Reinforcement learning, computer vision and AGI!
RAIL Lab and PRIME are co-hosting an IndabaX South Africa Roadshow event and you are cordially invited to join us! RAIL Lab and PRIME are two machine learning research groups from the Computer Science and Applied Math (CSAM / "kazam!") department at the University of the Witwatersrand.
RAIL Lab focuses primarily on learning in autonomous systems. In particular, we are interested in the acquisition of behaviours, as well as knowledge about the environment around a learning system. Our work draws on tools from multiple fields including decision theory, machine learning, and computer vision, using techniques including reinforcement learning, Bayesian models, deep neural networks, and Monte Carlo tree search.
The PRIME Lab is focused on problems in deep learning for computer vision, representation learning, and learning in situations where data and labels are scarce. We have projects based on computer vision, natural language processing, self-supervised and weakly-supervised learning, generative modelling, and latent space analysis; as well as in energy, education, and healthcare.
When: 17 Sept 2021 from 14h00 - 17h00 SAST (convert to your timezone)
Where: We'll be kicking off on Zoom HERE
SCHEDULE:
- RAIL and PRIME Intro talks (14h00 to 14h20)
- Prof Benjamin Rosman
- Dr Richard Klein
- Panel: "What might we still require to achieve AGI?" (14h20 to 15h00)
- Prof Christopher Cleghorn
- Geraud Nangue Tasse
- Dr Helen Robertson
- Dr Chris Fourie
- Parallel Tutorials (pick one to attend) (15h00 to 16h00)
- An Introduction to Reinforcement Learning (Geraud Nangue Tasse)
- An Introduction to Physical Reasoning Benchmarks (Divanisha Pillay)
- Google Earth Engine and its Opportunities for Geospatial Machine Learning (Geethen Singh)
- A How-To Guide on Getting Started with RoboCup's 3D Soccer Simulation (Branden Ingram)
- Learning without Labels: An Introduction to Visual Self-Supervised Representation Learning (David Torpey)
- Keynote Talk: "Signal to Symbol (via Skills)" (16h00 - 17h00)
- Prof George Konidaris (Brown University)
Keynote Talk
Prof George Konidaris
The John E. Savage Assistant Professor of Computer Science and director of the Intelligent Robot Lab at Brown University, which forms part of bigAI (Brown Integrative, General AI). I am also the Chief Roboticist of Realtime Robotics, a startup based on our research on robot motion planning. My research aims to build intelligent, autonomous, general-purpose robots that are generally capable in a wide variety of tasks and environments. I focus on understanding how to design agents that learn abstraction hierarchies that enable fast, goal-oriented planning. I develop and apply techniques from machine learning, reinforcement learning, optimal control and planning to construct well-grounded hierarchies that result in fast planning for common cases, and are robust to uncertainty at every level of control. I believe that it will take advances in all of these areas, and additionally advances in how to integrate these areas, to solve the AI problem.
Intro Talks
Prof Benjamin Rosman
Associate Professor in the School of Computer Science and Applied Mathematics at the University of the Witwatersrand, where he runs the Robotics, Autonomous Intelligence and Learning (RAIL) Laboratory and is the Director of the National E-Science Postgraduate Teaching and Training Platform (NEPTTP). He received his Ph.D. in Informatics (2014), and his M.Sc. in Artificial Intelligence (2009), both from the University of Edinburgh, UK. His research interests focus primarily on reinforcement learning and decision making in autonomous systems. He is a founder and organiser of the Deep Learning Indaba, a recipient of a Google Faculty Research Award (2017), and a Senior Member of the IEEE.
Dr Richard Klein
Senior Lecturer in the School of Computer Science and Applied Mathematics at the University of the Witwatersrand where he runs the PRIME Lab. He is the coordinator of the Undergraduate BSc in Computer Science as well as the MSc in Data Science and MSc in Artificial Intelligence. He received his Ph.D in Computer Science in 2017 and MSc in 2013. His research interests are in Deep Learning and Computer Vision, with a particular interest in techniques for model training in areas with data scarcity.
Panelists
Prof Christopher Cleghorn
Received his Masters and PhD degrees in Computer Science from the University of Pretoria, South Africa, in 2013 and 2017 respectively. He is an Associate Professor in the School of Computer Science and Applied Mathematics at the University of the Witwatersrand. His research interests include swarm intelligence, evolutionary computation, machine learning, and radio-astronomy with a strong focus of theoretical research. Prof Cleghorn annually serves as a reviewer for numerous international journals and conferences in domains ranging from swarm intelligence and neural networks to mathematical optimization.
Dr Helen Robertson
Lecturer in the School of Computer Science and Applied Mathematics at the University of the Witwatersrand. She received her PhD in Philosophy from University College London (UCL) in 2017, with prior postgraduate studies at the University of the Witwatersrand. She currently teaches the applied ethics of data science and has research interests in the epistemology and ethics of artificial intelligence.
Dr Chris Fourie
Medical doctor turned health data scientist, trying to use the powers of AI for social good. Currently busy with a Computer Science masters and interested in research at the intersections of theoretical / computational neuroscience and machine learning. I enjoy long-ish walks in the mountains and will board down any snowy stretch of hillside that happens to be nearby.
Geraud Nangue Tasse
I am an IBM Ph.D. fellow and a Ph.D. student in Computer Science at the University of the Witwatersrand. I have always been fascinated by the immense potential for good of creating intelligent systems---the pinnacle of which is artificial general intelligence (AGI). My main research interests lie in reinforcement learning (RL) since it is the subfield of machine learning which I believe has the most potential for achieving AGI. In particular, my current work focuses on principled approaches for efficient compositional generalisation of skills in RL.
