Highlight

The Knowledge Complexity of Interactive Proof Systems

Develops a complexity theory of knowledge in proofs and defines zero-knowledge proofs, giving examples for quadratic residuosity.

A proof of a theorem usually conveys more knowledge than the bare fact that it is true. This paper develops a computational complexity theory of the knowledge contained in a proof, defining zero-knowledge proofs as those that convey nothing beyond the correctness of the proposition. It presents zero-knowledge proof systems for the languages of quadratic residuosity and quadratic nonresiduosity, the first such examples for languages not known to be efficiently recognizable.

Based on: The Knowledge Complexity of Interactive Proof Systems · SIAM journal on computing (Print)

Curated by Aramai Editorial

Read summary →
Highlight

Learning Patterns of Activity Using Real-Time Tracking

Presents a real-time visual tracker using adaptive Gaussian-mixture background subtraction that learns site activity patterns from motion.

The paper develops a passive visual monitoring system that observes moving objects at a site and learns activity patterns, focusing on motion tracking. Motion segmentation uses adaptive background subtraction modeling each pixel as a mixture of Gaussians updated online, giving a stable real-time outdoor tracker robust to lighting changes, clutter, and scene changes. Since object identity is consistent across a sequence, joint co-occurrence statistics are accumulated and used to build a hierarchical binary-tree classification of sequences and activities.

Based on: Learning Patterns of Activity Using Real-Time Tracking · IEEE Transactions on Pattern Analysis and Machine Intelligence

Curated by Aramai Editorial

Read summary →
Highlight

FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness

FlashAttention: an IO-aware exact attention algorithm using tiling to cut GPU memory traffic, speeding Transformer training and enabling longer context.

Self-attention scales quadratically with sequence length, making Transformers slow on long sequences, and approximate methods cut compute but often lose quality without speedups. The authors argue the missing principle is IO-awareness, counting reads and writes between GPU memory levels. FlashAttention is an IO-aware exact attention algorithm that uses tiling to cut transfers between GPU HBM and on-chip SRAM, needing fewer HBM accesses. It trains Transformers faster (15% on BERT-large, 3x on GPT-2, 2.4x on long-range arena) and enables longer context and new capabilities.

Based on: FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness · Neural Information Processing Systems

Curated by Aramai Editorial

Read summary →
Highlight

Colorful Image Colorization

Presents a fully automatic CNN that colorizes grayscale photos by posing colorization as a classification task with class rebalancing.

The paper tackles automatic colorization of grayscale photos, an underconstrained problem prior methods solved with heavy user input or desaturated output. The authors propose a fully automatic CNN giving vibrant, realistic colors by posing colorization as classification with class rebalancing. Run as a feed-forward pass and trained on over a million images, it fools humans on 32% of a colorization Turing test, beating prior methods. It also serves as a strong self-supervised pretext task, reaching state-of-the-art on feature-learning benchmarks.

Based on: Colorful Image Colorization · European Conference on Computer Vision

Curated by Aramai Editorial

Read summary →
Highlight

Phytozome: a comparative platform for green plant genomics

Introduces Phytozome, a comparative web platform and database for green plant genome and gene family data and analysis.

As sequenced plant genomes proliferate via next-generation sequencing, Phytozome provides a comparative hub for plant genome and gene family data. It shows each gene's evolutionary history at the level of sequence, gene structure, gene family and genome organization, while giving access to annotations for 25 complete plant genomes spanning land plants and selected algae. Via a web portal, it offers the plant science community comparative genomics tools that link model systems to economically and ecologically important plants.

Based on: Phytozome: a comparative platform for green plant genomics · Nucleic Acids Res.

Curated by Aramai Editorial

Read summary →
Highlight

Dimensionality reduction for visualizing single-cell data using UMAP

Applies UMAP dimensionality reduction to single-cell data and shows it outperforms five other tools in speed, reproducibility, and cluster organization.

Single-cell technologies enable high-resolution dissection of tissue composition but generate many parameters that require dimensionality-reduction tools. The authors apply uniform manifold approximation and projection (UMAP), a nonlinear technique developed for high-dimensional data, to biology using three mass cytometry and single-cell RNA sequencing datasets. Comparing UMAP with five other tools, they find it gives the fastest run times, highest reproducibility, and most meaningful organization of cell clusters for improved single-cell visualization.

