Advanced Deep Learning

What you will be able to do

• Handle imbalanced classification.
• Check whether predicted probabilities are calibrated, and attach conformal prediction intervals to a model's predictions.
• Adapt pretrained models with transfer learning and fine-tuning.
• Learn from unlabelled data with generative models (autoencoders, GANs and normalizing flows), and explain what each one learns: a representation, a way to sample, or the data distribution.

Prerequisites

An introduction to deep learning, for example Introduction to Deep Learning, and Python experience.

Outline

Duration: 2 days, on-site or remote, with exercises.

Six topics, in this order. The last three learn from the data itself, without labels: autoencoders and GANs are examples of self-supervised learning, and normalizing flows learn the data distribution directly.

1. Imbalanced classification
2. Calibration and conformal prediction
3. Transfer learning and fine-tuning
4. Autoencoders: learning a representation
5. Generative adversarial networks (GANs): learning to sample
6. Normalizing flows: learning the data distribution

Sample materials

• Imbalanced data
• Calibration and conformal prediction
• Transfer learning
• Autoencoders
• Generative adversarial networks
• Normalizing flows

More in yoavram/DataSciPy.

Related workshops

• Introduction to Deep Learning
• Open-Set Learning