Ha Machine Learning Summer School | Buenos Aires

Tentative Schedule


Monday 06/18   Tuesday 06/19 Wednesday 06/20 Thursday 06/21
Morning 8:20 - 9:00 Registration      
Morning 9:00-9:30 Opening Morning 9-10:30 Adams -3 Broderick - 2 Doshi-Velez - 1
Morning 9:30-11:00 Candes -1   Coffee Break Coffee Break Coffee Break
  Coffee Break Morning 11-12:30 Gonzalez - 1 Broderick - 3 Doshi-Velez - 2
Afternoon 11:30-12:45 Candes -2 Afternoon 2:00-3:30 Gonzalez - 2 Cuturi - 1 Doshi-Velez - 3
Afternoon 2:00-3:30 Candes -3   Coffee Break Coffee Break Coffee Break
  Coffee Break Afternoon 4:00-5:00 Gonzalez - 3 Cuturi - 2 Break
Afternoon 4:00-5:30 Adams - 1 Night 5:00-6:00 Broderick - 1 Cuturi - 3 Levine - 1
Night 5:45-7:00 Adams - 2 Night 600-6:30 A - 9 Social Event (Transportation to the venue) Poster - Session
Friday 06/22 Saturday 06/23
Morning 9-10:30 Levine - 2 Morning 9:00 - 9:30 A - 6
  Coffee Break Morning 9:30-10:00 A - 7
Morning 11-12:30 Levine - 3 Morning 10:00-10:30 A - 8
Afternoon 2:00-3:00 A - 1   Coffee Break
  Coffee Break Morning 11:00 - 11:30 TBA. Sadosky Foundation.
Afternoon 3:30-4:15 A - 2 Noon 11:30 - 1:00 PM Delteil - 1
Afternoon 4:15 - 5:00 A - 3
Afternoon 5:00-5:45 A - 4
Night 5:45 - 6:00 A - 5
Night 6:00 - 7:00 Poster - Session
Monday 06/25   Tuesday 06/26 Wednesday 06/27 Thursday 06/28 Friday 06/29 Saturday 06/30
Morning 9:00-10:30 Larochelle - 1 Blei - 1 Blei - 3 Warde-Farley - 1 Warde-Farley - 3 Hsu - 3
  Coffee Break Coffee Break Coffee Break Coffee Break Coffee Break Coffee Break
Morning 11-12:30 Larochelle - 2 Blei - 2 Wood - 3 Warde-Farley - 2 Hazan - 2 Pfau - 2
Afternoon 2:00-3:30 Larochelle - 3 Wood - 1 Abadi - Isard Hazan - 1 Hazan - 3 Closing - 12:30 - 1:00
  Coffee Break Coffee Break Coffee Break Coffee Break Coffee Break  
Afternoon 4:00-5:30 Delteil - 2 Break Abadi - Isard Pfau - 1 Hsu - 1  
Night 5:45-7:00 Delteil - 3 Wood - 2   Diversity Hsu - 2  
Night 7:00-8:00     Social Event      

TALK 1

Sergey Levine
Deep Reinforcement Learning
Abstract:
I will discuss reinforcement learning and control algorithms that combine high-dimensional parametric models, such as deep neural networks, with decision making and control. In particular, the lectures will cover policy gradient, value function-based, and actor-critic algorithms for reinforcement learning with function approximation, model-based reinforcement learning, and a number of advanced topics, which may include: the connection between control and probabilistic inference, inverse reinforcement learning, transfer, multi-task, and meta-learning for control, and applications in areas such as robotics.
Slides
Daniel Hsu
Learning latent variable models using tensor decompositions
This tutorial surveys algorithms for learning latent variable models based on the method-of-moments, focusing on algorithms based on low-rank decompositions of higher-order tensors. The target audiences of the tutorial include (i) users of latent variable models in applications, and (ii) researchers developing techniques for learning latent variable models. The only prior knowledge expected of the audience is a familiarity with simple latent variable models (e.g., mixtures of Gaussians), and rudimentary linear algebra and probability. The audience will learn about new algorithms for learning latent variable models, techniques for developing new learning algorithms based on spectral decompositions, and analytical techniques for understanding the aforementioned models and algorithms. Advanced topics such as learning overcomplete representations may also be discussed.
