ML Ensemble #3
June 2nd 2018 Toronto

Share machine learning insights, techniques, methods, and observations with your technical peers

Find Out More

ML Ensemble concept

A one day invitation-only meeting for machine learning researchers and practitioners to learn from each other, sharing not just polished high-level demos, but also the gritty realities: technique details, systems used, failed approaches, and unanswered questions.

Academics & Industry

Learn about research frontiers from academia; learn about applied realities from industry

Technically In-Depth

This is not a marketing meeting. Talks will not shy away from necessary mathematical and technical details.


We are limiting the event size to allow for actual discussion, and leaving plenty of time for it in the schedule.





The ML Ensemble will take place in the beautiful Kensington Market offices of Normative

91 Oxford Street, Suite #200
Toronto, Ontario
M5T 1P2


Don't have the password? Reach out to and tell us about yourself

Register now

Past speakers

Sageev Oore

Deep Learning for Musical Language Generation

Dalhousie / Vector


Jake Cheng

Hubba: Recommendations and Beyond



Renjie Liao

Deep Learning on Graphs: Models and Applications

Uber ATG


Angela Schoellig

Machine Learning for Safe, High-Performance Control of Mobile Robots

University of Toronto


Kalu Kalu

Insights from Building and Applying Behavioural Image Prediction Models



Kathryn Hume

Ethical Algorithms: Bias and Explainability in Machine Learning

Integrate AI


Renat Gataullin

Learning to Grasp



Solmaz Shahalizadeh

Detecting Order Fraud on More Than 400k Stores: Scaling Machine Learning at Shopify



Afsaneh Fazly

Consumer Opinion Analysis: Lessons Learned

Thomson Reuters


Anna Goldenberg

Building ML Models When the Data is Scarce: The Case of Complex Human Diseases

Sick Kids


Xavier Snelgrove

Subjective Objective Functions: High-Resolution Neural Image Creation



Jimmy Ba

Progress and Challenges in Optimizing Neural Networks

University of Toronto / Vector


Yevgeniy Valis

Yevgeniy Valis photo

Leveraging privacy preserving machine learning methods to overcome the barriers to data sharing

In the B2B space, it is common for customers to restrict the use of their data by the service provider. As a result, existing customers' data cannot be used to bootstrap new customers predictive models, leading to a long data collection period before new customers have enough data in the system to train models that perform adequately well. We set out to overcome this limitation by building privacy preserving machine learning models over the data of the entire customer base while eliminating existing customers concerns about the use of their data to benefit others. In this talk we will discuss a set of methods called Differential Privacy that allowed us to solve the data rights and cold start problem in a privacy preserving way.


Yevgeniy Vahlis is the head of the applied machine learning group at Borealis AI. Prior to joining Borealis AI Yevgeniy lead an applied research team at Georgian Partners, a late stage venture capital fund, and worked at a number of tech companies including Amazon and Nymi. Yevgeniy started his career at AT&T Labs in New York as a research scientist after completing his PhD in computer science at UofT and spending a year at Columbia university as a postdoc.

Martin Snelgrove

Martin Snelgrove photo

Tensor Stasis: Inference without kilowatts

The energy consumed in doing heavy ML inference is a problem: it kills battery life and heats up data centres. Cloud computing fixes nothing, because the energy your device uses to ship the data is worse than what it needs to do the work itself.

I’ll cover a bit of the physics of where energy goes; but it turns out that almost all of it is wasted pumping data around. The trendy ways to do that are “in-memory” and “near-memory” computing, one of which works at scale.

From a software point of view, programming near memory is a serious mind-warp. A GPU runs around 1,000 processors in parallel; we run a million. On one chip. On a AAA cell.


Martin is a partly reformed academic: degrees and ECE Prof. at U. of Toronto, sabbatical time at AT&T Bell Labs and then an Industrial Research Chair at Carleton. He followed his lab when it started a chip shop called Philsar, then was CTO for a company that did early LTE-type wireless internet, then a stack of consulting. For the last ten years he’s been CEO of Kapik, which designs strange mixed-signal ICs in Toronto and thereabouts.

Back in the 90s he was involved in early massively-parallel computing work in “smart memory”, which was the most power-efficient but a highly specialized way to do video: and now it turns out that dot products are very good; burning energy is very bad; and transistors are very small; so the technology is back. He’s CEO of “Untether”, doing a chip with seed round funding and getting ready for A round and turmoil.

Julieta Martinez

Julieta Martinez photo

Scalable algorithms for large-scale retrieval via multi-codebook quantization

Multi-codebook quantization (MCQ) is a matrix factorization problem analogous to k-means, where the cluster centres arise from a combinatorial combination of entries in smaller codebooks.

The applications of MCQ include the compression of deep neural networks for mobile devices, large-scale approximate nearest neighbour search (ANN), and large-scale maximum inner product search.

Unfortunately, MCQ poses a series of challenging optimization problems that are often too expensive in practice for large-scale datasets.

