This module covers the fundamental concepts of deep learning methods and how to use them for extracting, modelling and visualizing the learned knowledge.
Topics include: neural networks with backpropagation, convolutional neural networks, recurrent neural networks, methods for representation learning, and how to use them under different learning regimes (supervised, unsupervised and reinforcement learning) and in variety of real-world applications ranging from computer vision, machine translation and medical image analysis.
This module also covers techniques of explainable AI (XAI) for understanding and visualizing how deep models make decisions and their generalization capabilities.
The learning objectives are:
a) to understand and use the main methodologies and techniques for learning from data
b) to understand the main methodologies to design and implement neural networks for real-world applications
c) to understand how to extract and learn knowledge in scenarios when supervision cannot be provided
d) to understand and foresee the reliability of machine learning methods in operational scenarios.
Knowledge and understanding
To understand the main concepts of learning from data
To understand concepts and tools for building intelligent systems using supervision and no supervision
To understand the most important machine learning and artificial intelligence methodologies and techniques used by industries to make sense of data in order to support the decision process
To understand what are the most appropriate techniques to be used in different real-world applications
To understand and interpret how machine learning models work in order to uncover inner mechanisms of black-box methods
Applying knowledge and understanding
To be able to effectively understand and use the main tools for creating, loading and manipulating datasets.
To design and implement from scratch a machine learning system following application-derived constraints in terms of modelling and data
To understand proper benchmarks and baselines and analysing achieved results and their generalization in real-world applications
To be able to apply methodologies and techniques to analyse data.
The module will focus on the implementations of various machine learning techniques and their applications in various domains. The primary tools used in the class are the Python programming language and several associated libraries.
The main teaching methods are as follows:
Lectures, to provide theoretical and methodological knowledge of the subject;
Hands-on exercises, to provide “problem solving” skills and to apply design methodology;
Laboratories, to learn and test the usage of related tools.
Paper reading and presentations to enhance understanding of the core concepts
Seminars by renowned experts (from both universities and industries) in the field to understand the current state of the art.
Should teaching be carried out in mixed mode or remotely, it may be necessary to introduce changes with respect to previous statements, in line with the programme planned and outlined in the syllabus.
Lectures, hands-on exercises, paper reading, student presentations and seminars
Should teaching be carried out in mixed mode or remotely, it may be necessary to introduce changes with respect to previous statements, in line with the programme planned and outlined in the syllabus.
Python programming language, statistical learning basic concepts
Python programming language, Linear Algebra
Strongly recommended. Attending and actively participating in the classroom activities will contribute positively towards the overall assessment of the final exam.
Strongly recommended. Attending and actively participating in the classroom activities will contribute positively towards the overall assessment of the oral exam.
The KD module consists of two parts: the first one will be addressing the general and modern techniques based on deep learning paradigm to create KD systems from data, while the second one on how to extract, represent and visualize knowledge from data and trained models.
Part I: Methods and Architectures
Neural Networks and Backpropagation
Derivatives and Gradient Descent
Neural Network Representation, Gradient descent for Neural Networks
Forward and Back Propagation
The revolution of depth: deep learning
Optimization algorithms: Mini-batch gradient descent, Exponentially weighted average, Gradient descent with momentum, RMSprop, Adam optimization algorithm, Learning rate decay
Training aspects of deep learning: Regularization, Dropout, Normalizing inputs, Vanishing / Exploding gradients, Weight Initialization for Deep Networks
Convolutional Neural Networks
Foundations: padding, strided convolution, dilation, 2D and 3D convolution, separable convolution, pooling
State of the art models: AlexNet, ResNets, DenseNets, Inception
Transfer Learning and Data Augmentation
Recurrent Neural Networks
LSTM and variants
Attention mechanisms
Part II: Knowledge Discovery from Data and Models
Unsupervised Learning with Deep Networks
Representation and Feature Learning
Autoencoders and Variational Autoencoders
Generative Adversarial Networks
Graph Neural Networks
Reinforcement Learning
Introduction to Reinforcement Learning
Policy Gradients
Actor-Critic Algorithms
Value Function Methods
Deep RL with Q-functions
Knowledge Discovery in Deep Models: Explainable AI (XAI)
Visualizing Convolutional Neural Decisions
Guided Backpropagation
