Rutgers ECE Associate Professor Anand D. Sarwate has been awarded a $70,000 research contract from the Pacific Northwest National Laboratories (PNNL) as part of a larger project on interpretability and the Mathematics of Artificial Reasoning Systems (MARS).
Prof. Sarwate will work with collaborators at PNNL to develop a structured framework for understanding the variability of the deep neural networks (DNNs) that drive contemporary ML/AI applications. Because DNNs are trained using stochastic optimization methods, there is inherent variability in the resulting predictive models: on one run they may be good and on another not so good. This is a form of process variation which only now being studied systematically in the machine learning literature. Prof. Sarwate will work with Rutgers ECE graduate student Sinjini Banerjee to develop statistical analyses of this performance variability in a longer-term effort to develop appropriate robustness/reliability measure for ML models. If successful, this pilot project will develop the foundations for a number of future investigations into assessing and characterizing the impact of different parameters and choices in neural network training on the stability of the training process itself.
Prof. Sarwate will work with collaborators at PNNL to develop a structured framework for understanding the variability of the deep neural networks (DNNs) that drive contemporary ML/AI applications. Because DNNs are trained using stochastic optimization methods, there is inherent variability in the resulting predictive models: on one run they may be good and on another not so good. This is a form of process variation which only now being studied systematically in the machine learning literature. Prof. Sarwate will work with Rutgers ECE graduate student Sinjini Banerjee to develop statistical analyses of this performance variability in a longer-term effort to develop appropriate robustness/reliability measure for ML models. If successful, this pilot project will develop the foundations for a number of future investigations into assessing and characterizing the impact of different parameters and choices in neural network training on the stability of the training process itself.
Congratulations to Anand!