Unlocking Image Understanding at Scale: Transformers for Image Recognition

Interpreting neural networks is like trying to understand a recipe without knowing what’s in it. Self-Explanatory Neural Networks (SENN) are the equivalent of a recipe book that tells you what’s inside and how it’s made, making it easier to understand. SENN is a new framework for neural networks that improves interpretability by providing detailed information about the concepts learned from the input data.
The article discusses three desiderata for concepts in SENN: Fidelity (preserving relevant information), Diversity (non-overlapping concepts), and Grounding (aligning with human interpretable concepts). The authors propose a method that exceeds stability requirements by using constant, quantized, and sparse weights.
Imagine you’re trying to make a cake recipe but need to understand what each ingredient does. SENN is like providing a list of all the ingredients with their roles and how they interact. This makes it easier to adjust the recipe or identify problem areas.
The authors also address the issue of feature disentangling, which can lead to reduced interpretability. Disentangling features can help increase alignment between learned concepts and human interpretable ones.
In summary, SENN is a framework that improves interpretability by providing detailed information about the concepts learned from input data. It does this by exceeding stability requirements, using constant, quantized, and sparse weights, and addressing feature disentangling issues. Understanding SENN is like having a recipe book that explains what’s in each dish, making it easier to create new ones or identify problem areas.