Dear Mila partners and researchers,
Mila is growing: In the second quarter, we signed 8 new partnerships, including Hitachi, Nuance Communications, and DiDi, filling 10 of our 12 corporate labs for the next 3 to 5 years. In the third quarter, we are poised to welcome an extra 150 students (intern, MSc, PhD and post-doc) and at least 5 new faculty, which will increase the total number of researchers to more than 500 by the end of September.
We will also continue talks with a few more potential partners, where we see especially good fits with Mila's values and research interests.
— Frédéric Laurin, Director of Partnerships
|
|
|
|
|
|
Finding roads in satellite images |
|
|
|
When AI tries to find a path through an unfamiliar area by classifying each pixel in a satellite image as either road or not road, it often ends up finding only disconnected pieces of a path, as the road's appearance changes. A new method presented at CVPR considers the angles of the roads at the same time that it learns their boundaries and then uses a multi-branch convolutional neural network to infer the connections between the pieces of the path. Mila PhD student Suriya Singh (India) shared first authorship with IIIT Hyderabad master's student Anil Batra on a team with IIIT Hyderabad professor C.V. Jawahar and Facebook research scientists Guang Pang, Saikat Basu, and Manohar Paluri.
|
|
|
|
Above: Suriya's team's method ("Ours," far right) traces the roads in a satellite picture ("Image," far left) as identified by human experts ("GT," i.e. "ground truth," second from left) better than the previous state-of-the-art methods ("DRM," "TL," "L34," "MatAN").
Read the full paper here.
|
|
|
|
Fighting back against hackers |
|
|
|
AI safety advocates are deeply concerned with the prospect of combatting adversarial attacks, in which hackers can cleverly make small changes to an image — invisible to the human eye — and trick AI into misclassifying it. At CVPR, Mila intern Saeid Asgari (Iran), working with Simon Fraser University master's student Kumar Abishek, Simon Fraser University professor Ghassan Hamarneh and University of British Columbia PhD student Shekoofeh Azizi, proposed a method to ward off these attacks through a function which transforms the features of each layer in a convolutional neural network into a new manifold, with less overlap between the classification categories.
|
|
|
|
Above: Saeid's team's method ("PROP," far right), unlike previous methods ("ORIG," "ADVT," "FSG," "FSM"), correctly locates the lesions on these pictures of patients' skin ("Legitimate," far left), even when the pictures have been altered by adversarial attacks ("Perturbed," second from left), in the same way that human experts do ("ISIC GT," third from left).
Read the full paper here.
|
|
|
|
One of the main barriers to the broader deployment of neural networks is their computational expense; as researchers at Mila, we have access to powerful computing resources which most organizations don't have. At ISCAS, Mila master's student Olexa Bilaniuk (Canada), IBM research scientist Sean Wagner, and Polytechnique Montréal professors Yvon Savaria and Jean-Pierre David proposed an architecture for a hardware accelerator which would allow the performance of a neural network to scale with the precision of its weights and activation functions through a process called bit-slicing, making it possible to run lower-precision neural networks more cheaply.
|
|
|
|
Left: Before performing matrix-vector multiplication, Olexa's team's method groups each factor's most significant bits (MSB), second-most significant bits, and so forth, down to the least significant bits (LSB).
Right: When performing multiplication, the method starts by multiplying the two most significant bits and then proceeds by a "zig-zag" order of encounter, computing all the products between pairs of bits whose significance rankings sum to the same total, before descending to the next-highest significance total.
Read the full paper here.
|
|
|
|
Above: Alex's team's model still struggles on pages which contain handwritten characters of multiple sizes. Blue shows the correct modern Japanese equivalent for each pre-modern character. Red shows the model's predictions, some of which are incorrect.
|
|
|
|
Above: Mila master's student Xavier Bouthillier (Canada) presents a paper which we did not summarize in our June newsletter called "Unreproducible Research is Reproducible," coauthored with Mila PhD student César Laurent (Switzerland) and Mila professor Pascal Vincent (France).
A quick overview of all of Mila's ICML papers is available on the Mila.Quebec blog.
|
|
|
|
AI4Good Summer Lab Demo Night: In our April newsletter, we told you about the AI4Good Summer Lab, founded by Mila professor Doina Precup (Romania) and backed by many of Mila's partners, which recruits women from STEM fields across Canada to come to Mila for seven weeks, study the basics of AI, divide into teams, and develop applications for real-word problems.
At the end of the seven weeks, the members of Montreal's AI ecosystem who attended the AI4Good Summer Lab Demo Night voted on their favorite project. The winner was A.I.D., a health care tool for chronic pain management which learns from a patient's reported schedule and reported pain levels which future activities are most likely to cause pain.
|
|
|
|
The audience also voted for two projects to receive honorable mentions: Elentless, a bias-detection program which mines LinkedIn data for correlations between the physical features of employees and the companies that they work for, and Climate Times, a filter which finds news coverage reporting on extreme weather events without attributing them to climate change.
|
|
|
|
Take a survey: Mila PhD student Adrianna Janik (Poland) is collecting data on the interpretability of latent spaces for her thesis. She's looking for people to spend two minutes playing with her interactive demo and then two minutes answering questions.
To help Adrianna, go here.
|
|
|
|
Upcoming events in Montreal |
|
|
|
IBM AI Horizons Network Seminar Series: July 10 @ 12pm, remote via Webex.
UCSD master's student and IBM intern Dustin Wright will present his paper "NormCo: Deep Disease Normalization for Biomedical Knowledge Base Construction," which won the Best Application award at the 2019 Automated Knowledge Base Construction (AKBC) conference.
To watch the talk, go here.
To watch previous talks in the series, go here.
|
|
|
|
|
|
Tea Talk: July 12 @ 10:30am, 6650 Rue Saint-Urbain (Mila Auditorium)
IIT Madras professor Balaraman Ravindran will present on a topic TBD.
|
|
|
|
NextAI Venture Showcase: July 18 @ 3:30-9:00pm, 6650 Rue Saint-Urbain (Mila Agora)
The first cohort of startups to come out of the NextAI-Montreal accelerator program will pitch their projects to the community over cocktails and light refreshments.
To register, go here.
|
|
|
|
Tea Talk: July 19 @ 10:30am, 6650 Rue Saint-Urbain (Mila Auditorium)
UdeM professor Paul Cisek will give a talk titled "Rethinking Behavior from an Evolutionary Perspective."
|
|
|
|
La Presse: Yoshua Bengio: intelligence profonde
This profile of Mila's founding father includes little-known details from his childhood in Paris and Montreal, and his appreciation of the Netflix series Black Mirror.
|
|
|
|
|