Lectures

Lectures


Reading the Genome Sequence using Neural Networks: Introduction
Reading the Genome Sequence using Neural Networks: Improving Resolution, Interpretability and Accuracy


Mathematics for Deep Learning 2/2
Mathematics for Deep Learning 1/2


Introduction to Deep Learning 1/6

Introductory course on deep learning methods and algorithms.

Legend
T: theory (slides and animations)
P: practice (Jupyter Notebooks)
Introduction and motivation.
T) Learning paradigms: supervised, unsupervised, and reinforcement learning
P) Getting started with the tools: Jupyter notebook, PyTorch tensors and auto differentiation
Introduction to Deep Learning 2/6

Introductory course on deep learning methods and algorithms.

Legend
T: theory (slides and animations)
P: practice (Jupyter Notebooks)
Classification and regression.
T+P) Neural net’s forward and backward propagation for classification and regression
Introduction to Deep Learning 3/6

Introductory course on deep learning methods and algorithms.

Legend
T: theory (slides and animations)
P: practice (Jupyter Notebooks)
Energy-based models.
T) Latent variable generative energy-based models (LV-GEBMs) part I: foundations
Introduction to Deep Learning 4/6

Introductory course on deep learning methods and algorithms.

Legend
T: theory (slides and animations)
P: practice (Jupyter Notebooks)
Geometric deep learning (grid and set).
T+P) Convolutional neural nets improve performance by exploiting data nature
Introduction to Deep Learning 5/6

Introductory course on deep learning methods and algorithms.

Legend
T: theory (slides and animations)
P: practice (Jupyter Notebooks)
Geometric deep learning (grid and set).
T+P) Recurrent nets natively support sequential data
T+P) Self/cross and soft/hard attention: a building block for learning from sets
Introduction to Deep Learning 6/6

Introductory course on deep learning methods and algorithms.

Legend
T: theory (slides and animations)
P: practice (Jupyter Notebooks)
Generative models.
T+P) LV-GEBMs part II: autoencoders, adversarial nets


Voting: axiomatic and algorithmic challenges
The goal of this lecture is to offer a general introduction to preference aggregation, desiderata for voting rules and computational complexity of classic preference aggregation problems.
Voting over restricted preference domains
In this lecture we discuss domain restrictions that make voting-related problems computationally tractable and avoid some of the classic social choice paradoxes. We focus on the classic domain on single-peaked preferences and consider the complexity of preference elicitation and learning the preference structure.


Towards Developmental Machine Learning

By and large, most studies of machine learning and pattern recognition are rooted in the framework of statistics. This is primarily due to the way machine learning is traditionally posed, namely by a problem of extraction of regularities from a sample of a probability distribution. This lecture promotes a truly different way of interpreting the learning of that relies on system dynamics. We promote a view of learning as the outcome of laws of nature that govern the interactions of intelligent agents with their own environment.This leads to an in-depth interpretation of causality along with the definition of the principles and the methods for learning to store events without their long-term forgetting that characterize state of the art technologies in recurrent neural networks. Finally, we reinforce the underlying principle that the acquisition of cognitive skills by learning obeys information-based laws based on variational principles, which hold regardless of biology.



State representation learning and evaluation in robotic interaction tasks

Efficient representations of observed input data have been shown to significantly accelerate the performance of subsequent learning tasks in numerous domains. To obtain such representations automatically, we need to design both i) models that identify useful patterns in the input data and encode them into structured low dimensional representations, and ii) evaluation measures that accurately assess the quality of the resulting representations. We present work that addresses both these requirements. We present a short overview of representation learning techniques and different structures that can be imposed on representation spaces.  We show into how these can be applied in complex robotics tasks considering physical interaction with the environment.



AI as a Tool for Investigating Human Intelligence 1/2

Artificial intelligence has a great potential to uncover the underlying mechanisms of human intelligence. Neural networks inspired by the brain can simulate how humans acquire cognitive abilities and thus reveal what enables/disables cognitive development. My lecture introduces a neuroscience theory called predictive coding. We have been designing neural networks based on predictive coding and investigating to what extent the theory accounts for cognitive development. The key idea is that the brain works as a predictive machine and perceives the world and acts on it to minimize prediction errors. Our robot experiments demonstrate that the process of minimizing prediction errors leads to sensorimotor and social cognitive development and that aberrant predictive processing produces atypical development such as developmental disorders. We discuss how these findings facilitate the understanding of human intelligence and provide a new principle for cognitive development.

