CLIP-Dissect: Automatic Description of Neuron Representations in Deep Vision Networks

TL;DR

CLIP-Dissect delivers a training-free, CLIP-based framework to automatically label hidden neurons with open-ended concepts, decoupling probe data from concept vocabularies for flexible, scalable interpretation. It demonstrates superior qualitative and quantitative performance over prior methods, including the ability to label concepts absent from the probing dataset and to adapt to future multimodal models. The approach achieves major speedups (e.g., labeling five ResNet-50 layers in ~4 minutes) and reveals that high-weight connections correlate with similar concepts across neurons, offering new insights into network representations. These contributions advance practical interpretability of deep vision models and provide a scalable tool for model auditing and bias discovery.

Abstract

In this paper, we propose CLIP-Dissect, a new technique to automatically describe the function of individual hidden neurons inside vision networks. CLIP-Dissect leverages recent advances in multimodal vision/language models to label internal neurons with open-ended concepts without the need for any labeled data or human examples. We show that CLIP-Dissect provides more accurate descriptions than existing methods for last layer neurons where the ground-truth is available as well as qualitatively good descriptions for hidden layer neurons. In addition, our method is very flexible: it is model agnostic, can easily handle new concepts and can be extended to take advantage of better multimodal models in the future. Finally CLIP-Dissect is computationally efficient and can label all neurons from five layers of ResNet-50 in just 4 minutes, which is more than 10 times faster than existing methods. Our code is available at https://github.com/Trustworthy-ML-Lab/CLIP-dissect. Finally, crowdsourced user study results are available at Appendix B to further support the effectiveness of our method.

Figures (22)

• Figure 1: Labels generated by our method CLIP-Dissect, NetDissect netdissect2017 and MILAN hernandez2022natural for random neurons of ResNet-50 trained on ImageNet. Displayed together with 5 most highly activating images for that neuron. We have subjectively colored the descriptions green if they match these 5 images, yellow if they match but are too generic and red if they do not match. In this paper we follow the torchvisiontorchvision naming scheme of ResNet: Layer 4 is the second to last layer and Layer 1 is the end of first residual block. MILAN(b) is trained on both ImageNet and Places365 networks, while MILAN(p) is only trained on Places365.
• Figure 2: Overview of CLIP-Dissect: a 3-step algorithm to dissect neural network of interest
• Figure 3: Example of a final layer neuron: we compare the descriptions generated by different methods and our metrics. Accuracy only evaluated for CLIP-Dissect with ImageNet labels as concept set since it is the only method where exact correct answer is a possible choice and therefore accuracy makes sense.
• Figure 4: Example of CLIP-Dissect correctly labeling neurons that detect the little blue heron and the great white heron based on pictures of dolphins and dinosaurs in CIFAR. CIFAR100 does not contain any bird images but CLIP-Dissect can still get correct concept.
• Figure 5: a) 3 highest weights of the final layer of ResNet-50 trained on ImageNet, we can see neurons connected by the highest weights are detecting very much the same concept. b) Cosine similarities between the concepts of neurons connected by highest weights. The higher the weight between neurons, the more similar a concept they represent.