Microsoft Research & Mila

Mila faculty engagement

Microsoft Research & Mila Call for Proposals

Semi-annually, Microsoft Research conducts a call for research proposals with the faculty at Mila to promote machine learning research collaborations between the Microsoft Research and Mila. The call for proposals aims to fund ambitious collaborative projects that have the potential to shape the future of artificial intelligence.

Please see below a list of selected proposals funded by Microsoft Research during the most recent call for proposals. The next round of call for proposals will take place during May-June, 2020.

March 2020

• Learning Representations from Physical Interaction

  Devon Hjelm (MSR Montreal), Mihai Jalobeanu (MSR AI), Yonatan Bisk (MSR AI), Liam Paull (Mila), Florian Golemo (Mila), Aaron Courville (Mila)

  In this project we explore the role of interaction on revealing underlying properties of the world, notably those properties not easily discernible through visual observation alone.  Specifically, we train models by curiosity and interaction within robotic control, as this setting allows us to concretely design experiments over which we can reasonably make progress on questions that relate interaction with representation learning. Our experimental setting will focus on learning invisible properties and affordances through interaction both in simulation and on real robot platforms.

• Parsimony-Biased Representation in Reinforcement Learning

  Rémi Tachet Des Combes (MSR Montreal), Pierre-Luc Bacon (Mila), Yoshua Bengio (Mila), Romain Laroche (MSR Montreal)

  We aim to explore the concept of parsimony bias (PB) in reinforcement learning. The PB states that agents should reason about a small number of concepts from the current state (state-space PB), and over a small number of meaningful time steps in its trajectory (temporal PB). We hope to demonstrate that an agent trained using the PB has strong credit assignment and planning capabilities and can adapt rapidly to new environments.

• Meta Learning by Inducing Logical Rules

  Will Hamilton (Mila/McGill), Xingdi Yuan (MSR Montreal), Marc-Alexandre Cote (MSR Montreal), Romain Laroche (MSR Montreal), Adam Trischler (MSR Montreal)

  The ability to learn and generalize from small amounts of data is a fundamental aspect of human intelligence. In the context of machine learning, this ability is often studied under the umbrella of «meta learning», in which models are trained to enable fast adaption to new tasks. Our project will explore a new approach to meta learning based on the concept of logical rule induction, where we will develop learning approaches that seek to extract the shared logical structure between related tasks.

• Towards Combining Statistical Relational Learning and Graph Neural Networks for Reasoning

  Jian Tang (Mila), Chenyan Xiong (MSR AI)

  Relational Reasoning, a key ingredient for System II reasoning, is important in a variety of complicated tasks such as question answering, image understanding, and knowledge graph completion. This problem has been extensively studied in traditional statistical relational learning community and recent growing graph representation learning community in an independent way. In this project, we aim to combine the advantages of both worlds and foster the collaboration between the two different communities.

• Open-Domain Conversational Question Answering

  Shehzaad Dhuliawala (MSR Montreal), Kaheer Suleman (MSR Montreal), Siva Reddy (Mila/McGill) 

  In this proposal, we plan to develop next-generation search engines which have the ability to answer conversational questions. The main challenge is building a joint inference model that understands conversations, perform information retrieval and read documents to answer questions.

• Two-sided Recommendation with Fairness

  Xue Liu (Mila), Fernando Diaz (MSR Montreal)

  Two-sided recommendation systems bring together two populations such as buyers and sellers, listeners and musicians, and employers and job seekers.  While many existing recommendation systems have focused on optimizing metrics such as user satisfaction, they often do not incorporate the satisfaction of content producers.  As a result, the platforms become less attractive and content quality can suffer.  Moreover, even when multiple objectives are considered, they can result in unfair allocation of exposure for content from certain groups.  This project will develop models for multi-objective, fair recommendation.

• Mitigating Gender Pronoun Bias in Text

  Carolyne Pelletier (Mila), Sandeep Subramanian (Mila), Samira Shabanian (MSR Montreal), Yoshua Bengio (Mila)

  This project investigates ways in which Gender Pronoun Bias (a type of implicit bias) can be mitigated when deploying natural language processing systems. This is important because, much like how humans incorporate their own biases into the text they write, language understanding models like BERT have been shown to propagate these types of biases that exist in their training data to downstream tasks. First, we propose a modification to BERT’s pre-training task in order to mitigate for Gender Pronoun Bias. Second, we propose to fine-tune on an augmented dataset that aims to reduce Gender Pronoun Bias in tasks like Masked Language Modeling. Finally, we explore the possibility of training a model to identify and correct for Gender Pronoun Bias at the sentence level by using a pre-trained language understanding model like BERT and fine-tuning it on the Gender Pronoun Bias dataset which is being collected at Mila with BiaslyAI. This could be used as an end-user application or used to correct for Gender Pronoun Bias in training data.

May 2019

• Towards characterizing compositionality and systematic generalization for natural language representations

  Alessandro Sordoni (MSR Montreal) and Professor Timothy O’Donnell (Mila)

• Learning and Reasoning with Graph Structures in Interactive Text Environments

  Marc-Alexandre Cote (MSR Montreal), Eric Yuan (MSR Montreal), Professor William Hamilton (Mila) and Professor Jian Tang (Mila)

• Learning Dynamics of Adversarial Machine Learning Formulations

  Remi Tachet Des Combes (MSR Montreal) and Professor Ioannis Mitliagkas (Mila)

MSR Collaborating Faculty Members at Mila

• Professor Yoshua Bengio (opens in new tab)
• Professor Aaron Courville (opens in new tab)
• Professor Ioannis Mitliagkas (opens in new tab)
• Professor Jackie Cheung (opens in new tab)
• Professor William L. Hamilton (opens in new tab)
• Professor Timothy J. O’Donnell (opens in new tab)
• Professor Reihaneh Rabbany (opens in new tab)
• Professor Jian Tang (opens in new tab)
• Professor Pierre-Luc Bacon (opens in new tab)
• Professor Xue Liu (opens in new tab)
• Professor Liam Paull (opens in new tab)
• Professor Siva Reddy (opens in new tab)

MSR Distinguished AI Lecture Series at Mila

Once a month Microsoft Research hosts a technical lecture featuring a distinguished faculty/academic or Microsoft Researcher. Faculty, students and the local academic community are welcome to attend. Pizza is provided afterwards for faculty and the academic community to network. See upcoming and past speakers on our event page.

Once a year the faculty and students at Mila come together with Microsoft researchers to share recent research advancements and breakthroughs to build upon existing collaborations.