In this talk, I will share recent advancements from my lab in explainable AI and interpretable machine learning for computer vision. I will present a series of works that aim to bring interpretability into deep learning across three fronts: (i) unveiling the inner workings of neural networks through automated mechanistic interpretability techniques [1-3], (ii) developing inherently interpretable architectures such as concept bottleneck models [4-7], and (iii) establishing a unified framework for evaluating neuron-level explanations, including practical guidelines and a set of reliable evaluation metrics [8-9]. Together, these efforts mark important steps toward building trustworthy and transparent deep learning systems for computer vision. 

Reference: 

[1] Oikarinen and Weng, CLIP-Dissect: Automatic Description of Neuron Representations in Deep Vision Networks, ICLR 23 (spotlight)

[2] Oikarinen and Weng, Linear Explanations for Individual Neurons, ICML 24

[3] Bai & Iyer etal, Describe-and-Dissect: Interpreting Neurons in Vision Networks with Language Models, TMLR 25

[4] Oikarinen etal, Label-Free Concept Bottleneck Models, ICLR 23

[5] Srivastava & Yan etal, VLG-CBM: Training Concept Bottleneck Models with Vision-Language Guidance, NeurIPS 24

[6] Sun etal, Concept Bottleneck Large Language Models, ICLR 25 

[7] Kulkarni etal, Interpretable Generative Models through Post-hoc Concept Bottlenecks, CVPR 25

[8] Oikarinen etal, Evaluating Neuron Explanations: A unified Framework with Sanity Checks, ICML 25 

[9] Oikarinen etal, Rethinking Crowd-Sourced Evaluation of Neuron Explanations, arxiv preprint 25