| |
|
| |
| DETAILS |
|
As machine learning algorithms increasingly pervade our everyday life, it becomes imperative that we better understand the algorithms we deploy. In this talk I will present my work that uses classical kernel methods to study empirical phenomena in machine learning.
I will first present structured kernel constructions that provide competitive performance on a range of scientific tasks ranging from computational biology to heliophysics, but fall short on standard machine learning benchmarks such as Cifar-10 & ImageNet.
Then I take a substantial detour to understand whether the high predictive performance of neural networks on standard benchmarks is fundamental, or simply due to overfitting on test sets. This leads to a line of work involving constructing novel test sets for Cifar-10 & ImageNet & precisely measuring human & model performance on these datasets.
Finally I return to improve my structured kernel constructions to achieve significantly higher performance on standard machine learning benchmarks.
Bio: Vaishaal Shankar is a final year PhD student working with Ben Recht at UC Berkeley. He broadly works on experimental analysis of phenomena in machine learning. A majority of his research has revolved around understanding the fundamental limitations of deep neural networks & their connection to classical kernel methods. He will be joining a special projects team at Amazon in Fall 2020.
|
|
|
|
|
|
|
|