AI as a Tool for Investigating Human Intelligence 2/2

Artificial intelligence has a great potential to uncover the underlying mechanisms of human intelligence. Neural networks inspired by the brain can simulate how humans acquire cognitive abilities and thus reveal what enables/disables cognitive development. My lecture introduces a neuroscience theory called predictive coding. We have been designing neural networks based on predictive coding and investigating to what extent the theory accounts for cognitive development. The key idea is that the brain works as a predictive machine and perceives the world and acts on it to minimize prediction errors. Our robot experiments demonstrate that the process of minimizing prediction errors leads to sensorimotor and social cognitive development and that aberrant predictive processing produces atypical development such as developmental disorders. We discuss how these findings facilitate the understanding of human intelligence and provide a new principle for cognitive development.



Computational Approaches for Solving Systems of Nonlinear Equations
Finding one or more solutions to a system of nonlinear equations (SNE) is a
computationally hard problem with many applications in sciences, engineering, machine learning and artificial intelligence.
First, we will briefly discuss classical approaches for addressing (SNE).
Then, we will discuss the various ways that a SNE can be transformed
into an optimization problem, and we will introduce techniques that can be utilized to search for solutions to the global optimization problem that arises when the most common reformulation is performed.
In addition, we will present computational results using different heuristics.


Automated Machine Learning: the state of the art

Automated machine learning is the science of learning how to build machine learning models in a data-driven, efficient, and objective way. It replaces manual (and often frustrating) trial-and-error with automated, principled processes. It also democratizes machine learning, allowing many more people to build high-quality machine learning systems.

In the first lecture, we will explore the state of the art in automated machine learning. We will cover the best techniques for neural architecture search, as well as learning complete machine learning pipelines. We explain how to design model search spaces, and how to efficiently search for the best models within this space. We’ll also cover useful tips and tricks to speed up the search for good models, as well as pitfalls and best practices.

Automated Machine Learning: learning to learn

In the second lecture, we’ll cover techniques to continually learn how to build better machine learning models. Just as human experts get ever better at building better models, automated machine learning systems should also get better every time they run. We’ll cover research on the intersection of automated machine learning, meta-learning, and continual learning that enables us to learn and capture which models work well, and transfer that knowledge to build better machine learning models, faster.




 

Tutorials


Statistical Physics view of theory of Machine Learning 1/2
The past decade has witnessed a surge in the development and adoption of machine learning algorithms to solve day-a-day computational tasks. Yet, a solid theoretical understanding of even the most basic tools used in practice is still lacking, as traditional statistical learning methods are unfit to deal with the modern regime in which the number of model parameters are of the same order as the quantity of data – a problem known as the curse of dimensionality. Curiously, this is precisely the regime studied by Physicists since the mid 19th century in the context of interacting many-particle systems. This connection, which was first established in the seminal work of Elisabeth Gardner and Bernard Derrida in the 80s, is the basis of a long and fruitful marriage between these two fields.
 
In this tutorial I will motivate and review the connections between Statistical Physics and problems in high-dimensional Statistics, such as the ones stemming from Machine Learning. My goal is to introduce the students to this rich literature with an overview of the tools from Statistical Physics which have found fruitful applications in ML theory, and cover at least one detailed application.
Statistical Physics view of theory of Machine Learning 2/2
The past decade has witnessed a surge in the development and adoption of machine learning algorithms to solve day-a-day computational tasks. Yet, a solid theoretical understanding of even the most basic tools used in practice is still lacking, as traditional statistical learning methods are unfit to deal with the modern regime in which the number of model parameters are of the same order as the quantity of data – a problem known as the curse of dimensionality. Curiously, this is precisely the regime studied by Physicists since the mid 19th century in the context of interacting many-particle systems. This connection, which was first established in the seminal work of Elisabeth Gardner and Bernard Derrida in the 80s, is the basis of a long and fruitful marriage between these two fields.
 
In this tutorial I will motivate and review the connections between Statistical Physics and problems in high-dimensional Statistics, such as the ones stemming from Machine Learning. My goal is to introduce the students to this rich literature with an overview of the tools from Statistical Physics which have found fruitful applications in ML theory, and cover at least one detailed application.


PyTorch 1/5

Introduction.

PyTorch 2/5

Introduction.

PyTorch 3/5

Optimization.

PyTorch 4/5

Optimization.

PyTorch 5/5

Deployment.