Vishal Keshav

I'm a Senior Engineer at Apple, focusing on Generative AI quality with a special emphasis on automating systems and processes.

Previously, I worked at Quora as an ML Engineer, on Quora.com (Ranking, Moderation) and Poe.com (RAG). Before that, in my short time at Amazon, I worked with the internationalization team, and at Samsung Research, I focused on on-device AI.

I obtained my master's in computer science (AI focus) from University of Massachusetts Amherst and a Bachelor's in Mathematics & Computing from IIT Guwahati.

Posts / Email / GitHub / Google Scholar / LinkedIn

Projects/Papers/Patents

	Project - Structuring parameter uncertainity for efficient incremental learning pdf / code / We exploit the computational structure of convolutions to re-parametrize the weight uncertainty (under the Bayesian framework) in an incremental learning setting. We validate our hypothesis “structured adaptability bias” through empirical experimentations. Our method results in both parameter efficiency and accuracy improvements over state-of-the-art.
	Project - Cost aware inference with early exit pdf / code / slides / For real-time non-critical continuous machine learning applications, we propose a novel approach to dynamically adjust the model execution time to strike the right balance between accuracy and runtime efficiency. We take advantage of early exit architectures where execution can be halted without losing the final predictions.
	Project - Distributed Multi-threaded Map-Reduce Library pdf / code / An implementation of the map-reduce protocol in C++. The implementation supports the execution of user-defined map-reduce functions in a distributed environment and ensures fault tolerance.
	Paper - Self-supervised visual feature learning with curriculum paper / code / Training on a fixed pre-text task has the disadvantage of either letting network take shortcuts or miss semantically valuable information. We take inspiration from the curriculum learning technique to progressively remove low-level signals and show that it significantly increases the speed of convergence of the downstream tasks.
	Paper - Decoupling Semantic Context and Color Correlation with Multi-class Cross Branch Regularization paper / code / poster / We present a novel design methodology for architecting a light-weight and faster DNN architecture for vision applications. We propose to adopt a multi-branch architecture that decouples pixel level and semantic level signals from an image using efficient point-wise and depth-wise convolution operators. The fused results are then used for global prediction. We achieve computational efficiency by a magnitude of 30x while achieving the same accuracy as state-of-the-art architectures.
	Patent - Method and system for DNN based imaging patent / The method comprises of simultaneously extracting per-pixel color distribution along the RGB channel dimension and semantic information from each of the color channels, later, fusing the processed signals to output the predictive outcome based on the task.
	Patent - Method and system for context based task scheduling in a multi-core processor patent / patent #2 / The patents provide methods for managing the execution of tasks by symmetric and asymmetric multi-core processors.
	Project - Optimisation of power consumption in wireless network pdf / code / slides / This is my Bachelor’s thesis aimed at modeling a wireless network with M/M/1 queue, define a multi-variate power function and apply optimization techniques such as backtrack line search algorithm to minimize the power consumption of the wireless network. We theoretically show the improvements in power consumption given our proposed model.