ELLIS against Covid-19

COVID-19 research at MILA

ellis Yoshua Bengio 01 April 2020 - 01 April 2020
01 April 2020 • 15:10 - 15:25

Yoshua Bengio (Mila – Quebec Artificial Intelligence Institute)

Prof. Bengio presented the projects related to COVID-19 in his lab called MILA at Université de Montréal, Canada. His lab has recently published a survey of how different machine learning methods can be helpful to fight COVID-19. In particular, MILA is working on two main projects related to Contract tracing and drug discovery. It is known that tracing people using phones has a significant impact on controlling the infection rate of the virus compared with the current method based on interviews. MILA is collaborating with people around the world to develop a mobile app with high emphasis on privacy and machine learning. The app tries to find the intersection among paths of individuals using the bluetooth protocol and use this information to estimate the risk of infection for each person. The peer-to-peer exchange of information is used to preserve privacy by eliminating the need to upload trajectory information to a central storage or processor. Machine learning helps with combining several fuzzy pieces of information and estimates a person’s risk of infection. The belief of one phone about being infected depends on the belief of other phones over the past 14 days. This dependence encourages the use of algorithms such as loopy belief propagation such that several estimates converge to a consistent joint belief among all estimators. To take care of data scarcity, epidemiological models built using the collected data from COVID-19 can be used to generate data for pe-training the machine learning models of the app. MILA is also working on neural networks to predict the binding energy of candidate drugs with the target proteins which the virus needs. Such neural networks can be pre-trained by physical docking engines. In another project, an RL framework similar to AlphaZero is being developed that uses Monte Carlo Tree Search to find those drugs which are synthesizable. Low-cost simulators are used to estimate the reward for this RL algorithm which is otherwise hard to estimate by real biological experiments.


Question & Answers

Link to the recording of the live Questions & Discussion session for this talk. 

  • Q: According to you, how can AI help minimize the outbreak? How can applied AI be productionised quickly to combat the outbreak in time?

    • A: AI can be used to target at-risk of infection populations for isolation, plus in many other ways I discussed.

  • Q: In the future, how can we detect COVID-19 kind of viruses at the very early stage using AI/ML?

    • A: We are investigating the effect of the virus on the breathing patterns in the early days of infection before symptoms become obvious.

  • Q: Why do you think that AI research pan globe is more focused on not so posit topics like self-driving cars, playing games etc where it should be focusing on issues more pressing to make human life liveable for everyone rather than a select few so that a Condition like today’s could have been better dealt with?

    • A: Because research is more driven by profit-seeking rather than by seeking the common good, unfortunately.

  • Q: How would you handle false positives in a contact tracing app? Namely people who state that they have tested positive when in reality they didn't?

    • A: By working in probability space, not hard decisions. It is not a problem.

  • Q: Which exact proteins of virus do you target during drug discovery (e.g. Spike proteins, membrane proteins etc.) and what's the reason for that?

    • A: Protease. I am not the expert but it seems crucial to the virus, allowing to 'cut' things in the cell.

  • Q: How Important is planning in the RL framework to get good results in drug discovery?

    • A: I am not sure but we have good results with MCTS+AlphaZero which has some slight planning.

  • Q: Will the app and server code be open source?

    • A: Yes!

  • Q: What is you take on DL-based approaches on CT imaging and Xrays for COVID-19 screening?

    • A: It's great.

  • Q: Aren't you afraid that the techniques you're developing now might be used for personal surveillance by restrictive governments in the future?

    • A: Absolutely. Hence the very strong privacy protections we are putting. With our method, even if governments seized the files in the servers they could not do anything with them to recover trajectories and contacts.


Thumb ticker yoshua bengio
(Mila & University of Montreal)