Slides
Martin Abadi, Michael Isard
Tensorflow
TensorFlow is a machine learning system that operates at large scale and in heterogeneous environments. Its computational model is based on dataflow graphs with mutable state. Graph nodes may be mapped to different machines in a cluster, and within each machine to CPUs, GPUs, and custom ASICs. TensorFlow supports a variety of applications, but it particularly targets training and inference with deep neural networks. It serves as a platform for research and for deploying machine learning systems across many areas, such as speech recognition, computer vision, robotics, information retrieval, and natural language processing. In these lectures, we will describe TensorFlow's programming models, some aspects of its implementation, and some of the underlying theory. TensorFlow is joint work with many other people in the Google Brain team and elsewhere. More information is available at tensorflow.org.
Some of the material is based on the following papers: https://dl.acm.org/citation.cfm?id=3026899, https://dl.acm.org/citation.cfm?id=3088527, and https://dl.acm.org/citation.cfm?id=3190551
Elad Hazan
Optimization for Machine Learning
In this tutorial we'll survey the optimization viewpoint to learning. We will cover optimization-based learning frameworks, such as online learning and online convex optimization. These will lead us to describe some of the most commonly used algorithms for training machine learning models.
Slides
Marco Cuturi
A Primer on Optimal Transport
Optimal transport (OT) provides a powerful and flexible way to compare probability measures, discrete and continuous, which includes therefore point clouds, histograms, datasets, parametric and generative models. Originally proposed in the eighteenth century, this theory later led to Nobel Prizes for Koopmans and Kantorovich as well as Villani’s Fields Medal in 2010. OT recently has reached the machine learning community, because it can tackle challenging learning scenarios including dimensionality reduction, structured prediction problems that involve histogram outputs, and estimation of generative models such as GANs in highly degenerate, high-dimensional problems. Despite very recent successes bringing OT from theory to practice, OT remains challenging for the machine learning community because of its mathematical formality. This tutorial will introduce in an approachable way crucial theoretical, computational, algorithmic and practical aspects of OT needed for machine learning applications.
Slides
Javier Gonzalez Hernandez
Gaussian processes for Uncertainty Quantification
In these three lectures we will cover different theoretical and practical aspects of Gaussian processes (GPs) and how they can be used for making decisions under uncertainty. The first lecture will be an introduction to GPs. We will review the basic concepts of GPs, explore some connections with other related techniques and show how to use them in practice. The second lecture will present different ways in which GPs can be used for decision making. We will focus on cases in which taking into account uncertainty coming from the predictions of the GPs is a key element. In particular, we will describe how GPs can be used for optimization, quadrature and experimental design. In the third lecture students will take part in a hands-on lab, using different Python libraries to see these methods at work. There are no pre-requisites for the lectures but a general background in machine learning is recommended. For the practical sessions a laptop with some version of Python (preferably Conda) is desirable.
Slides
Ryan Adams
A Tutorial on Deep Probabilistic Models
I will give a tutorial on the interface between probabilistic modeling and deep neural networks.  The three primary topics of interest will be Bayesian neural networks, Boltzmann machines, and neural density models such as variational autoencoders.  I will provide an introduction to inference and learning in these models, and give an overview that connects modern approaches to the long history of probabilistic modeling with neural function approximation.
Slides
Tamara Broderick
Nonparametric Bayesian Methods: Models, Algorithms, and Applications
This tutorial introduces Bayesian nonparametrics (BNP) as a tool for modern data science and machine learning. BNP methods are useful in a variety of data analyses---including density estimation without parametric assumptions and clustering models that adaptively determine the number of clusters. We will demonstrate that BNP allows the data analyst to learn more from a data set as the size of the data set grows and see how this feat is accomplished. We will study popular BNP models such as the Dirichlet process, Chinese restaurant process, Indian buffet process, and hierarchical BNP models---and how they relate to each other.
