Source-linked AI summary
Deep Features Analysis with Attention Networks
Shipeng Xie, Da Chen, Rong Zhang, Hui Xue
TL;DR
Because neural networks are difficult to interpret, this paper visualizes internal features with heatmaps and relates attention setups to classification accuracy. It reports that shared attention yields less-noisy salient features, while early-stage attention gives the best attention-based heatmap and image-classification performance.
Problem
Neural networks are difficult to interpret, leaving the reasons for their predictions and strong performance hard to understand.
Source: p. 1Method
The paper visualizes internal feature representations with heatmaps and compares attention configurations and their corresponding image-classification accuracy.
Source: pp. 1–2Results
The authors report that early-stage attention gives the best attention-based heatmap and image-classification performance, while the proposed shared-attention method produces salient features with less noise.
Source: p. 2, Conclusion and Discussion · p. 3Takeaways & Limitations
Heatmaps provide a way to assess whether a network focuses on salient image regions and to compare attention-focused with original features.
Source: p. 2Future work includes exploiting attention loss and reducing the method’s FLOPs.
Source: Conclusion and Discussion · p. 3
Abstract
from arXiv · showhide
Deep neural network models have recently draw lots of attention, as it consistently produce impressive results in many computer vision tasks such as image classification, object detection, etc. However, interpreting such model and show the reason why it performs quite well becomes a challenging question. In this paper, we propose a novel method to interpret the neural network models with attention mechanism. Inspired by the heatmap visualization, we analyze the relation between classification accuracy with the attention based heatmap. An improved attention based method is also included and illustrate that a better classifier can be interpreted by the attention based heatmap.