REVIEW 24 cited by
Quantifying Attention Flow in Transformers
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
read the original abstract
In the Transformer model, "self-attention" combines information from attended embeddings into the representation of the focal embedding in the next layer. Thus, across layers of the Transformer, information originating from different tokens gets increasingly mixed. This makes attention weights unreliable as explanations probes. In this paper, we consider the problem of quantifying this flow of information through self-attention. We propose two methods for approximating the attention to input tokens given attention weights, attention rollout and attention flow, as post hoc methods when we use attention weights as the relative relevance of the input tokens. We show that these methods give complementary views on the flow of information, and compared to raw attention, both yield higher correlations with importance scores of input tokens obtained using an ablation method and input gradients.
Forward citations
Cited by 24 Pith papers
-
Transient Reserves, Sink Dampers, and the Failure of Eigenvalue Reasoning in the Attention Propagator
Resolvent analysis of trained causal attention shows sinks act as transient dampers, routing heads carry excess Kreiss reserve, and eigenvalue depth predictions fail by 7–11 orders of magnitude.
-
AGNFormer I: Reconstruction of AGN spectra using a probabilistic transformer model
An uncertainty-aware transformer reconstructs masked AGN broad lines and spectral halves with 4-16% flux errors and beats eleven purpose-built Lyα-reconstruction algorithms on a blind benchmark.
-
Feature-level Interaction Explanations in Multimodal Transformers
FL-I2MoE separates unique, synergistic, and redundant cross-modal evidence at the token/patch level and uses SII and redundancy-gap scores to rank pairs whose removal degrades performance more than random masking.
-
Full-Frequency Temporal Patching and Structured Masking for Enhanced Audio Classification
Replacing square spectrogram patches with full-frequency temporal patches plus patch-aligned masking improves audio classification accuracy and reduces compute for Transformer and Mamba models.
-
DSS-Prompt: Dynamic-Static Synergistic Prompting for Few-Shot Class-Incremental Learning
DSS-Prompt combines static prompts with instance-aware dynamic prompts generated from BLIP multi-modal features to achieve state-of-the-art few-shot class-incremental learning on four benchmarks without incremental training.
-
Towards White-Box Deep Wireless Sensing
RF-CRATE derives a fully complex-valued white-box transformer for RF sensing from the sparse rate reduction principle and shows it matches black-box baselines across five datasets.
-
Self-Guided Masked Autoencoder
A Masked Autoencoder that masks the object cluster found by its own early patch-clustering signal learns better representations than random masking, with no external labels or models.
-
CytoSAE: Interpretable Cell Embeddings for Hematology
CytoSAE learns sparse, expert-validated morphological concepts from blood-cell images that generalize across datasets and can classify AML subtypes at patient level with F1 0.83.
-
VIP: Visual Information Protection through Adversarial Attacks on Vision-Language Models
A perturbation computed from early attention and value matrices can make LLaVA, Instruct-BLIP, and BLIP2-T5 fail to detect objects inside a specified image region while keeping the rest of the image usable.
-
Foveation-Guided Dynamic Token Selection for Robust and Efficient Vision Transformers
FDT adds foveation and binary fixation modules to DeiT so multi-scale tokens are selected dynamically in one pass, improving ImageNet100 accuracy, MACs, and robustness without adversarial training.
-
From Features to Actions: Explainability in Traditional and Agentic AI Systems
Attribution explanations that work for static classifiers do not diagnose failures in multi-step AI agents; trace-grounded rubric evaluation does, with state-tracking inconsistency 2.7x more common in failed agent runs.
-
Robust Representation Learning in Masked Autoencoders
Masked Autoencoders build class-separable representations across depth and keep their embeddings directionally stable under blur and occlusion, which tracks their robust classification.
-
Revisiting 2D Foundation Models for Scalable 3D Medical Image Classification
A frozen 2D vision foundation model with lightweight LoRA adapters and attention-based slice fusion achieves state-of-the-art 3D medical image classification across 12 tasks with about 1M trainable parameters per task.
-
An Autoencoder and Vision Transformer-based Interpretability Analysis of the Differences in Automated Staging of Second and Third Molars
An autoencoder-plus-ViT pipeline improves dental staging accuracy and uses latent space, reconstructions, and attention maps to attribute the weaker third-molar performance to high intra-class data variability.
-
Attention Maps in 3D Shape Classification for Dental Stage Estimation with Class Node Graph Attention Networks
CGAT, a graph attention network with a CLS node, achieves 0.76 weighted F1 on Demirjian stage classification of 3D third-molar meshes and generates attention maps that highlight roots and furcation regions.
-
Attention of a Kiss: Exploring Attention Maps in Video Diffusion for XAIxArts
A method and case study for visualizing cross-attention maps in Wan video diffusion transformers, showing token-region alignment over time and their use as artistic material.
-
Decoding the Multimodal Maze: A Systematic Review on the Adoption of Explainability in Multimodal Attention-based Models
A systematic review of 55 papers finds explainability for multimodal attention-based models is dominated by attention-weight visualizations, while evaluation remains mostly qualitative and non-standardized.
-
User Experience Estimation in Human-Robot Interaction Via Multi-Instance Learning of Multimodal Social Signals
A multimodal Transformer with multi-instance learning estimates user experience (UX) questionnaire ratings from facial expressions and voice during human-robot interaction, reporting accuracy above third-party human raters.
-
Fair-FLIP: Fair Deepfake Detection with Fairness-Oriented Final Layer Input Prioritising
Fair-FLIP improves fairness parity in deepfake detection by reweighting final-layer features based on between-ethnicity variance, with negligible accuracy loss.
-
Safer Skin Lesion Classification with Global Class Activation Probability Map Evaluation and SafeML
A pixel-level argmax over per-class Grad-CAM maps, combined with a selective predictor, is proposed to detect unreliable skin lesion classifications.
-
Generalizable Federated Learning using Client Adaptive Focal Modulation
The abstract describes AdaptFED, a claimed federated learning method, but the full text is an unrelated graph theory paper, so the claimed results are absent from the submission.
-
PiPViT: Patch-based Visual Interpretable Prototypes for Retinal Image Analysis
PiPViT combines vision transformers and prototype learning to classify retinal OCT scans while showing the spatial extent of the biomarker that drove the decision.
-
Learning from Limited and Imperfect Data
A doctoral thesis compiling nine peer-reviewed papers on long-tailed image generation, long-tailed recognition, semi-supervised learning, and domain adaptation.
-
The Hitchhiker's Guide to Agentic AI: From Foundations to Systems
A survey-style reference book mapping the full agentic-AI stack from transformer internals to production deployment, with no new research result.
Discussion (0). Sign in to comment.