Slides
David Blei
Variational Inference: Foundations and Innovations
One of the core problems of modern statistics and machine learning is to approximate difficult-to-compute probability distributions. This problem is especially important in probabilistic modeling and Bayesian statistics, which frame all inference about unknown quantities as calculations about conditional distributions. In this tutorial I will review and discuss variational inference (VI), a method a that approximates probability distributions through optimization. VI has been used in myriad applications in machine learning and tends to be faster than more traditional methods, such as Markov chain Monte Carlo sampling. I will first review the basics of variational inference. Then I will describe some of the pivotal tools for VI that have been developed in the last few years: Monte Carlo gradients, black box variational inference, stochastic variational inference, and variational autoencoders. Last, I will discuss some of the unsolved problems in VI and point to promising research directions.
Slides
Hugo Larochelle
Deep Neural Networks
In this lecture, I'll start by covering the basic concepts behind feedforward neural networks. I'll present forward propagation and backpropagation in neural networks. Specifically, I'll discuss the parameterization of feedforward nets, the most common types of units, the capacity of neural networks and how to compute the gradients of the training loss for classification with neural networks. I'll discuss the training of neural networks by gradient descent and then discuss the more recent ideas that are now commonly used for training deep neural networks. Then, I'll discuss various types of neural network architectures, designed to address a variety of learning problems (supervised, multi-task, one-shot, zero-shot, meta-learning). Finally, I'll end with a discussion of some of the intriguing properties of neural networks that are the object of a lot of research today.
Slides
Finale Doshi-Velez
Introduction to Reinforcement Learning (+ bonus material)
We will begin with an introduction to fundamental concepts in reinforcement learning: policies, value functions, and planning in discrete environments.  Armed with this groundwork, we'll do some hands-on practicals to gain an intuition of these fundamental concepts/understand how even simple, discrete environments can exhibit interesting subtleties.   In the final session, we'll dive into a very specific use-case of reinforcement learning known as off-policy evaluation: if we have already collected data under one policy, can we reuse that data to guess how well a different decision-making strategy will perform?  I'll link this to some of our current work in applying reinforcement learning in healthcare.
Slides
Frank Wood
Probabilistic Programming
How do we engineer machines that reason? This is a question that has long vexed humankind. The answer to this question is fantastically valuable. There exist various hypotheses. One major division of hypothesis space delineates along lines of assertion: that random variables and probabilistic calculation are more-or-less an engineering requirement [Ghahramani, 2015, Tenenbaum et al., 2011] and the opposite [LeCun et al., 2015, Goodfellow et al., 2016]. The field ascribed to the former camp is roughly known as Bayesian or probabilistic machine learning; the latter as deep learning. The first requires inference as a fundamental tool; the latter optimization, usually gradient-based, for classification and regression. Probabilistic programming languages are to the former as automated differentiation tools are to the latter. Probabilistic programming is fundamentally about developing languages that allow the denotation of inference problems and evaluators that “solve” those inference problems. It can be argued that the rapid exploration of the deep learning, big-data/regression approach to artificial intelligence has been triggered largely by the emergence of programming languages tools that automate the tedious and troublesome derivation and calculation of gradients for optimization. Probabilistic programming aims to build and deliver a toolchain that does the same for probabilistic machine learning; supporting supervised, unsupervised, and semi-supervised inference. Without such a toolchain one could argue that the complexity of inference-based approaches to artificial intelligence systems are too high to allow rapid exploration of the kind we have seen recently in deep learning. These lectures introducing probabilistic programming will cover the basics of probabilistic programming from language design to evaluator implementation with the dual aim of explaining existing systems at a deep enough level that students should have no trouble adopting and using any of both the languages and systems that are currently out there and making it possible for the next generation of probabilistic programming language designers and implementers to use this as a foundation upon which to build.
