CENDRe: Concept Extraction with Natural Domain Representations

Imagine you're working on a machine learning model that analyzes time-series data, like predicting equipment failures based on sensor readings. You want to understand not just the predictions but also the underlying patterns that lead to those predictions. Currently, many methods focus on analyzing the time domain, which means they miss out on important frequency information. This can lead to misinterpretations of what the model is actually learning, especially in critical applications where understanding the 'why' behind a prediction is essential. This is what's called a lack of concept extraction, where the model's reasoning remains opaque.

To tackle this, CENDRe offers a fresh approach to concept extraction for convolutional neural networks. It clusters latent representations of the data in two stages, automatically determining the number of concepts based on how well they fit together. By using gradients to create masks that highlight the most relevant areas in both time and frequency domains, CENDRe provides a clearer picture of what drives the model's predictions. This means that instead of just getting a prediction, you can also see which specific features in the data are influencing that prediction, making it easier to trust and validate the model's decisions. Compared to previous methods, CENDRe not only improves the accuracy of concept extraction but also aligns better with the actual regions of interest in the data, which is crucial for applications like fault diagnosis.

CENDRe introduces a novel approach to concept extraction in CNNs by addressing limitations in existing methods.

The method shows solid performance on both synthetic and real-world data, with clear metrics for evaluation.