About Micha
Micha is a machine learning researcher (a post-doc fellow in University of Toronto, and Vector Institute), focusing on latent variable models, representation learning, Bayesian inference, and semi-supervised/unsupervised learning. He recently graduated his PhD under the supervision of prof. David Fleet in the Computational Vision Group / Artificial Intelligence Lab, part of the Department of Computer Science at University of Toronto, and in the Vector Institute, Toronto. Micha has been working on machine learning, computer vision, physical inference, and optimization problems for the last decade.
In the past, he focused mainly on 3D video tracking combined with physical models. Currently he has been focusing on representation learning, generative models, Bayesian modelling (always go Bayesian!), and in particular the Mutual Information Machine (MIM). MIM is trained with a new symmetric variational inference, which learns a latent representation with high mutual information, and low marginal entropy. Importantly, MIM is not prone to suffer from posterior collapse, unlike VAEs. If you use VAE, you might find MIM useful!
Micha holds a BSc. in Electrical Engineering and a BSc. in Physics from the Technion, Israel Institute of Technology, both with Summa Cum Laude honors (top 3% of his class). He obtained his MSc. from University of Toronto under David's supervision, where he researched inferring attributes, such as gender, weight, happiness and anxiety level, from motion capture data and from video tracking. During his PhD Micha worked with MERL (Mitsubishi Electric Research Labs), Disney Research Pittsburgh (i.e., which has been closed since then), and Adobe CTL (Creative Technologies Labs) Seattle, and Creative Destruction Labs (Toronto), among others.
Micha also founded a Computer Vision/Machine Learning lab, with the goal of pushing ahead CV/ML research, while giving smaller start-ups (who typically cannot afford to hire ML researchers) accessibility to one of the biggest revolutions that humanity is experiencing. He also believes that the social/moral aspects of that revolution are ignored, by large. As part of his efforts for a better future for all, he is working to encourage an open discussion about the social implications of the AI revolution, in order reduce fears from AI, and replace it with openness and better understanding of what possibilities lies ahead.
In his spare time he enjoys hiking, canoeing, SUP, bike riding, skiing, ice skating, or basically every outdoor/indoor sport activity he finds the time to put into.
You can find Micha's CV here.
Micha is always looking for the next opportunity! If you find his research interests aligned with yours, please do not hesitate to reach out to him.
News and Updates
- 01/07/202 - I started a post-doc fellowship in University of Toronto, and Vector Institute.
- 18/02/2020 - SentenceMIM preprint (see below) is available - we break 2 SOTA records when training a probabilistic autoencoder with MIM learning!
Contact Information
Dept. of Computer Science University of Toronto 10 King's College Road, Rm. 3303 Toronto, ON M5S 3G4
Office
Pratt Building, Rm. 263A 6 King's College Road Toronto, ON M5S 3G8
Or
Vector Institute MaRS Centre, West Tower 661 University Ave., Suite 710 Toronto, ON M5G 1M1
Publications
- Thesis: “Symmetric Variational Inference with High Mutual Information"
- “High Mutual Information in Representation Learning with Symmetric Variational Inference”
- “TzK Flow - Conditional Generative Model”
- “Walking on thin air: Environment-free physics-based markerless motion capture”
- M. Livne, L. Sigal, M. A. Brubaker, and D. J. Fleet
- Computer and Robot Vision, no. 15th, 2018.
- project page pdf bibtex
- “Spinal cord segmentation by one dimensional normalized template matching: A novel, quantitative technique to analyze advanced magnetic resonance imaging data”
- “Characterizing the location of spinal and vertebral levels in the human cervical spinal cord”
- “Human attributes from 3d pose tracking”
- “Human attributes from 3d pose tracking”
- L. Sigal, D. J. Fleet, N. F. Troje, and M. Livne
- ECCV 2010 European Conference on Computer Vision. European Conference on Computer Vision (ECCV), Heraklion, Greece, 2010
- bibtex
Preprints
- “SentenceMIM: A Latent Variable Language Model”
- M. Livne, K. Swersky, and D. J. Fleet
- Preprint, 2020
- github repo model rankings
- pdf bibtex
- “MIM: Mutual Information Machine”
- M. Livne, K. Swersky, and D. J. Fleet
- Preprint, 2019
- project page presentation github repo
- pdf bibtex
- “TzK: Flow-Based Conditional Generative Model”
- M. Livne, and D. J. Fleet
- Preprint, 2019
- TzK: Flow-Based Conditional Generative Model
Open Source Contributions
Micha created an open source research-support framework to help small companies, independent researchers, and students to concentrate on their research work. The frameworks are meant to supply powerful utilities, classes, and tools to be used with many existing research frameworks. It is a work in progress, and all feedback/requests are welcomed. If you would like to be involved please contact Micha via email.
NOTE: repos are temporarily down for maintenance
- Settings - the base for all other libraries. Contains many useful scripts, and helps in deploying multiple research environments on local/remote machines.
- Common - The heart of the framework. It contains many powerful classes and utilities that Micha developed over the course of my PhD.
- Deployment - A collection of helper scripts to install external research frames on Linux and OS X.
Research Highlights
Here you can find a collection of videos that highlights some of Micha's research projects.
MIM: Mutual Information Machine
MIM
VAE
PNN - Physical Neural Network
Performance Capture
Spine Segmentation
Physical Tracking
Fun
Random projects that are guaranteed to possibly put a smile on your face!
- Mike and the City - The exciting tale of Mike and the City, based on (mostly) true events.