Deep Generator Networks
CNN Visualization: Activation based and gradient based methods
Deep Learning Frameworks:
Overview of the most used DL frameworks
PyTorch and Jupyter Notebooks
Applications:
Computer vision
Medical Image Analysis
Machine translation
Introduction to the Course
Introduction to Machine Learning
Brief Python review and an overview of Numpy and Pandas
Review of data Characteristics of Data and Preparation and Preprocessing
Supervised Learning
Classification and Prediction using K-Nearest-Neighbor
Classifying with Probability Theory; Naïve Bayes
Building Decision Trees
Regression models
Evaluating predictive models
Ensemble Models: Bagging and Boosting
Unsupervised Learning
Clustering using K-Means
Hierarchical Clustering
Association Rule discovery
Principal Component Analysis and Dimensionality Reduction
Singular Value Decomposition
Brief note on Advance Topics
Matrix Factorization
Support Vector Machines
Search and Optimization Techniques
Markov models; time series analysis, sequential pattern mining
Real application domains
Text Mining and document analysis/filtering
Content analysis, TFxIDF transformation, text categorization, document clustering
Recommender systems
Neighborhood methods (user- and item-based)
Matrix factorization
Marketing and finance data analysis
Deep Learning. I. Goodfellow, Y. Bengio and A. Courville, MIT Press, 2016
Programming PyTorch for Deep Learning, I. Pointer, O'Reilly Media
Teaching materials and reading paper list provided by the instructor
Introduction to Machine Learning, Fourth Edition, By Ethem Alpaydin, MitPress ISBN: 9780262043793. 2020
Python Data Science Essentials - Third Edition by Alberto Boschetti, Luca Massaron, Packt Publishing, ISBN: 9781789537864, 2020
Teaching materials and reading paper list provided by the instructor
All teaching material will be published both on Studium and on www.perceivelab.com/teaching
All teaching material will be published both on Studium and on Microsoft Team
KNOWLEDGE DISCOVERY | ||
Argomenti | Riferimenti testi | |
1 | Neural networks: derivatives, gradient descent, back-propagation | 3 |
2 | Deep Learning: basic concepts, optimization algorithms, training procedures | 1, 3 |
3 | Convolutional Neural Networks | 1,3 |
4 | Recurrent Neural Networks | 1,3 |
5 | Unsupervised Learning with Deep Networks: Representation and Feature Learning | 1, 3 |
6 | Autoencoders and Variational Autoencoders | 1, 3 |
7 | Generative Adversarial Networks | 1, 3 |
8 | Graph Neural Networks | 3 |
9 | Reinforcement Learning: Deep Q-Networks and Policy Gradient | 3 |
10 | Explainable AI (XAI): Guided Backpropagation, Deep Generator Networks, CNN Visualization | 3 |
11 | Deep Learning Frameworks: PyTorch and Jupyter Notebooks | 2, 3 |
ADVANCED MACHINE LEARNING | ||
Argomenti | Riferimenti testi | |
1 | Introduction to Machine Learning | 1,3 |
2 | Python review | 3 |
3 | Pandas, Numpy, Matplotlib | 2,3 |
4 | Classification and Prediction: K-Nearest-Neighbor | 1 |
5 | Classification and Naive Bayes | 1 |
6 | Decision Tree | 1,3 |
7 | Regression models | 1 |
8 | Evaluating predictive models | 1 |
9 | Ensemble Models: Bagging and Boosting | 1 |
10 | Clustering using K-Means | 1 |
11 | Hierarchical Clustering | 1 |
12 | Association Rule discovery | 1 |
13 | Dimensional reduction | 1 |
14 | Singular Value Decomposition | 1 |
15 | Advanced topic | 1,3 |
16 | Sckit learn | 3 |
17 | ML in NLP | 3 |
18 | ML for Reccomender System | 3 |
The final exam consists of the development of a project in Pytorch, addressing one of the topics discussed during classes, together with a final report (structured as a scientific paper) discussing motivation, models, datasets and results used in the project.
The exam is evaluated according to the ability to create a deep learning model from scratch for extracting and learning knowledge from data on a given real-world problem, to understand how to properly measure its performance and to motivate the devised solutions.
The vote on the knowledge discovery module will account for 50% of the total grade for the entire course.
The module also foresees intermediate assignments. These assignments (between two to four) include: a) python scripts to solve simple basic learning problems on datasets discussed during with the instructor in order to avoid overlap and b) quizzes to verify the correct understanding of the presented techniques.
The grading policy for the KD module is:
50%: Final project
35%: Programming assignments
15%: Quizzes
Learning assessment may also be carried out on line, should the conditions require it.
There will be two assignments and one final exam. The assignments will contain written questions that require some Python programming. The final exam consists of quizzes and a final assignment.
The final assignment concerns comparative analysis on a given problem that must be presented in a final report and discussed in an oral discussion. The vote on the advanced machine learning module will account for 50% of the total grade for the entire course.
The grading policy for the AML module is:
50%: Final assignments
30% Intermediate assignments
20%: Quizzes
Learning assessment may also be carried out on line, should the conditions require it.
Examples of questions and exercises are available on the Studium platform and on the course website.
Examples of questions and exercises are available on the Studium platform and on the course website.