Towards Technically Responsible AI via Self-Healing Inference and Tensor-Compressed Training

February 17, 2023, ESB 2001

Zheng Zhang

UCSB, Electrical and Computer Engineering


As modern AI is reshaping almost every aspect of our daily life, responsible AI has become an increasingly important topic: we should design and deploy AI with good intentions and positive impacts, both technically and socially. If we rethink the current AI ecosystem from the responsible AI perspective, there are many concerns in the inference and training process. This talk will focus on two issues: the trustworthiness and sustainability issues. One fundamental challenge of neural networks is their vulnerability: the accuracy of a neural network classifier can drop significantly when the input data is perturbed by an invisible amount. This is a big concern in safety-critical applications such as medical imaging and autonomous driving. Adversarial training and defense techniques often rely on some strong assumptions of the perturbation or attack information. In practice, it is hard to know such information exactly in advance, and a “robustly” designed neural network may still fail to work easily. In this talk, we present a self-healing method, which can detect and fix the possible errors of a neural network automatically in the inference. This method is formulated as a closed-loop control. Since this method does not need a-priori attack information, it can handle a broad class of unforeseen attacks or perturbations that conventional methods cannot handle. Another increasing concern is the sustainability issue of large-scale AI training. Record-breaking GPU-hours have been used to train large deep learning models, causing ever increasing carbon emissions. The huge model size also prevents energy-efficient on-device training. Can we train large AI models with a low computing, memory and finally environmental cost? In the second part, we will show our efforts in tensor-compressed training. With a proposed rank-adaptive tensor-compressed training, it is possible to reduce the training variables of large models by orders of magnitudes. This can significantly reduce the hardware cost of training, and enable energy-efficient and communication-efficient training in cloud, edge and federated settings.

Speaker's Bio

Dr. Zheng Zhang is an Assistant Professor of Electrical and Computer Engineering at University of California, Santa Barbara. He received his PhD degree in Electrical Engineering and Computer Science from MIT in 2015. His research is focused on uncertainty quantification and tensor computation for semiconductor chip design automation and for responsible AI systems. He is a recipient of NSF CAREER award, 3 best journal paper awards from IEEE Transactions, and two best dissertation awards from ACM SIGDA and MIT Microsystems Technology Labs.

Video URL: