Topic outline

  • Intelligent Systems for Pattern Recognition - 9 CFU

    Code: 760AA, Credits (ECTS): 9, Semester: 2, Official Language: English

    Instructor: Davide Bacciu 

    Contact: email - phone 050 2212749

    Office: Room 367, Dipartimento di Informatica, Largo B. Pontecorvo 3, Pisa

    Office Hours: (email to arrange meeting)

  • Course Information

    Weekly Schedule

    The course is held on the second term. The tentative schedule for A.A. 2022/23 is provided in table below.

    The first lecture of the course will be ON THE WEEK OF FEBRUARY 20-24 (exact date TBD). The course will be hybrid, both in person and online on the dedicated MS Team.

    Recordings of the lectures will be made available to the students following the course.

    Day Time
    Tuesday 11.00-12.45 (Room C1)
    Wednesday 16.15-18.00 (Room E)
    Thursday 14.15-16.00 (Room C)


    Course Prerequisites

    Course prerequisites include knowledge of machine learning fundamentals (e.g. covered through ML course). Knowledge of elements of probability and statistics, calculus and optimization algorithms are also expected. Previous programming experience with Python is a plus for the practical lectures.

    Course Overview

    The course introduces students to the analysis and design of advanced machine learning and deep learning models for modern pattern recognition problems and discusses how to realize advanced applications exploiting computational intelligence techniques.

    The course is articulated in five parts. The first part introduces basic concepts and algorithms concerning traditional pattern recognition, in particular as pertains sequence and image analysis. The next two parts introduce advanced models from two major learning paradigms, that are deep neural networks and generative models,  their combinaton and their use in pattern recognition applications. The fourth part will introduce fundamentals of reinforcement learning and deep reinforcement learning. The final part of the course will present selected recent works, models and applications of learning models.

    Presentation of the theoretical models and associated algorithms will be complemented by introductory classes on the most popular software libraries used to implement them.

    The course hosts guest seminars by national and international researchers working on the field as well as by companies that are engaged in the development of advanced applications using machine learning models.

    The official language of the course is English: all materials, references and books are in English. Lecture slides will be made available here, together with suggested readings.

    Topics covered -Bayesian learning, graphical models, learning with sampling and variational approximations, fundamentals of deep learning (CNNs, AE, DBN, GRNs), deep learning for machine vision and signal processing, advanced deep learning models (transformers, NTMs), generative deep learning (VAE, GANs, diffusion models, score-based models) deep graph networks, reinforcement learning and deep reinforcement learning, signal processing and time-series analysis, image processing, filters and visual feature detectors, pattern recognition applications (machine vision, bio-informatics, robotics, medical imaging, etc), introduction to programming libraries and frameworks.

    Textbooks and Teaching Materials

    The course does not have an official textbook covering all its contents. However, a list of reference books covering parts of the course is listed at the bottom of this section (note that all of them have an electronic version freely available online).

    [BRML] David Barber, Bayesian Reasoning and Machine Learning, Cambridge University Press (PDF)

    [DL] Ian Goodfellow and Yoshua Bengio and Aaron Courville , Deep Learning, MIT Press (ONLINE)

    [RL] Richard S. Sutton and Andrew G. Barto, Reinforcement Learning: An Introduction, Second Edition, MIT Press, 2018 (PDF)

    • Introduction (2h)

      Introduction to the course philosophy, its learning goals and expected outcomes. We will discuss prospectively the overall structure of the course and the interelations between its parts. Exam modalities and schedule are also discussed.
      Date Topic  References  
       Additional Material 
      Introduction to the course
      Motivations and aim; course housekeeping (exams, timetable, materials); introduction to modern pattern recognition applications

      • Fundamentals of Pattern Recognition (6h)

        The module will provide a brief introduction to classical pattern recognition for signal/timeseries and for images. We will cover approaches working on the spatial (temporal) and frequency (spectral) domain, presenting methods to represent temporal and visual information in static descriptors, as well as approaches to identify relevant patterns in the data (feature descriptors). Methodologies covered include correlation analysis, Fourier analysis, wavelets, intensity gradient-based descriptors and detectors, normalized cut segmentation.

        Date Topic  References  
         Additional Material 
        Signal processing
        Timeseries; time domain analysis (statistics, correlation); spectral analysis; fourier analysis.

        3 17/02/2022
         Image Processing I
         Spatial feature descriptors (color histograms, SIFT); spectral analysis.
          Additional readings
         [1] Survey on visual descriptors
         4 22/02/2022
         Image Processing II
         Feature detectors (edge, blobs); image segmentation; wavelet decompositions
          Additional readings 
        [2] Survey on visual feature detectors


        A wavelet browser to visualize some popular wavelet families and their instances, powered by the PyWavelet library.

        • Generative Learning (20h)

          The module introduces probabilistic learning, causal models, generative modelling and Bayesian learning. We will discuss fundamental algoritms and concepts, including Expectation-Maximization, sampling and variational approximations, and we will study relevant models from the three fundamental paradigms of probabilistic learning, namely Bayesian networks, Markov networks and dynamic models.  Models covered include: Bayesian Networks, Hidden Markov Models, Markov Random Fields, Boltzmann Machines,  Latent topic models.

          Date Topic  References  
           Additional Material 
          Introduction to Generative and Graphical Models.
          robability refresher; graphical model representation; directed and undirected models
          [BRML] Ch. 1 and 2 (Refresher)
          [BRML] Sect. 3.1, 3.2 and 3.3.1
          (conditional independence)
          • Pyro - Python library based on PyTorch
          • PyMC3 - Python library based on Theano
          • Edward - Python library based on TensorFlow
          • TensorFlow Probability - Probabilistic models and deep learning in Tensorflow
          Conditional independence and causality - Part I
          Bayesian networks; Markov networks; conditional independence;

          [BRML] Sect. 3.3 (Directed Models)
          [BRML] Sect. 4.1, 4.2.0-4.2.2 (Undirected Models)
          [BRML] Sect. 4.5 (Expressiveness)

          Conditional independence and causality - Part II
          d-separation; structure learning in Bayesian Networks
          [BRML] Sect. 9.5.1 (PC algorithm)
          [BRML] Sect. 9.5.2 (Independence testing)
          [BRML] Sect. 9.5.3 (Structure scoring)
          Additional readings
          [3] A short review of BN structure learning
          [4] PC algorithm with consistent ordering for large scale data
          [5] MMHC - Hybrid structure learning algorithm

          - A selection of BN structure learning libraries in Python: pgmpy, bnlearn, pomegranate.
          - bnlearn: the most consolidated and efficient library for BN structure learning (in R)
          - Causal learner: a mixed R-Matlab package integrating over 26 BN structure learning algorithms.

          Hidden Markov Models  - Part I
          learning in directed graphical models; forward-backward algorithm;  generative models for sequential data
          [BRML] Sect. 23.1.0 (Markov Models)
          [BRML] Sect. 23.2.0-23.2.4 (HMM and forward backward) 
          Additional Readings
          [6]  A classical tutorial introduction to HMMs
          Hidden Markov Models - Part II
          EM algorithm, learning as inference, Viterbi Algorithm
          [BRML] Sect. 23.2.6 (Viterbi)
          [BRML] Sect. 23.3.1-23.3.4 (EM and learning)

          HMM III + Markov Random Fields
          learning in undirected graphical  models; conditional random fields; pattern recognition applications

          [BRML] Sect. 4.2.2, 4.2.5 (MRF)
          [BRML] Sect. 4.4 (Factor Graphs)
          [BRML] Sect. 5.1.1 (Variable Elimination and Inference on Chain) 
          [BRML] Sect. 9.6.0, 9.6.1, 9.6.4, 9.6.5 (Learning in MRF/CRF)
          Additional Readings
          [7,8] Two comprehensive tutorials on CRF ([7] more introductory and [8] more focused on vision)
          [9] A nice application of CRF to image segmentation

          Check out pgmpy: it has Python notebooks to introduce to working with MRF/CRF
          Bayesian Learning I
          Principles of Bayesian learning; EM algorithm objective; principles of variational approximation
          [BRML] Sect. 11.2.1 (Variational EM)
          Bayesian Learning II
          latent topic models; Latent Dirichlet Allocation; machine vision application of latent topic models
          [BRML] Sect. 20.4-20.6.1  (LDA) Additional Readings
          [10] LDA foundation paper
          [11] A gentle introduction to latent topic models
          [12] Foundations of bag of words image representation
          ROOM C1
          Bayesian Learning III
          sampling methods; ancestral sampling; Gibbs sampling and Monte Carlo methods

          Guest lecture by Daniele Castellana 

          [BRML] Sect. 27.1-27.3, 27.4.1, 27.6.2 Additional Readings
          [13] A step-by-step derivation of collapsed Gibbs sampling for LDA
          Boltzmann Machines
          bridging neural networks and generative models; stochastic neuron; restricted Boltzmann machine; contrastive divergence
           [DL] Sections 20.1 and 20.2 Additional Readings
          [14] A clean and clear introduction to RBM from its author

          Matlab code for Deep Belief Networks (i.e. stacked RBM) and Deep Boltzmann Machines.

          • Deep Learning (24h)

            The module presents the fundamental concepts, challenges, architectures and methodologies of deep learning. We introduce the learning of neural representations from vectorial, sequential and image data, covering both supervised and unsupervised learning, and hinting at various forms of weak supervision. We close the gap between neural networks and probabilistic learning by discussing generative deep learning models. Models covered include: deep autoencoders, convolutional neural networks, long-short term memory, gated recurrent units, deep reservoir computing, sequence-to-sequence, neural attention, neural Turing machines, variational autoencoders, generative adversarial networks. Methodological lectures will be complemented by introductory seminars to Keras-TF and Pytorch.

            Date Topic  References  
             Additional Material 
            Convolutional Neural Networks I
            Introduction to the deep learning module; i
            ntroduction to CNN; basic CNN elements
            [DL] Chapter 9
            Additional Readings
            [15-19] Original papers for LeNet, AlexNet, VGGNet, GoogLeNet and ResNet.
             Convolutional Neural Networks II
            CNN architectures for image recognition; convolution visualization; advanced topics (deconvolution, dense nets); applications and code
             [DL] Chapter 9 Additional Readings
            [20] Complete summary of convolution arithmetics
            [21] Seminal paper on batch normalization
            [22] CNN interpretation using deconvolutions
            [23] CNN interpretation with GradCAM
             17  25/03/2022
            ROOM D1 + ONLINE
             Deep Autoencoders
            parse, denoising and contractive AE; deep RBM
             [DL] Chapter 14, Sect 20.3, 20.4.0 (from 20.4.1 onwards not needed)  Additional Readings
            [24] DBN: the paper that started deep learning
            [25] Deep Boltzmann machines paper
            [26] Review paper on deep generative models
            [27] Long review paper on autoencoders from the perspective of representation learning
            [28] Paper discussing regularized autoencoder as approximations of likelihood gradient
             18  28/03/2022
            ROOM C
             Gated Recurrent Networks I
            Deep learning for sequence processing; gradient issues;
             [DL] Sections 10.1-10.3, 10.5-10.7, 10.10, 10.11  Additional Readings
            [29] Paper describing gradient vanish/explosion
            [30] Original LSTM paper
            [31] An historical view on gated RNN
            [32] Gated recurren units paper
            [33] Seminal paper on dropout regularization
             19  29/03/2022
             Gated Recurrent Networks II
            long-short term memory; gated recurrent units; generative use of RNN
             20 30/03/2022
             Coding practice I - Tensorflow
             21 31/03/2022
             Coding practice II - PyTorch
             22  05/04/2022
            Deep Randomized Networks - Guest lecture by Claudio Gallicchio
            reservoir computing; randomized models; echo state networks
             23  06/04/2022
            Advanced Recurrent Architectures and Attention
            sequence-to-sequence;  attention models; multiscale network; hierarchical models
             [DL] Sections 10.12, 12.4.5  Additional Readings
            [34,35] Models of sequence-to-sequence and image-to-sequence transduction with attention
            [36,37] Models optimizing dynamic memory usage (clockwork RNN, zoneout)
            [38] Transformer networks: a paper on the power of attention without recurrence
             24  07/04/2022
            Neural Reasoning
            memory networks; neural Turing machines
               Additional Readings
            [39] Differentiable memory networks
            [40,41] Neural Turing Machines and follow-up paper on pondering networks
             25  12/04/2022
            Unsupervised and Generative Deep Learning I
            explicit distribution models; neural ELBO; variational autoencoders
            [DL] Sections 20.9, 20.10.1-20.10.3  Additional Readings
            [42] PixelCNN - Explict likelihood model
            [43] Tutorial on VAE

