Thursday 20 March 2025
Deep learning models have revolutionized many fields, but their inner workings remain shrouded in mystery for most people. A new approach aims to change that by providing concept-based explanations of a model’s predictions.
The researchers behind this effort have developed a method that identifies the key concepts or features that a deep neural network uses to make a prediction. This is particularly useful when dealing with complex medical diagnoses, like distinguishing between basal cell carcinoma and trichoblastoma based on skin images.
To understand how this works, let’s take a step back. Traditional explanation methods focus on feature attribution, which highlights the most important features of an input image that contribute to the model’s prediction. However, these approaches often fail to provide meaningful insights into what the model is actually looking at. The new method takes a different tack by using non-negative matrix factorization (NMF) to identify underlying concepts that are used to explain a model’s predictions.
The researchers tested their approach on several medical image classification tasks, including distinguishing between basal cell carcinoma and trichoblastoma. They found that the concept-based explanations provided valuable insights into what features of an image were most important for the model’s prediction. For example, in the case of basal cell carcinoma, the model was using concepts related to palisading cell edges with different orientations to make its predictions.
The researchers also demonstrated how their approach can be used to explain why a model is making a particular prediction by contrasting it with another class. In the case of basal cell carcinoma vs trichoblastoma, they found that the model was relying on different concepts for each class, including features related to white areas inside tumors and epidermis.
One of the most impressive aspects of this approach is its ability to provide explanations for single-class predictions as well. For example, the researchers showed how their method could be used to explain why a model predicts a particular image of a fireboat or gazelle with high confidence.
The implications of this work are far-reaching. By providing transparent and interpretable explanations of deep learning models, researchers can gain a better understanding of what features they are relying on and how they are making predictions. This could lead to more accurate and reliable diagnoses in medical imaging and other fields where AI is being used.
In addition, the ability to provide concept-based explanations for single-class predictions opens up new possibilities for analyzing and improving the performance of deep learning models.
Cite this article: “Unlocking the Black Box: Concept-Based Explanations for Deep Learning Models”, The Science Archive, 2025.
Deep Learning, Medical Imaging, Concept-Based Explanations, Non-Negative Matrix Factorization, Feature Attribution, Model Interpretation, Ai Transparency, Diagnosis Accuracy, Reliability, Neural Networks.
Reference: Rudolf Herdt, Daniel Otero Baguer, “Concept Based Explanations and Class Contrasting” (2025).
