Karthik Rao

Avi Schwarzschild, Max Cembalest, Karthik Rao et al.

As neural networks increasingly make critical decisions in high-stakes settings, monitoring and explaining their behavior in an understandable and trustworthy manner is a necessity. One commonly used type of explainer is post hoc feature attribution, a family of methods for giving each feature in an input a score corresponding to its influence on a model's output. A major limitation of this family of explainers in practice is that they can disagree on which features are more important than others. Our contribution in this paper is a method of training models with this disagreement problem in mind. We do this by introducing a Post hoc Explainer Agreement Regularization (PEAR) loss term alongside the standard term corresponding to accuracy, an additional term that measures the difference in feature attribution between a pair of explainers. We observe on three datasets that we can train a model with this loss term to improve explanation consensus on unseen data, and see improved consensus between explainers other than those used in the loss term. We examine the trade-off between improved consensus and model performance. And finally, we study the influence our method has on feature attribution explanations.

Karthik Rao, William Song, Yorai Wardi et al.

The consistent demand for better performance has lead to innovations at hardware and microarchitectural levels. 3D stacking of memory and logic dies delivers an order of magnitude improvement in available memory bandwidth. The price paid however is, tight thermal constraints. In this paper, we study the complex multiphysics interactions between performance, energy and temperature. Using a cache coherent multicore processor cycle level simulator coupled with power and thermal estimation tools, we investigate the interactions between (a) thermal behaviors (b) compute and memory microarchitecture and (c) application workloads. The key insights from this exploration reveal the need to manage performance, energy and temperature in a coordinated fashion. Furthermore, we identify the concept of "effective heat capacity" i.e. the heat generated beyond which no further gains in performance is observed with increases in voltage-frequency of the compute logic. Subsequently, a real-time, numerical optimization based, application agnostic controller (TRINITY) is developed which intelligently manages the three parameters of interest. We observe up to $30\%$ improvement in Energy Delay$^2$ Product and up to $8$ Kelvin lower core temperatures as compared to fixed frequencies. Compared to the \texttt{ondemand} Linux CPU DVFS governor, for similar energy efficiency, TRINITY keeps the cores cooler by $6$ Kelvin which increases the lifetime reliability by up to 59\%.