Tutorials
An Introduction to Physical Reasoning Benchmarks
Physical reasoning is an important ability of intelligent agents — an intelligent agent that understands the physical laws of an environment and how these laws relate to its own actions and the actions of other agents will be able to act in a more effective and flexible way in a complex, uncontrolled setting. In this tutorial, we provide an overview of two recently established commonsense physical reasoning benchmarks – PHYRE and CREATE. These benchmarks provide a diverse range of physics puzzles, which are relevant to multiple research areas including: reinforcement learning for physical reasoning, predictive modelling and model-based RL, multi-task learning, causal reasoning, generalisation to unseen actions, as well as understanding tool functionality and usage.
Additional Links: https://www.clvrai.com/create/ and https://phyre.ai/#tasks
Presented by Divanisha Patel
I am a PhD candidate in Computer Science at the University of the Witwatersrand, with a research focus on abstract model-based reinforcement learning for physical reasoning. More specifically, I am investigating decomposing physics scenes into objects and their interactions to achieve improved generalization. I am passionate about tech for social good and its potential to break barriers through providing previously inaccessible opportunities and services to the masses.
Google Earth Engine and its Opportunities for Geospatial Machine Learning
With a drastic increase in the amount of freely available satellite imagery, there has been a corresponding increase in the need for the storage and computational infrastructure to be able to extract meaningful ecological insights. The recent Google Earth Engine (GEE) - a cloud computing platform that has provided an unprecedented opportunity for researchers to carry out large scale analysis. In this introduction to GEE, we (1) briefly consider recent and ongoing research that leverages GEE, (2) highlight the strengths, weaknesses, opportunities, and threats of GEE and (3) walk through getting started with GEE in python and some of GEE’s powerful capabilities.
Optional: register for a free GEE account (https://signup.earthengine.google.com/).
Presented by Geethen Singh
Geethen Singh is a Ph.D. candidate at the University of the Witwatersrand, South Africa. He is interested in applying machine learning to satellite data to gain ecological insight. He is currently working on satellite-based monitoring to aid in the management of water hyacinth - the world’s worst invasive aquatic alien plant.
A How-To Guide on Getting Started with RoboCup's 3D Soccer Simulation
This tutorial will walk you through the basics on how to install/run and use all the components required to get started in RoboCup's 3D soccer simulation domain. RoboCup being an annual Robotics competition with multiple leagues from soccer to rescue: https://www.robocup.org/
[tutorial slides]
Presented by Branden Ingram
I am a PhD candidate in Computer Science at the University of the Witwatersrand. My research interests involve machine learning and its application in video games where I am currently focused on Play-style Identification. However, I have also taken a keen interest in the RoboCup domain which is a multifaceted problem where I serve as a leader of the WITS RoboCup Team.
An Introduction to Reinforcement Learning
Reinforcement learning (RL) is a powerful framework in artificial intelligence that enables agents to learn desired behaviors by maximising the rewards received through interaction with an environment. It is relevant to an enormous range of tasks, including robotics, game playing, consumer modeling and healthcare. This tutorial will provide an introduction to RL, with an emphasis on agents that can make near-optimal decisions in a timely manner with incomplete information and limited computational resources. We will cover Markov decision processes (problem formulation), temporal difference learning (Q-learning), and function approximation (deep Q-learning).
Presented by Geraud Nangue Tasse
I am an IBM Ph.D. fellow and a Ph.D. student in Computer Science at the University of the Witwatersrand. I have always been fascinated by the immense potential for good of creating intelligent systems---the pinnacle of which is artificial general intelligence (AGI). My main research interests lie in reinforcement learning (RL) since it is the subfield of machine learning which I believe has the most potential for achieving AGI. In particular, my current work focuses on principled approaches for efficient compositional generalisation of skills in RL.
Learning without Labels: An Introduction to Visual Self-Supervised Representation Learning
Self-supervised learning has recently come to the fore as a viable alternative to fully supervised pre-training for computer vision models that can perform on par, and often better, than these supervised baselines on various downstream tasks. The main advantage of these models is that no human annotation is required to train them - only a large collection of unlabelled images. We discuss self-supervised learning, the state-of-the-art architectures in the domain, and run through an example implementation of a self-supervised model.
Presented by David Torpey
David is a computer vision researcher at Shutterstock Ireland, and PhD candidate at the University of the Witwatersrand. He uses machine learning and deep learning to solve large-scale, real-world computer vision problems, leveraging the semantic and technical properties of images and video for various use cases in the business. His PhD research is focused on sample-efficient unsupervised representation learning.
Additional Organising Team Members
Devon Jarvis
I am a Masters student in Computer Science at the University of the Witwatersrand. I am supervised by Benjamin Rosman, Richard Klein and Andrew Saxe, and a member of both RAIL Lab and Prime. My interests are in learning theory, with a particular focus on the generalizability of neural networks.
Steven James
Google PhD fellow and associate lecturer at the University of the Witwatersrand, where I've spent my entire academic career. I'm also part of the RAIL Lab research group. My research interests revolve around artificial intelligence, and reinforcement learning in particular. I'm currently interested in developing agents capable of learning symbolic state representations that can be transferred between tasks. I count myself extremely fortunate to be supervised by George Konidaris and Benjamin Rosman.