In last month's newsletter, we told you about the twenty-two Mila papers accepted to the International Conference on Machine Learning (ICML), held in Long Beach, California this June. The week after ICML, some Mila students stuck around Long Beach for the smaller Conference on Computer Vision and Pattern Recognition (CVPR). In this month's newsletter, we'll look at two Mila papers from CVPR and a third from the International Symposium on Circuits and Systems (ISCAS) in Sapporo, Japan.

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.

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.

Followers of the Mila.Quebec blog may recall the adventures of Mila PhD student Alex Lamb (USA), who — when he's not mixing up his manifolds or reifying his hidden states — has been using deep learning to preserve classical Japanese literature. With his collaborators Tarin Clanuwat and Asanobu Kitamoto at the Center for Open Data in the Humanities (CODH), Alex has had a new paper accepted to  the International Conference on Document Analysis and Recognition (ICDAR), which will take place in September in Sydney, Australia. The new paper presents an end-to-end model called KuroNet which tackles some of the outstanding challenges in reading raw pages from these old manuscripts.

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.

Mila PhD student Vikram Voleti (India) got camera-happy during the poster sessions at the International Conference on Machine Learning (ICML).

Above: Mila intern Nasim Rahaman (Germany) and Mila postdoc Devansh Arpit (India) present their paper "On the Spectral Bias of Neural Networks," which we summarized in the Research section of our June newsletter under the heading "When there's more than one right answer."

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.

Prix D'excellence: Mila scientific director Yoshua Bengio (Canada) is one of two recipients of the 2019 Prix D'excellence from the Fonds de recherche du Québec - Nature et technologies (FRQNT). The prize recognizes outstanding contributions from academic researchers in the natural sciences, mathematics, or engineering and comes with a $10,000 cash award.

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.

Above: Mila professor Doina Precup (Romania) and NVIDIA Canadian Sales Manager Eric Dahan with the members of the winning team: Université de Sherbrooke physics master's student Maude Lizaire, McGill undergraduate Jazlyn Hellman, Ericsson software developer Mahsa Khosravi, and McGill undergraduate Ran Tao. 

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.

Above: Mila professor Doina Precup (Romania) with the two honorable mention teams. 

Elentless (left) was made by University of British Columbia undergraduates Kavya Kaluarachi and Naadiya Ahmedabadi, University of Toronto computer science PhD student Audrina Zhou, and McMaster undergraduate Alyssia Jovellanos. 

Climate Times (right) was made by University of New Brunswick undergraduate Vaishnavi Anupindi, McGill undergraduates Rosie Zhao and Emi Baylor, and University of Manitoba undergraduate Camila Bustos.

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.

Write your own disentanglement algorithm: Mila scientific director Yoshua Bengio (Canada) is on the organizing team for the 2019 Conference on Neural Information and Processing Systems (NeurIPS) Disentanglement Challenge, which is now offering 20,000€  in prize money for new algorithms which can transfer from simulated data to real data.

To download the dataset, go here.

Tea Talk: July 5 @ 10:30am, 6650 Rue Saint-Urbain (Mila Auditorium)

Mila PhD student Ankesh Anand (India) and Mila master's student Evan Racah (USA) will jointly present their paper "Unsupervised State Representation Learning in Atari," submitted to the 2019 Conference on Neural Information and Processing Systems (NeurIPS) and coauthored with Mila PhD student Sherjil Ozair (India), Mila scientific director Yoshua Bengio (Canada), Mila associate member Devon Hjelm (USA), and Microsoft researcher Marc-Alexandre Côté.

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.

The Verge: Here's how AI can help fight climate change according to the field's top thinkers

MIT Technology Review: Here are 10 ways AI could help fight climate change

These two articles both summarize the findings of a new survey paper coauthored by a team of twenty-two elite researchers in machine learning spread across sixteen institutions, including four Mila members: postdoc Kris Sankaran (USA), Sasha Luccioni (Canada), PhD student Tegan Maharaj (Canada), and scientific director Yoshua Bengio (Canada).

New Scientist: Humans should worry us more than machines, says founding father of AI

This interview with Mila scientific director Yoshua Bengio (Canada) emphasizes health care applications as one of the best effects of AI.

Mobile Syrup: Samsung opens AI Lab in Montreal in partnership with local research centre

This feature piece covers the new Samsung Advanced Institute of Technology (SAIT) lab in Montreal, directed by Mila professor Simon Lacoste-Julien (Canada).