Slides
Thomas Delteil
Learning Deep Learning from Scratch with MXNet/Gluon
This lecture introduces MXNet Gluon, a flexible new imperative interface that pairs MXNet’s speed with a user-friendly frontend. It allows users to seamlessly transition from imperative code to a symbolic graph representation. This allows for faster execution and deployment on a wide range of devices, including embedded ones. In the first part of the lecture, we will cover the fundamentals of Gluon: NDArray data structure, Block layer and automatic differentiation. We will show how to define neural networks at the atomic level and through Gluon’s predefined layers. We will demonstrate how to load data asynchronously, serialize models and build dynamic graphs. In the second part, we will focus on a specific deep learning task: semantic segmentation. We will show how you can implement and train a Fully Convolutional Network (FCN) to segment an image according to a set of classes. In the final part, we will go through the GluonCV (Computer Vision) and GluonNLP (Natural Language Processing) toolkits. You will learn how to leverage these two libraries to quickly replicate the results of state-of-the-art models in several tasks.
Slides
David Pfau
Discovering the Geometry of Data Manifolds with Spectral and Deep Learning
A central tenet of machine learning is that complex high-dimensional data can be described by the structure of a low-dimensional latent manifold. However, most machine learning methods do little to exploit the rich toolkit of techniques for analyzing these manifolds. In this tutorial I will give a tour of how ideas from differential geometry and spectral analysis can be brought to bear on problems in machine learning. We will cover basic ideas in differential geometry such as curvature, metrics and geodesics, and go over how they relate to problems in spectral theory like Laplacian operators and computing low-rank matrix decompositions. We will survey applications to machine learning, including recent works on generalizing convolutional neural networks to graph- and manifold-structured input, analyzing the structure of latent spaces in deep generative models, and embedding hierarchical structure in continuous spaces. Finally, I will discuss spectral inference networks, a framework for unsupervised learning that uses the algorithmic tools of deep learning and stochastic optimization to solve large-scale spectral decomposition problems that would otherwise be intractable.
Slides
Emmanuel Candes
Modern Approaches to False Discovery Rate Control and Inference in High Dimensional Models.
Generative adversarial networks (GANs) are a recently developed approach to learning generative models, in particular generative models capable of highly realistic synthesis. This tutorial will explore the GAN approach to generative models, comparing and contrasting it with more traditional generative model paradigms as well as other modern approaches based on neural networks. We will explore the practical considerations for effectively stabilizing GAN learning dynamics, as well as approaches to quantitatively evaluating the resulting models, and the current set of challenges at the frontiers of GAN and generative model research. We will also motivate the study of GANs from the perspective of successful applications to date, including domain adaptation, image-to-image translation, and single-image super-resolution.
Slides
David Warde-Farley
Generative Adversarial Networks
Generative adversarial networks (GANs) are a recently developed approach to learning generative models, in particular generative models capable of highly realistic synthesis. This tutorial will explore the GAN approach to generative models, comparing and contrasting it with more traditional generative model paradigms as well as other modern approaches based on neural networks. We will explore the practical considerations for effectively stabilizing GAN learning dynamics, as well as approaches to quantitatively evaluating the resulting models, and the current set of challenges at the frontiers of GAN and generative model research. We will also motivate the study of GANs from the perspective of successful applications to date, including domain adaptation, image-to-image translation, and single-image super-resolution.

TALK 2

A - 1
Recommendations in the largest e-commerce platform of Latin America.
Creating a product from scratch faces many different challenges from both business perspective and algorithmic/infrastructure complexity. In this presentation, we will share how the largest e-commerce platform of LATAM develops its recommender engine from its birth in 2017 up to present day.
Speakers: Pablo Zivic ML Researcher, Martin Pozzer Senior Product Development Manager. Mercado Libre.
