Robust Video Instance Segmentation and Multi-Object Tracking 

Collaboration: Huawei

Current self-driving cars relying on neural networks for detection are vulnerable to a variety of noises like weather noises, and digital corruption noises. Our objective is to develop a robust architecture for VIS and MOT for self-driving cars. 

Bio-Inspired Gaze Augmentation for Transformers

In this project, we are researching bio-inspired sampling strategies for Vision Transformers. We developed Random Scanning and Biased Scanning strategies inspired by saccadic eye movement in humans. We also observed that our bio-inspired scanning enabled transformers to have adaptive attention in the given image.

Evolving Robust Neural Networks

In this project, we are researching to evolve deep neural networks through neuro-evolution. Our objective is to transfigure a more robust architecture than the current deep neural networks and also study the developed relatively resilient models for characteristics in the architecture which contribute towards robustness.

Understanding Adversarial Attacks and Defences

In this project, we are researching existing adversarial attacks for deep neural networks. We developed our own L0 and L∞ adversarial attacks based on Evolutionary Strategies (DE and CMA-ES) which are some of the most extreme adversarial attacks with minimal perturbations. We analysed a variety of adversarial attacks and different neural architectures to link representational bias with attack accuracies. We found that the models which had high representational bias tend to be more susceptible to adversarial attacks. Also, we generated the saliency and attention maps of the adversarial samples to understand the change of saliency and attention at the affected portion of the adversarial example.

Last Updated: September 2024