[Paper Review] Visualizing and Understanding Convolutional Neural Networks

Matthew D. Zeiler, Rob Fergus|arXiv (Cornell University)|Nov 12, 2013

Anomaly Detection Techniques and Applications451 citations

TL;DR

This paper introduces a novel visualization technique to interpret intermediate layers in Convolutional Neural Networks (CNNs), revealing how features are learned and classified. Through ablation studies and architectural improvements, the authors achieve superior ImageNet performance and demonstrate strong generalization to Caltech-101 and Caltech-256 by retraining the softmax classifier.

ABSTRACT

Large Convolutional Neural Network models have recently demonstrated impressive classification performance on the ImageNet benchmark \cite{Kriz12}. However there is no clear understanding of why they perform so well, or how they might be improved. In this paper we address both issues. We introduce a novel visualization technique that gives insight into the function of intermediate feature layers and the operation of the classifier. We also perform an ablation study to discover the performance contribution from different model layers. This enables us to find model architectures that outperform Krizhevsky \etal on the ImageNet classification benchmark. We show our ImageNet model generalizes well to other datasets: when the softmax classifier is retrained, it convincingly beats the current state-of-the-art results on Caltech-101 and Caltech-256 datasets.

Motivation & Objective

• To develop a visualization method that provides insight into the function of intermediate feature layers in CNNs.
• To understand the contribution of individual layers to overall classification performance through ablation studies.
• To identify model architectures that outperform existing models on the ImageNet benchmark.
• To evaluate the generalization capability of the proposed model on other datasets such as Caltech-101 and Caltech-256.

Proposed method

• Proposes a novel visualization technique to interpret feature maps and filters in intermediate convolutional layers.
• Performs ablation studies by systematically removing or modifying layers to assess their contribution to classification accuracy.
• Uses the ImageNet dataset to train and evaluate multiple CNN architectures with varying depth and filter configurations.
• Retrains the final softmax classifier on Caltech-101 and Caltech-256 datasets to evaluate transfer performance.
• Analyzes feature representations learned by the network to understand how they support classification decisions.
• Employs visualization to interpret the operation of the classifier and the role of feature hierarchies in recognition.

Experimental results

Research questions

• RQ1 How do intermediate convolutional layers in deep CNNs contribute to image classification, and what features do they learn?
• RQ2 What is the relative performance contribution of different layers in a CNN architecture to final classification accuracy?
• RQ3 Can architectural modifications based on ablation analysis lead to improved performance on the ImageNet benchmark?
• RQ4 How well does a trained ImageNet model generalize to other image classification datasets like Caltech-101 and Caltech-256?

Key findings

• The proposed visualization technique reveals that early layers learn simple edge and texture features, while deeper layers learn more complex, semantically meaningful patterns.
• Ablation studies show that removing certain convolutional layers significantly degrades performance, highlighting their critical role in feature learning.
• The authors identify a modified CNN architecture that outperforms Krizhevsky et al.'s model on the ImageNet classification benchmark.
• When the softmax classifier is retrained on Caltech-101 and Caltech-256, the model achieves state-of-the-art performance on both datasets.
• The model demonstrates strong generalization, indicating that features learned on ImageNet are transferable to other visual recognition tasks.
• The visualization method provides actionable insights into how filters and feature maps contribute to classification decisions.