             26  13/04/2022
            Unsupervised and Generative Deep Learning II
            generative adversarial networks; adversarial autoencoders

            [DL] Section 20.10.4
             Additional Readings
            [44] Tutorial on GAN (here another online resource with GAN tips)
            [45] Wasserstein GAN
            [46] Tutorial on sampling neural networks
            [47] Progressive GAN
            [48] Cycle Gan
            [49] Seminal paper on Adversarial AEs


            • Reinforcement Learning (14h)

              We formalise the reinforcement learning problem by rooting it into Markov decision processes and we provide an overview of the main approaches to design reinforcement learning agents, including model-based, model-free, value and policy learning. We link classical approaches with modern deep learning based approximators (deep reinforcement learning). We overview the main programming frameworks available. Methodologies covered include: dynamic programming, MC learning, TD learning, SARSA, Q-learning, deep Q-learning, policy gradient and deep policy gradient, MC tree search. 

              Date Topic  References  
               Additional Material 
               28 21/04/2022
              Reinforcement learning fundamentals
              reinforcement learning problems; environment; agent; actions and policies; taxonomy of approaches
               [RL] Chapter 1 Software
              Open AI gym for RL environments and tasks
               29 26/04/2022
              Markov Decision Processes
              formal model of RL probelms; rewards; returns; Bellman expectation and optimality
               [RL] Chapter 3  
               30 27/04/2022
              Model-Based Planning
              dynamic programming; policy evaluation; policy iteration; value iteration
               [RL] Chapter 4
              Dynamic programming demo on Gridworld in Javascript (with code)
               31 28/04/2022
               Model-free reinforcement learning
              model-free predition; model-free control; Monte Carlo methods; TD learning;  SARSA; Q-learning
              [RL] Section 5.1-5.6, 6.1-6.6, 7.1, 7.2, 12.1, 12.2, 12.7
               Additional reading:
              [50] The original Q-learning paper


               NO LECTURE    
              NO LECTURE    
               32 05/05/2022
               Value-function Approximation
              linear incremental methods; batch value function approximation; deep Q-learning; linear least-squares control
               [RL] Section 9.1-9.5, 9.8, 10.1, 11.1-11.5 Additional Reading:
              [51] Original DQN paper
              [52] Double Q-learning
              [53] Dueling Network Architectures
              [54] Prioritized Replay
               33 09/05/2022
              (16-18 - ROOM C
               Policy gradient methods  [RL] Chapter 13  Additional Reading:
              [55] Original REINFORCE paper
              [56] Learning with the actor-critic architecture
              [57] Accessible reference to natural policy gradient
              [58] A3C paper
              [59] Deep Deterministic Policy Gradient
              [60] TRPO paper
              [61] PPO paper
               34 10/05/2022
               Integrating Learning and Planning  [RL] Chapter 8, Sect 16.6  Additional Reading:
              [62] UCT paper: the introduction of Monte-Carlo planning
              [63] MoGo: the grandfather of AlphaGo (RL using offline and online experience)
              [64] AlphaGo paper
              [65] AlphaGo without human bootstrap

