Bhushan Atote bio photo

Bhushan Atote

PhD candidate
University of Warwick
Coventry, England, UK

Twitter LinkedIn Google Scholar Github

Research

My interests broadly lie in the fields of machine learning, computer vision and explainable AI. You can refer my current and completed research projects here!

Projects

### Generative Explainable Models

Project: How were you created? Explaining synthetic face images generated by diffusion models

Abstract: Diffusion probabilistic models have been shown to surpass the capabilities of Generative Adversarial Networks (GANs), setting a new standard in the field of generative models. While the emphasis has been predominantly on producing high-quality images, the explainability of diffusion models has often been overlooked. In this work, we introduce a state-of-the-art approach designed to explain face images generated by diffusion models. To this end, we introduce a novel methodology, named the Explainable DIffusion PRobabilistic (EDIPR) model, based on a classification framework. EDIPR consists of three stages: an initial clustering stage, which serves as the pre-processing step to discover groups of similar face images in the training set; a synthesizing stage, carried out by a diffusion model; and an explaining stage, which allows determining which training images contributed the most to the generation of a new face image. To provide explainability, we also introduced two influence scores as quantitative metrics: the Normalized Influence Score (NIS) and the class-Normalized Influence Score (cNIS). These scores provide the probability that a specific training image, or class, contributes to the generation of a new synthetic image. Our results on synthetic images generated based on the real images of the FFHQ dataset show that EDIPR provides robust and plausible explanations linking the training images to the synthetic images at three levels of granularity: the region, the image, and the class level.

Discriminative Explainable Models

Project: Enhanced Prototypical Part Network (EPPNET) for explainable image classification via prototypes

Abstract: Explainable Artificial Intelligence (xAI) has the potential to enhance the transparency and trust of AI-based systems. Although accurate predictions can be made using Deep Neural Networks (DNNs), the process used to arrive at such predictions is usually hard to explain. In terms of perceptibly human-friendly representations, such as word phrases in text or super-pixels in images, prototype-based explanations can justify a model’s decision. In this work, we introduce a DNN architecture for image classification, the Enhanced Prototypical Part Network (EPPNet), which achieves strong performance while discovering relevant prototypes that can be used to explain the classification results. This is achieved by introducing a novel cluster loss that helps to discover more relevant human-understandable prototypes. We also introduce a faithfulness score to evaluate the explainability of the results based on the discovered prototypes. Our score not only accounts for the relevance of the learned prototypes but also theperformance of a model. Our evaluations on the CUB-200-2011 dataset show that the EPPNet outperforms state-of-theart xAI-based methods, in terms of both classification accuracy and explainability.

Extra Project: Robust Multiview Multimodal Driver Monitoring System Using Masked Multi-Head Self-Attention

Abstract: Driver Monitoring Systems (DMSs) are crucial for safe hand-over actions in Level-2+ self-driving vehicles. State-of-the-art DMSs leverage multiple sensors mounted at different locations to monitor the driver and the vehicle’s interior scene and employ decision-level fusion to integrate these heterogenous data. However, this fusion method may not fully utilize the complementarity of different data sources and may overlook their relative importance. To address these limitations, we propose a novel multiview multimodal driver monitoring system based on feature-level fusion through multi-head self-attention (MHSA). We demonstrate its effectiveness by comparing it against four alternative fusion strategies (Sum, Conv, SE, and AFF). We also present a novel GPU-friendly supervised contrastive learning framework SuMoCo to learn better representations. Furthermore, We fine-grained the test split of the DAD dataset to enable the multi-class recognition of drivers’ activities. Experiments on this enhanced database demonstrate that 1) the proposed MHSA-based fusion method (AUC-ROC: 97.0%) outperforms all baselines and previous approaches, and 2) training MHSA with patch masking can improve its robustness against modality/view collapses. The code and annotations are publicly available.