דניאל סודרי (הנדסת חשמל, טכניון)
יום שני, 4.11.2019, 12:30
In many common datasets, neural networks can achieve zero training loss yet generalize well to unseen data. Recent works suggest that this because standard training algorithms (e.g., GD or SGD) have an implicit regularization which is biased towards specific solutions, which tend to have good generalization. I will review such "algorithmic biases", how they affect the functional capabilities of the neural networks, and how this relates to generalization in simple models.