Cohere For AI - Shreyash Arya, Research Staff and Collaborator

B-cosification: Transforming Deep Neural Networks to be Inherently Interpretable
Abstract: B-cos Networks have been shown to be effective for obtaining highly human interpretable explanations of model decisions by architecturally enforcing stronger alignment between inputs and weight. B-cos variants of convolutional networks (CNNs) and vision transformers (ViTs), which primarily replace linear layers with B-cos transformations, perform competitively to their respective standard variants while also yielding explanations that are faithful by design. However, it has so far been necessary to train these models from scratch, which is increasingly infeasible in the era of large, pre-trained foundation models. In this work, inspired by the architectural similarities in standard DNNs and B-cos networks, we propose ‘B-cosification’, a novel approach to transform existing pre-trained models to become inherently interpretable. We perform a thorough study of design choices to perform this conversion, both for convolutional neural networks and vision transformers. We find that B-cosification can yield models that are on par with B-cos models trained from scratch in terms of interpretability, while often outperforming them in terms of classification performance at a fraction of the training cost. Subsequently, we apply B-cosification to a pretrained CLIP model, and show that, even with limited data and compute cost, we obtain a B-cosified version that is highly interpretable and competitive on zero shot performance across a variety of datasets. We release our code and pre-trained model weights at https://github.com/shrebox/B-cosification.

Full name: Shreyash Arya
Affiliation: Scientific Research Staff at CISPA / Research Collaborator at Max Planck Institute for Informatics

Intro:
I recently completed my Master’s in Computer Science at Saarland University, Saarbrücken, and am set to join CISPA's startup, Detesia, as an AI Engineer specializing in Machine Learning Explainability. In my previous role at the Max Planck Institute for Informatics, I focused on Explainable Machine Learning in Computer Vision, with my master’s thesis, “B-cosification: Transforming Deep Neural Networks to be Inherently Interpretable,” recently accepted at Neural Information Processing Systems (NeurIPS) 2024 conference. The skills I developed through this work, from designing interpretable models to deepening my understanding of AI transparency, will now drive my contributions at Detesia. Outside of work, I enjoy badminton, basketball, swimming, hiking, and mobile photography, and I love connecting with new people.