The authors generalize the self-play reinforcement learning of AlphaGo Zero into a single algorithm, AlphaZero, applicable to many games. Unlike top chess programs that rely on sophisticated search, domain-specific tuning, and human-crafted evaluation functions, AlphaZero starts from random play with only the rules provided. Learning purely through self-play, it achieved superhuman performance and convincingly defeated world-champion programs in chess, shogi, and Go, marking a step toward general game-playing systems.

Based on: A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play · Science

Curated by Aramai Editorial

Read summary →
Highlight

SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition

Introduces SpecAugment, a simple data augmentation applied to speech feature inputs via time warping and masking of frequency and time blocks.

SpecAugment is a simple augmentation method applied directly to a neural network's filter bank feature inputs. Its policy warps the features and masks blocks of frequency channels and time steps. Applied to Listen, Attend and Spell networks for end-to-end recognition, it reaches state-of-the-art on LibriSpeech 960h and Switchboard 300h. On LibriSpeech test-other it hits 6.8% WER without a language model and 5.8% with shallow fusion, beating the prior 7.5% hybrid system.

Based on: SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition · Interspeech

Curated by Aramai Editorial

Read summary →
Highlight

Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators

Introduces DeepONet, a deep operator network that learns nonlinear operators, grounded in the universal approximation theorem of operators.

A neural network with a single hidden layer can approximate not just continuous functions but any nonlinear continuous operator, a result the authors extend to deep networks. They design DeepONet, a deep operator network with small generalization error, built from a branch net that encodes the discrete input function space and a trunk net that encodes the output domain. DeepONet learns explicit operators such as integrals and fractional Laplacians and implicit operators representing deterministic and stochastic differential equations, studied across 16 diverse applications.

Based on: Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators · Nature Machine Intelligence

Curated by Aramai Editorial

Read summary →
Highlight

Machine Learning in Medicine

A review introducing basic machine learning concepts for medicine and examining why ML has had limited impact on clinical care.

Advances in computing power and data have enabled machines to master complex tasks from poker to video games, and analytic companies increasingly turn to healthcare. This review explores which medical problems might benefit from machine learning and uses literature examples to introduce basic concepts. Although suitable datasets and algorithms have existed for decades and thousands of papers apply ML to medical data, very few have improved clinical care. The author identifies obstacles to changing medical practice through statistical learning and how to overcome them.

Based on: Machine Learning in Medicine · Mach. Learn. under Resour. Constraints Vol. 3

Curated by Aramai Editorial

Read summary →
Highlight

Improved Baselines with Momentum Contrastive Learning

Shows that adding an MLP head and stronger augmentation to MoCo yields improved contrastive learning baselines that surpass SimCLR.

Contrastive unsupervised learning has progressed with methods like MoCo and SimCLR. This note verifies two of SimCLR's design improvements by implementing them in the MoCo framework: an MLP projection head and stronger data augmentation. With these simple modifications, the authors establish stronger baselines that outperform SimCLR while not requiring large training batches, aiming to make state-of-the-art unsupervised learning research more accessible. Code will be released.

Based on: Improved Baselines with Momentum Contrastive Learning · arXiv.org

Curated by Aramai Editorial

Read summary →
Highlight

Recurrent Models of Visual Attention

Introduces a recurrent visual attention model that adaptively selects image regions to process, trained end-to-end with reinforcement learning.

Applying convolutional networks to large images is costly because computation scales with pixel count. This paper presents a recurrent neural network that extracts information from an image or video by adaptively selecting a sequence of regions and processing only those at high resolution. It has built-in translation invariance, yet its computation is controlled independently of input size. Being non-differentiable, it is trained via reinforcement learning; it beats a CNN baseline on cluttered image classification and learns to track an object without explicit supervision.

Based on: Recurrent Models of Visual Attention · Neural Information Processing Systems

Curated by Aramai Editorial

Read summary →