              • Advanced Topics and Applications (8h)

                The module covers some recent and interesting development and research topics in the field of machine learning. Topics choice is likely to vary at each edition. Example topics include: deep learning for graphs, learning with structured data, continual learning, distributed learning, learning-reasoning integration, edgeAI,etc.. The module concludes with a final lecture which discusses the course content retrospectively and details the exam modalities, topics and deadlines.

                   Date  Topic  References     Additional Material 
                 27  20/04/2022
                 Continual Learning - Guest lecture by Vincenzo Lomonaco

                 35 11/05/2022

                Deep learning for graphs
                - PyDGN: our in-house DLG library
                - PyTorch geometric
                - Deep graph library

                Additional readings
                [66-67] Seminal works on neural networks for graphs
                [68] Recent tutorial paper
                Final lecture

                (11-13) ROOM D1
                Research seminars by Ph.D. students

                Draft programme:
                Andrea Cossu - The reasonable effectiveness of pre-trained models in Continual Learning
                Michele Resta - Continual Incremental Language Learning for Neural Machine Translation
                RIccardo Massidda - Ontology-Driven Semantic Alignment of Artificial Neurons
                Danilo Numeroso - Neural Algorithmic Reasoning
                Francesco Landolfi - Graph Pooling with Maximum Weight k-Independent Sets

                • Course Grading and Exams

                  Typical course examination (for students attending the lectures) is performed in 2 stages: midterm assignments and an oral exam. Midterms waive the final project.

                  Midterm Assignment

                  Midterms consist in short assignments involving with one of the following tasks:

                  • A quick and dirty (but working) implementation of a simple pattern recognition algorithm
                  • A report concerning the experience of installing and running a demo application realized using available deep learning and machine learning libraries
                  • A summary of a recent research paper on topics/models related to the course content.

                  The midterms can consist in either the delivery of code (e.g. colab notebook) or a short slide deck (no more than 10 slides) presenting the key/most-interesting aspects of the assignment. 

                  Students might be given some amount of freedom in the choice of assignments, pending a reasonable choice of the topic. The assignments will roughly be scheduled every 3/4 weeks.

                  Oral Exam

                  The oral examination will test knowledge of the course contents (models, algorithms and applications).

                  Exam Grading (with Midterms)

                  The final exam vote is given by the oral grade. The midterms only wave the final project but do not contribute to the grade. In other words you can only fail or pass a midterm. You need to pass all midterms in order to succesfully wave the final project.

                  Alternative Exam Modality (No Midterms / Non attending students)

                  Working students, those not attending lectures, those who have failed midterms or simply do not wish to do them, can complete the course by delivering a final project and an oral exam.  Final project topics will be released in the final weeks of the course: contact the instructor by mail to arrange choice of the topics once these are published.

                  The final project concerns preparing a report on a topic relevant to the course content or the realization of a software implementing a non-trivial learning model and/or a PR application relevant for the course. The content of the final project will be discussed in front of the instructor and anybody interested during the oral examination. Students are expected to prepare slides for a 15 minutes presentation which should summarize the ideas, models and results in the report. The exposition should demonstrate a solid understanding of the main ideas in the report.

                  Grade for this exam modality is determined as

                   \( G = 0.5 \cdot (G_P + G_O) \)

                  where \( G_P \in [1,32] \) is the project grade and \( G_O \in [1,30] \) is the oral grade

                  • Midterms and Projects