ExplainabilityTransparencyAccountabilitycs.LG
Automatic Concept Embedding Model (ACEM): No train-time concepts, No issue!
Rishabh Jain
September 7, 2023
arXiv.org
Key Takeaways
-
Concept-based explanations make AI more understandable: The paper confirms that explaining AI decisions through human-like concepts (rather than raw data features) helps people understand how models work, which is especially important for safety-critical applications like healthcare or autonomous vehicles.
-
Previous concept models required expensive manual labeling: Existing Concept Embedding Models could explain their decisions well without sacrificing accuracy, but they required humans to manually annotate concepts in all training data—a process that becomes impractical for large datasets.
-
Automatic Concept Embedding Models eliminate the labeling bottleneck: ACEM removes the need for manual concept annotations by learning them automatically during training, making concept-based explanations feasible for real-world applications at scale.
-
This advancement addresses a key barrier to responsible AI deployment: By reducing the cost and effort required to create interpretable AI systems, this work helps remove practical obstacles to implementing explainability in production systems, supporting the "right to explanation" mentioned in regulations like GDPR.
Abstract
Interpretability and explainability of neural networks is continuously increasing in importance, especially within safety-critical domains and to provide the social right to explanation. Concept based explanations align well with how humans reason, proving to be a good way to explain models. Concept Embedding Models (CEMs) are one such concept based explanation architectures. These have shown to overcome the trade-off between explainability and performance. However, they have a key limitation -- they require concept annotations for all their training data. For large datasets, this can be expensive and infeasible. Motivated by this, we propose Automatic Concept Embedding Models (ACEMs), which learn the concept annotations automatically.