ViT-B/16
PulseAugur coverage of ViT-B/16 — every cluster mentioning ViT-B/16 across labs, papers, and developer communities, ranked by signal.
- instance of Vision Transformers 90%
- used by ResNet-50 70%
- instance of ResNet-50 70%
- used by CIFAR-100 70%
- instance of ViT-L/16 70%
- instance of CIFAR-100 70%
- used by ImageNet ILSVRC-2012 70%
- used by ResNet-18 70%
- used by Vision Transformers 70%
- instance of machine unlearning 70%
- used by Imagenet 1k 70%
- competes with ResNet-50 60%
8 day(s) with sentiment data
-
DISCERN method enhances incremental machine unlearning for transformers
Researchers have developed DISCERN, a novel method for incremental machine unlearning that addresses the challenge of sequentially removing data classes from a trained model while preserving remaining knowledge. DISCERN…
-
Astronomical image classifiers' accuracy influenced by context, not just morphology
A new study published on arXiv evaluates the performance of small-data astronomical image classifiers, revealing that factors beyond object morphology significantly influence accuracy. Researchers developed a dataset to…
-
Federated AI for Breast Imaging Faces Confounding and Privacy Challenges
Researchers explored a privacy-aware federated learning framework for breast cancer imaging using vision-language and vision-only models. They found that while models achieved high discrimination for normal versus sick …
-
New methods reveal functional content of CLIP fine-tuning updates
Researchers have developed new methods to understand the functional content of fine-tuning updates in CLIP models. By employing techniques like reversion and retention, they analyzed how specific weight changes contribu…
-
Recursive CLIP models explore parameter efficiency via recurrent blocks
Researchers have developed new vision-language models, rCLIP and r^2CLIP, that utilize recurrent transformer blocks instead of independently parameterized ones. This approach increases parameter sharing while maintainin…
-
LoRA Superposition technique boosts model accuracy by 30% · arXiv paper
Researchers have introduced a novel technique called LoRA superposition, which allows for the summation of LoRA (Low-Rank Adaptation) models trained independently on different pretrained backbones for the same task. Thi…
-
New GM-GReFEL Model Enhances Facial Expression Recognition
Researchers have introduced GM-GReFEL, a novel spatiotemporal state-space network designed for dynamic facial expression recognition. This model addresses challenges like quadratic computational costs with clip length a…
-
New foundation models and adaptation techniques for brain MRI research unveiled
Two new research papers explore the adaptation of foundation models for medical imaging, specifically focusing on structural brain MRI. The first paper introduces Brain2Vec, a self-supervised 3D foundation model pre-tra…
-
New ALPS visual encoder unifies prediction, language, and pixel tasks for world models
Researchers have introduced ALPS, a novel visual state encoder designed to serve as a foundational component for world models. Unlike previous encoders that specialize in prediction, language alignment, or pixel reconst…
-
New SAGA method mitigates catastrophic forgetting in AI models
Researchers have developed a new method called Sparse Autoencoder Geometric Anchoring (SAGA) to combat catastrophic forgetting in continual learning for AI models. SAGA uses sparse autoencoders trained at task boundarie…
-
New TADrop method improves model merging with adaptive sparsification
Researchers have developed TADrop, a novel method for model merging that addresses the limitations of uniform sparsity ratios. Unlike previous approaches, TADrop determines tensor-specific keep ratios based on magnitude…
-
New defense method shields ML models from black-box query attacks
Researchers have developed a new defense mechanism called Gradient Confusion Defense (GCD) to protect machine learning models from black-box score-based query attacks. This method, which does not require retraining the …
-
New method certifies detection of adversarial AI attacks
Researchers have developed a new method for detecting adversarial attacks on deep learning classifiers, focusing on certifying that malicious inputs are flagged rather than ensuring model robustness. This approach lever…
-
New WoodBR-Cross dataset improves AI wood species identification accuracy
Researchers have introduced WoodBR-Cross, a new dataset designed to improve the accuracy of wood species identification in images. This dataset contains over 4,800 images across 38 species, crucially linking images to i…
-
New ComCLIP Framework Enhances CLIP Vision Encoder for VLMs
Researchers have proposed ComCLIP, a new framework for refining CLIP's vision encoder, which is foundational for models like LLaVA. Contrary to previous findings, they demonstrate that contrastive post-training can be e…
-
New research explores optimal adapter placement in vision transformers
Researchers have explored two distinct methods for optimizing the placement of task-specific adapters in continual learning models, specifically within vision transformers. One approach, placement search, involves train…
-
New SRTS-BCE method improves AI classifier calibration with limited data
Researchers have developed a new method called Signal-Routed Temperature Scaling (SRTS-BCE) to improve the calibration of AI classifiers, particularly when limited validation data is available. This technique introduces…
-
New method learns functional subspaces for efficient neural network compression
Researchers have developed a new method called Learnable Subspace Projections (LSP) to compress large neural networks, particularly transformers and LLMs. Unlike previous techniques that use local criteria, LSP optimize…
-
New LSP method enhances neural network compression by learning subspaces end-to-end
Researchers have developed a new method called Learnable Subspace Projections (LSP) for compressing neural networks, particularly transformers. Unlike previous techniques that used local criteria, LSP learns the subspac…
-
Weight Pair Encoding reduces neural network grammar size
Researchers have developed a novel technique called Weight Pair Encoding (WeightPE) that aims to reduce the size of neural network weights by inducing a smaller grammar. This method incorporates a Re-Pair compressor wit…