2021/2022

2022/2023

DM270

ING-INF/05 (SISTEMI DI ELABORAZIONE DELLE INFORMAZIONI)

DIPARTIMENTO DI INGEGNERIA INDUSTRIALE E DELL'INFORMAZIONE

COMPUTER ENGINEERING

Computer Science and Multimedia

2°

Secondo Semestre (06/03/2023 - 19/06/2023)

6

45 ore di attività frontale

English

ORALE

PIASTRA MARCO (titolare) - 6 CFU

Foundations of linear algebra and multivariable calculus. Practical experience with at least one programming language. Basic knowledge of Python, acquaintance with Numpy.

The course follows a conceptual pathway that starting from simple linear regression to the sophisticated aspects of state-of-art of deep convolutional neural networks, deep recurrent networks and deep reinforcement learning. A unifying mathematical approach is followed throughout this path, to encompass and make it possible to understand the basic features of modern software frameworks for deep learning, such as TensorFlow.

1) Deep Supervised Learning

Algebraic model, foundations of tensor calculus.
Learning as representation, evaluation and optimization.
Single-layer networks as universal approximators.
Datasets in tensorial representation, for calculus.
Flow diagrams, automatic differentiation.
Regression vs. classification, softmax.
Deep layered representation, modularity.
Layer-wise gradient computation.

2) Deep Convolutional Neural Networks

Convolutional layers and complex architectures.
Data augmentation and Transfer learning.
Layered learning, different optimization processes.
Fallibility and adversarial models.
Autoencoders.
Classification, object detection, segmentation

3) Deep Recurrent Networks

Temporal unfolding, shared-parameters layers
Long-Short Term Memory (LSTM).
Attention and transformers.

4) Deep Reinforcement Learning

On-policy and off-policy learning
Actor critic and advance function
Neural MCTS: AlphaZero e MuZero

Lectures (hours/year in lecture theatre): 30
Practical class (hours/year in lecture theatre): 16
Practicals / Workshops (hours/year in lecture theatre): 0

See the home page of the course (http://vision.unipv.it/DL) for lecture slides, suggested readings and software for the exercises.

Software notebooks (Google Colab) used for the exercises will be published on KIRO.

The evaluation includes the realization of a project, to be agreed in advance. The final exam is an interview about the theory, together with the discussion of the project.