A - 2
AI: From Inception to Production
The recent developments in artificial intelligence (AI), supported by the analysis and exploitation of large data sets, open up a new era in which the application of AI promises the creation of innovative products and services. However, even though major players in the technology marketplace are creating software that takes AI out of the laboratories and makes it more accessible, the process of incorporating AI to the products or processes of a company still poses great challenges. Our aim in this talk is to present, drawing from our own experience, the process of carrying out an AI project from its inception, passing through the stages of development and quality control, until its eventual realization.
Speakers: Tomás Tecce, Pasquinel Ubani. Globant.
A - 3
IQA @ Despegar
What is a good quality photo?  During this talk we are going to discuss possible answers to this question by diving into the world of Image Quality Assesment. We will present the state of the art in Machine Learning technics applied to this field as well as industry caveats in order to turn prototypes into performant and scalable software
Speakers: Pedro Carossi y Alejandro Alvarez. Despegar.
A - 4
ML at OLX: Disrupting the Future of Classifieds World"
Machine learning (ML) technologies are becoming key differentiators for companies across all industries. Some of the biggest two-sided marketplace platforms in the world such as Uber, Airbnb and OLX use ML to deliver disruption in transportation, housing and classifieds marketplaces respectively. Founded in 2006, OLX is a global company that operates in more than 40 countries and provides one of the largest online classifieds marketplaces in the world. At the moment, OLX’s ML-powered platform facilitates matching buyers and sellers by delivering personalized relevant content to more than 300 million users monthly. In this talk I will introduce challenging and unique problems that we encounter while building recommendation, personalization and search systems at this scale at OLX and how we are tackling them using cutting-edge ML technologies.
Speakers: Vladan Radosavljevic, Ph.D., Head of Data Science at OLX.
A - 5
Next - Ai Canada
Are you interested in commercializing your research and starting a business? NextAI is Canada’s premiere AI startup accelerator located in Toronto and Montreal - two global hotspots for AI research and commercialization. NextAI is for entrepreneurs, researchers and scientists looking to launch AI-enabled ventures.
Speakers: Jon French.
A - 6
Machine Learning and Earth Observation
Satellogic aims to capture every square meter of the Earth's surface to derive insights and enable better decision making for industries, governments, and individuals. Addressing some of humanity’s most pressing challenges, such as providing food or distributing energy for nine billion people without depleting resources, requires real-time planet-scale data. Satellogic has created small, inexpensive satellites, that transmit real-time data and images back home. Our constellation already has several high-resolution satellites in orbit and is growing to 300 in the next few years to provide new insights about our planet. The Satellogic data science team works with our satellite images and other data sources in order to transform this data into knowledge. In this talk we will explain how our company has designed from scratch a fleet of satellites that cost 1000 times less than conventional earth observation satellites and what are the main challenges that our machine learning engineers face when developing remote sensing solutions through real cases.
A - 7
A brief survey of Data Science techniques applied to the analysis of Bank and Mobile Phone Datasets.
Unified Machine Learning Approaches for Heterogenous Data
Speakers: Charles Sarraute. Grandata.
A - 8
Unified Machine Learning Approaches for Heterogenous Data
Speakers: David Stevens and Santiago Hernandez. Jampp.
A - 9
Smart Learning for fraud prevention
How is the best way to investigate fraud? What happens when we find a new pattern or modus operandi? After the investigation of a fraud, and when we have found the most relevant that can impact the organization.It is necessary to be clear in what way and what information is vital to be able to feed our learning engine.
Speakers: Daniel Guzman. IBM.
Diversity - Panel
Discussion Session: Diversity and Inclusion in ML
Top local speakers will discuss main initiatives, outstanding projects and their personal experiences in mixed teams focusing on the importance of diversity to foster a more ethic ML. They will tell us about how to emphasize a more inclusive Machine Learning, Data Science and Information and Communications Technology focusing on Argentina and the region.
Speakers: Delfina Daglio (IBM Argentina), Fernando Schapachnik (Fundación Sadosky) and Luciana Ferrer (ICC UBA-CONICET). Moderator: Guadalupe Dorna.
Computational Challenges
Sadosky Fundation
Speakers: Lenadro Lombardi. Sadosky Foundation.