Computer vision models can now explain their predictions more accurately, thanks to a new method from MIT that extracts concepts directly from what a model has already learned. MIT News AI reports that researchers developed a technique to convert any pretrained vision model into one that can articulate its reasoning in plain language.
The core problem is straightforward: in fields like medical diagnostics, people need to know why an AI made a specific call. Concept bottleneck models (CBMs) were designed for exactly this. They force AI to identify human-readable concepts (like “clustered brown dots” in a skin lesion) before making a final prediction. But traditional CBMs rely on concepts defined in advance by humans or large language models, and those concepts often miss the mark for a specific task.
What the Researchers Did Differently
Instead of feeding the model pre-defined concepts, the MIT team flipped the approach. They extract the concepts the model has already learned during training and translate them into language humans can understand.
The pipeline works in two steps:
- A sparse autoencoder pulls out the most relevant features the model learned and reconstructs them into a small set of concepts
- A multimodal LLM then describes each concept in plain language and annotates images in the dataset, flagging which concepts appear in each image
Those annotated images train a concept bottleneck module that gets plugged back into the original model. The model is then forced to make predictions using only those extracted concepts, with a hard cap of five concepts per prediction. That constraint does double duty: it keeps explanations concise and pushes the model to pick the most relevant ones.
Results
When tested against state-of-the-art CBMs on bird species identification and skin lesion diagnosis, the new method delivered:
- Highest accuracy among all concept bottleneck approaches tested
- More precise explanations tied directly to the task
- Better concept relevance to the actual images in the dataset
In a sense, we want to be able to read the minds of these computer vision models.
Why This Matters for Practitioners
Explainability isn’t a nice-to-have in healthcare, legal tech, or any domain where AI decisions carry real consequences. This method addresses two persistent issues with current explainable AI:
- Concept quality – pre-defined concepts often don’t match what the model actually uses, creating a gap between the explanation and reality
- Information leakage – even with defined concepts, models sometimes secretly rely on unintended features. Restricting predictions to five extracted concepts limits this
For teams deploying vision models in regulated industries, this approach offers a practical path: take your existing pretrained model and bolt on interpretability without starting from scratch.
Limitations Worth Noting
De Santis was candid about the tradeoffs. Non-interpretable black-box models still outperform this method in raw accuracy. The gap between explainability and peak performance hasn’t fully closed.
The team plans to tackle the information leakage problem further, possibly by stacking additional bottleneck modules. They also want to scale up by using larger multimodal LLMs to annotate bigger training datasets, which could push accuracy higher.
The research will be presented at the International Conference on Learning Representations (ICLR). What stands out here is the elegance of the core insight: stop telling the model what concepts to use and start asking it what it already knows. For more details, check the original coverage from MIT News AI.