Full session (30 minutes)
Data Science
Machine Learning
Deep learning

While Deep Learning (DL) is proven effective black box, without the ability to peek into the inner workings of our model, to understand and guide the learning process, we quickly reach a glass ceiling.

We tackle this with an adversarial game: A DL model that learns DL! While it can be generalized, we'll focus on NLP: Given an RNN that learns to embed text, we introduce an "Editor" model - by observing the RNN’s hidden states (without access to the input text) it predicts how editing each word will affect the RNN’s performances - forcing it to analyze the RNN’s perception of language and guide it’s learning process.

We'll demonstrate the Editor as a highly effective analytical tool that reveals crucial patterns in our model's behavior, dataset and dramatically improve generalization.

Asi Sheffer