Concept Inconsistency: Using Concept Level Information Learned by Deep Learning Models to Improve Training

Deep Learning (DL) has brought a revolution in the realm of machine learning (ML) by allowing us to model more complex problems. It extracts features from data which are not directly interpretable by humans. However, interpretable meta information about data classes can be used in various applications such as curriculum learning. There are numerous methods to extract interpretable model behavior. However, they have limited effectiveness on deep learning models. Using feature extraction methods also do not capture class level information. Our study explores the parameter space of a DL model to understand the concepts it encapsulates. We propose a novel metric to measure concept inconsistency by analyzing task vectors, which represent the differences between model parameters before and after training on a dataset. We also design a framework to generate relevant task vectors to measure concept inconsistency. Our experiments show that this metric effectively captures concept similarities across image data classes. We cluster data classes based on our proposed metric and train a classifier. The classification accuracy on these clusters is significantly higher than on randomly formed clusters. We also utilize our metric in training models using curriculum learning. Further results demonstrate that our proposed metric can be used in a curriculum learning setting to improve the accuracy of models. In conclusion, this metric can help us to make more informed decisions when developing and applying DL models.