Based on: Dimensionality reduction for visualizing single-cell data using UMAP · Nature Biotechnology

Curated by Aramai Editorial

Read summary →
Highlight

Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings

Shows word embeddings encode gender stereotypes and proposes a geometric method to remove them while preserving useful structure.

Word embeddings, even those trained on Google News, encode female/male gender stereotypes to a disturbing degree, and their widespread use tends to amplify these biases. Geometrically, gender bias is captured by a direction in the space, and gender-neutral words are linearly separable from gender-definitional words. The authors use these properties to modify embeddings, removing stereotypical associations like receptionist-female while keeping legitimate ones like queen-female. Evaluations show significant bias reduction while preserving clustering and analogy-solving ability.

Based on: Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings · Neural Information Processing Systems

Curated by Aramai Editorial

Read summary →
Highlight

Learning to Prompt for Vision-Language Models

Introduces CoOp, which replaces hand-crafted prompts for CLIP-like vision-language models with learnable context vectors for image recognition.

Large pre-trained vision-language models like CLIP align images and texts in a shared space, enabling zero-shot transfer via prompting, but prompt engineering is time-consuming and needs expertise since small wording changes greatly affect performance. Inspired by NLP prompt learning, the authors propose Context Optimization (CoOp), which models a prompt's context words as learnable vectors while keeping the pre-trained model frozen. Across 11 datasets, CoOp beats hand-crafted prompts with just one or two shots and gains around 15% on average with 16 shots.

Based on: Learning to Prompt for Vision-Language Models · International Journal of Computer Vision

Curated by Aramai Editorial

Read summary →
Highlight

BEiT: BERT Pre-Training of Image Transformers

Introduces BEiT, a self-supervised vision Transformer pretrained with a masked image modeling task inspired by BERT.

BEiT is a self-supervised vision model that adapts BERT-style pretraining to images via a masked image modeling task. Each image is treated as two views: 16x16 patches and discrete visual tokens from tokenization. Random patches are masked and fed to a backbone Transformer trained to recover the original visual tokens, and the encoder is then fine-tuned on downstream tasks. Base BEiT reaches 83.2% top-1 on ImageNet-1K, beating from-scratch DeiT, and large BEiT reaches 86.3% using only ImageNet-1K.

Based on: BEiT: BERT Pre-Training of Image Transformers · International Conference on Learning Representations

Curated by Aramai Editorial

Read summary →
Highlight

Learning Important Features Through Propagating Activation Differences

Introduces DeepLIFT, a method that explains neural network predictions by backpropagating each neuron's contribution relative to a reference activation.

Neural networks' black-box nature limits adoption where interpretability matters. DeepLIFT (Deep Learning Important FeaTures) decomposes a network's output for a specific input by backpropagating the contributions of all neurons to every input feature. It compares each neuron's activation to a reference activation and scores the difference, and separating positive and negative contributions reveals dependencies other methods miss. Scores compute in a single backward pass, and on MNIST and simulated genomic data it shows significant advantages over gradient-based methods.

Based on: Learning Important Features Through Propagating Activation Differences · International Conference on Machine Learning

Curated by Aramai Editorial

Read summary →
Highlight

GhostNet: More Features From Cheap Operations

Proposes the Ghost module, generating extra feature maps via cheap linear operations, to build the efficient GhostNet for embedded devices.

GhostNet targets the difficulty of deploying CNNs on embedded devices with limited memory and compute. Observing that feature-map redundancy is characteristic of successful CNNs, the authors propose a Ghost module that generates more feature maps from cheap operations: from a set of intrinsic feature maps, inexpensive linear transformations produce additional ghost features. Stacking these modules into Ghost bottlenecks forms the lightweight GhostNet, which reaches 75.7% top-1 accuracy on ImageNet, exceeding MobileNetV3 at similar cost.

Based on: GhostNet: More Features From Cheap Operations · Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →
Highlight

Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour

Shows how to train ImageNet ResNet-50 in one hour using large 8192 minibatches with a linear LR scaling rule and warmup scheme.

This paper tackles the long training times of deep networks by scaling distributed synchronous SGD across many parallel workers, which requires large minibatches. The authors show that large minibatches cause optimization difficulties on ImageNet, but once these are addressed, networks generalize with no loss of accuracy up to a minibatch size of 8192. Key techniques are a hyper-parameter-free linear scaling rule for learning rates and a new warmup scheme. Their Caffe2 system trains ResNet-50 in one hour on 256 GPUs with about 90% scaling efficiency.

Based on: Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour · arXiv.org

Curated by Aramai Editorial

Read summary →