In this paper, we propose Describe-and-Dissect, a novel method to describe
the roles of hidden neurons in vision networks.
Describe-and-Dissect (DnD) utilizes recent advancements in multimodal deep learning to produce complex natural language
descriptions, without the need for labeled training data or a predefined set of concepts to choose from. Additionally,
Describe-and-Dissect is training-free, meaning we don’t train any new models and can easily leverage more capable general
purpose models in the future. We show on a large scale user study that our method outperforms the state-of-the-art baseline
methods including CLIP-Dissect, MILAN, and Network Dissection. Our method on average provides the highest
quality labels and is more than 2× as likely to be selected as the best explanation for a neuron than the best baseline.
Overview of Describe and Dissect neuron explanation pipeline. We highlight the three primary steps in the algorithm: 1) Probing Set Augmentation, 2) Candidate Concept Generation, and 3) Best Concept Selection.
Overview of Describe and Dissect pipeline.
Step 1: Probing Set Augmentation
Salient regions are cropped from a neuron's activating images and augmented to the original probing dataset.
Step 2: Candidate Concept Generation
An Image-to-Text model (BLIP) and an LLM (GPT)
generate and summarize neuron descriptions.
Step 3: Best Concept Selection
Neuron descriptions are fined-tuned on synthetic images to determine the best explanation.
We qualitatively analyze results of randomly selected neurons from various layers of ResNet-50, ResNet-18, and ViT-B-16.
Labels for each method are color coded by whether we believe they are accurate, somewhat correct,
or vague/imprecise. Compared to baseline models, we observe that DnD captures higher level concepts in a
more semantically coherent manner.
Below, we show examples from Layers 1 through 4 of ResNet-50 and present additional results in
Neuron Examples.
Example ResNet-50 DnD descriptions. We observe that DnD outperforms baseline models.
We follow CLIP-Dissect [1] to quantitatively analyze description quality on the last layer neurons, which have known ground truth labels (i.e. class name) to allow us to evaluate the quality of neuron descriptions automatically. Our results show that DnD outperforms MILAN [2], having a greater average CLIP cosine similarity by 0.0518, a greater average mpnet cosine similarity by 0.18, and a greater average BERTScore by 0.008.
Textual similarity between DnD/MILAN labels and ground truths on ResNet-50 (Imagenet). We can see DnD outperforms MILAN.
We also performed quantitative evaluation by calculating the textual similarity between a method's label
and the corresponding MILANNOTATIONS. Our analysis found that
if every neuron is described with the same constant concept:
'depictions', it will achieve better results than any
explanation method on this dataset, but this is not a useful
nor meaningful description. Thus, the dataset is unreliable to serve as ground
truths and can't be relied on for comparing different methods.
Textual similarity between descriptions produced by methods and MILANNOTATIONS. Simply labeling every neuron as ”depictions” outperforms all other methods, demonstrating the unreliability of MILANNOTATIONS as an evaluation method.
Our experiment compares the quality of labels produced by DnD against 3
baselines: CLIP-Dissect, MILAN, and Network Dissection [3]. For both models we evaluated
4 of the intermediate layers (end of each residual block),
with 200 randomly chosen neurons per layer for ResNet50
and 50 per layer for ResNet-18. Each neurons description is evaluated by 3 different workers.
We outline specifics of the experiment below:
Results for individual layers of ResNet-50. We observe that DnD is the best method across all layers in ResNet-50.
Results for individual layers of ResNet-18. DnD performs significatly better across every layer in ResNet-18 when compared to the baseline methods
To showcase a potential use case for neuron descriptions, we experimented with using neuron descriptions to find a good classifier for a class missing from the training set. We use neurons from Layer 4 of ResNet-50 (Imagenet) to find neurons in this layer that could serve as the best classifiers for an unseen class, specifically the classes in CIFAR-10 and CIFAR-100 datasets. Our setup is as follows:
The average classification AUC on out of distribution dataset when using neurons with similar description as a classifier. We can see that our DnD clearly outperforms MILAN, the only other generative description method.
For further insight, we conduct comprehensive ablation studies analyzing the importance of each step in the DnD pipeline. We perform the following experiments with additional details shown here:
[1] Oikarinen and Weng, "Clip-dissect: Automatic description of neuron representations in deep vision networks." International Conference on Learning Representations, 2023.
[2] Hernandez etal, "Natural language descriptions of deep visual features." International Conference on Learning Representations, 2022.
[3] Bau etal, "Network dissection: Quantifying interpretability of deep visual representations." Computer Vision and Pattern Recognition, 2017.
[4] Kalibhat etal, "Identifying Interpretable Subspaces in Image Representations." International Conference on Machine Learning, 2023.
@article{bai2025interpret,
title={Interpreting Neurons in Deep Vision Networks with Language Models},
author={Bai, Nicholas and Iyer, Rahul A and Oikarinen, Tuomas and Kulkarni, Akshay and Weng, Tsui-Wei},
journal={TMLR},
year={2025}
}