Emerging Evolutionary Concepts: Clustering-Based Refinement of Concept Bottleneck Model Embeddings for Interpretable Machine Learning

The work projected in this research introduces a novel approach towards interpretable machine learning systems with the development of an evolutionary algorithm-based concept refinement pipeline for handwritten digit classification. The work presents an important obstacle in Explainable Artificial Intelligence (XAI) in automatically discovering optimal granularities for interpretable concepts while actually maintaining classification performance. The foundation of the pipeline relies on three important components: visual concept annotator, concept bottleneck model and evolutionary concept generator model. The visual concept annotator transforms raw MNIST digit images into binary concept annotations for five fundamental visual concepts (loops, vertical lines, horizontal lines, diagonal lines and curves) using tailored computer vision techniques that include the calculation of convexity for loop detection, Sobel transformation for vertical and horizontal lines, Hough line transformations for diagonal lines and curvature analysis in order to identify any curves. We then apply adaptive thresholding based on 75th percentile in order to convert continuous measurements to binary annotations. The concept bottleneck model represents a bottleneck architecture where all classification decisions pass through learned concept representations. The neural network has a multi-stage design: an image encoder generating 64-dimensional concept embeddings, concept prediction layers with sigmoid activation, and a minimal digit classifier accepting concept predictions as input. Multi-objective training simultaneously optimizes concept prediction and digit classification accuracies using binary cross-entropy and sparse categorical cross-entropy losses. The training results have presented excellent digit classification with the presence of the bottleneck constraint, while concept prediction accuracy was at most average, this clearly indicates that there is a notable room for improvement in the concept realm. The evolutionary algorithm presents an important innovation in this research, as it implements complex approaches to discover significant granularity of the concept through cluster refinement. Each individual inside a generation encodes a clustering configuration specifying strategies for all concepts simultaneously. The system utilizes both K-Means and DBSCAN clustering algorithms and a fitness evaluation that uses Random Forest classification on newly generated concepts. Experimental results have shown a significant difference between K-mean clustering that were able to get higher results over the original concept, while the DBSCAN clustering technique achieved less impressive improvements on the original concepts. The research itself has demonstrated that evolutionary algorithms are able to automatically discover refined concepts that can represent a correct classification of the dataset. The significant improvement made by K-means clustering represents a great progress towards reducing the gap present between interpretability and performance in neural networks, as they provided interpretability on a structural level with concepts that are visually understandable by humans. This work provides a great foundation on evolutionary concept optimization as a promising direction for interpretable machine learning, as it offers a coherent approach to emerging concept discovery that reduces the reliance on concepts that are written by humans, with future researches that can focus on semantic-level concepts.