Imagenet 1k
PulseAugur coverage of Imagenet 1k — every cluster mentioning Imagenet 1k across labs, papers, and developer communities, ranked by signal.
8 day(s) with sentiment data
-
Architecture search methods break down at high compression, researchers find
Researchers have identified a critical limitation in blockwise distillation methods used for architecture search, particularly in models like DeiT-S and DeiT-B. These methods, which score individual replacement blocks a…
-
New distillation methods boost discrete diffusion model efficiency and quality
Researchers have developed new methods for distilling discrete diffusion models to improve efficiency and quality. One approach, Bidirectional Centered Score Matching (BCSM), offers an efficient objective that reduces c…
-
BDNet introduces fast-slow thinking for efficient image classification
Researchers have introduced BDNet, a novel hierarchical image-classification backbone designed for enhanced efficiency and robustness. This network utilizes a "fast-slow thinking" approach, employing a baseline deformat…
-
WithinVAE achieves 4x compression in visual autoencoders via self-distillation
Researchers have developed WithinVAE, a novel visual autoencoder that achieves high reconstruction fidelity and compactness without relying on external semantic supervision. This model utilizes a sequential bottleneck a…
-
New LocaLeJEPA framework improves self-supervised learning with geometric priors
Researchers have introduced LocaLeJEPA, a new framework for self-supervised learning that addresses representation collapse issues in Joint-Embedding Predictive Architectures (JEPA). Unlike previous methods tied to Eucl…
-
New training method enables efficient on-device training of binary and ternary AI models
Researchers have developed a new training method called \"$\Sigma\Delta$-Sampler\" that leverages Ising solvers to train binary and ternary AI models more efficiently. This approach embeds a stochastic sampling and anne…
-
TRACE framework enhances distributed Transformer inference on edge devices
Researchers have developed TRACE, a novel framework for communication-efficient distributed inference of Transformers on edge devices. TRACE utilizes a Vector Quantizer (VQ) in each Transformer layer to encode activatio…
-
New research explores adaptive visual perception for AI systems · 2 sources tracked
Two recent arXiv papers explore advancements in active visual perception, a field where AI systems dynamically control how they acquire visual evidence to improve recognition. The first paper provides a comprehensive re…
-
New variational filtering algorithm enhances deep classifier accuracy and calibration
Researchers have developed a new variational filtering algorithm designed to improve the accuracy and calibration of deep learning classifiers. This method treats the last layer's parameters as a state-space model, assi…
-
Vision Transformers Enhanced by Register Tokens Improve Image Reconstruction
Researchers have developed a method to improve image reconstruction and generation by incorporating register tokens, which encode global scene properties, into Vision Transformer (ViT) decoders. By jointly processing th…
-
New research explores vision-language models for image recognition
A new arXiv paper explores the capabilities of vision-language models (VLMs) for image recognition, comparing their performance in both language-guided and vision-only classification tasks. The research evaluates establ…
-
Recurrent Vision Transformer Achieves Full-Depth Encoder Accuracy with Single Block
Researchers have developed a novel recurrent Vision Transformer (reViT) that utilizes a single Transformer block applied recurrently to achieve accuracy comparable to full-depth vision encoders. This approach represents…
-
ViTok model enhances dense semantics in computer vision distillation
Researchers have developed ViTok, a new model that improves dense semantics in multi-teacher distillation for computer vision tasks. By combining insights from SigLIP2 and DINOv3-L, ViTok addresses the trade-off between…
-
New KCCA method enhances VLM visual perception
Researchers have revisited the visual representation enhancement of vision-language models (VLMs) by proposing a new method based on Kernel Canonical Correlation Analysis (KCCA). This approach characterizes representati…
-
New Spiking Self-Attention Method Reduces Train-Inference Mismatch
Researchers have developed a new method called Parallel Time-Aligned Spiking Self-Attention (PT-SSA) to address discrepancies in training and inference for spiking neural networks. This approach reconstructs virtual spi…
-
Function-Space Transformer Adapts Representation for Scientific and Visual Tasks
Researchers have introduced the Function-Space Transformer (FST), a novel framework designed to learn from functions represented by discrete samples. Unlike traditional methods that use fixed grids, FST employs a spatia…
-
VisionHOPE: New Visual Backbone Model Released with Pretrained Weights
A new research paper introduces VisionHOPE, a visual backbone model designed as a self-modifying learning system. The paper provides official pretrained weights for VisionHOPE checkpoints in various sizes (T/S/B). These…
-
Diffusion models advance with new architectures and training methods · 10 sources tracked
Researchers are exploring advanced diffusion models across various domains, including tri-modal generation for text, images, and audio by Apple, and data-free training methods for lattice field theories using diffusion …
-
New Fusion Method Merges Dissimilar Vision Models
Researchers have developed a novel method called Riemannian--Lorentz Parameter Fusion (RLPF) to merge independently trained vision models, even when their architectures differ. This technique addresses the challenges of…
-
New CTOAC method improves low-bit quantization for Visual State Space Duality models
Researchers have developed a new post-training quantization (PTQ) method called Channel-wise Token-balanced Output-Aware Clipping (CTOAC) to address the low-bit quantization challenges in Visual State Space Duality (VSS…