Agenda

Dalila Ressi, DAIS

Abstract:
This seminar is the first in a series of informal lectures regarding the practical aspects of modern AI techniques applied to real-world recognition and classification tasks.

The series, named "Deep Learning in Action", will be presented by Dalila Ressi, a freshman PhD student working in tight contact with industrial partners. The indented audience is not only Computer Vision people, but any researcher (and student) interested in knowing more about this increasingly popular topic.

In fact, in the last years Deep Learning adoption has grown exponentially thanks to the results achieved by neural networks in a wide range of applications. Even though this success, being able to apply and to make work these techniques on real data is not as simple as it might seem. Neural networks are complex structures that are not easy to understand, and they might present problems bigger than the one we are trying to solve.

In particular, this first seminar will focus on the main difficulties usually met using Deep Learning, like:

• problems arising from using the gradient (vanishing gradient, exploding gradient, gradient clipping)
• regularization, or how to prevent overfitting (weight decay, bagging/dropout, early stopping, learning rate and momentum, optimizers)
• hyperparameters serach and initialization
• dataset preparation (how to preprocess the dataset in order to avoid data dependencies or detect outliers)
• what to do when everything tried failed

The audience will finally be provided with main guide lines coming from the personal experience of the speaker.