Drawing from the literature on the stochastic optimization of hard combinatorial problems, I'll introduce three improvements to the standard optimization procedure of MCQ that obtain state-of-the-art compression rates on a range of computer vision tasks. I will also introduce Rayuela.jl, an open-source, GPU-optimized library for large-scale MCQ written in Julia.


Julieta is a computer vision researcher focused on machine learning, deep learning, and large-scale retrieval. She is currently a researcher with Uber ATG Toronto. She enjoys coming up with simple, scalable, and easy-to-understand algorithms that address challenging research problems. Her work has been published in the top three computer vision venues (CVPR/ECCV/ICCV). Julieta obtained her PhD from UBC and has interned at Disney Research and the Max Planck Institute.

Oren Kraus

Oren Kraus photo

Classifying and segmenting microscopy images with deep multiple instance learning

High-content screening (HCS) technologies have enabled large scale imaging experiments for studying cell biology and for drug screening. These systems produce hundreds of thousands of microscopy images per day and their utility depends on automated image analysis. Recently, deep learning approaches that learn feature representations directly from pixel intensity values have dominated object recognition challenges. These tasks typically have a single centered object per image and existing models are not directly applicable to microscopy datasets. Here we develop an approach that combines deep convolutional neural networks (CNNs) with multiple instance learning (MIL) in order to classify and segment microscopy images using only whole image level annotations. We introduce a new neural network architecture that uses MIL to simultaneously classify and segment microscopy images with populations of cells. We base our approach on the similarity between the aggregation function used in MIL and pooling layers used in CNNs. To facilitate aggregating across large numbers of instances in CNN feature maps we present the Noisy-AND pooling function, a new MIL operator that is robust to outliers. Combining CNNs with MIL enables training CNNs using whole microscopy images with image level labels. We show that training end-to-end MIL CNNs outperforms several previous methods on both mammalian and yeast datasets without requiring any segmentation steps.


I'm the co-founder of Phenomic AI; a startup accelerating image-based drug discovery with AI. Previously, I completed my PhD in Brendan Frey's lab. My research focused on applying cutting edge machine learning techniques (specifically deep learning) to high throughput microscopy screens of cell biology. In collaboration with Charlie Boone and Brenda Andrews at the Donnelly Centre for Cellular and Biomolecular Research (CCBR), I generated datasets and trained deep learning models on millions of individual cell objects from genome wide microscopy screens. Previously, I completed my BASc and MASc in mechanical and biomedical engineering at the University of Toronto. I also interned as a machine learning researcher at Apple in California and Borealis AI in Toronto.

Jennifer Listgarten

Jennifer Listgarten photo

From genetics to gene editing with machine learning

Molecular biology, healthcare and medicine have been slowly morphing into large-scale, data driven sciences dependent on machine learning and applied statistics. This talk will highlight two examples of work at this intersection of disciplines: (1) Understanding the genetic underpinnings of disease is important for screening, treatment, drug development, and basic biological insight. Genome-wide associations, wherein individual or sets of genetic markers are systematically scanned for association with disease are one window into disease processes. Naively, these associations can be found by use of a simple statistical test. However, a wide variety of confounders lie hidden in the data, leading to both spurious associations and missed associations if not properly addressed. Listgarten will discuss how these artifacts arise and how we can fix them. (2) Once we uncover such genetic weaknesses, genome editing—which is about deleting or changing parts of the genetic code—will one day let us fix the genome in a bespoke manner. Editing will also help us to understand mechanisms of disease, enable precision medicine and drug development, to name just a few more important applications. In 2012, a revolutionary new molecular biology technique called CRISPR emerged, which dramatically changed the field of gene editing. The presentation will discuss the development of a state-of-the-art predictive model to enable more effective CRISPR gene editing.


Since Jan. 2018 Jennifer Listgarten is a Professor in UC Berkeley EECS department and Center for Computational Biology, a member of the steering committee for the Berkeley AI Research (BAIR) Lab, and a Chan Zuckerberg investigator. From 2007 to 2017 she was at Microsoft Research, through Cambridge, MA; Los Angeles and Redmond, WA. Before that she did her PhD in the machine learning group at the University of Toronto, in her wonderful hometown of Toronto. Her undergraduate studies in Physics and Computer Science were at Queen's University.

Jennifer’s area of expertise is in machine learning and applied statistics for computational biology. She’s interested in both methods development as well as application of methods to enable new insight into basic biology and medicine.

Ozge Yeloglu

Ozge Yeloglu photo

Customer Experience Analysis by using NLP and Deep Learning Techniques

In this talk, I will be focusing on my team’s experience with working with financial services customers and the challenges and learnings we have in building AI applications with them. I will also talk about a specific project that we are currently working on by using web clickstream data to understand customer experience and satisfaction. We are using NLP and Deep Learning techniques as a solution.


Ozge is the Data & AI Lead at Microsoft Canada. She is leading a team of Data Architects and Data Scientists to build intelligent solutions for Microsoft’s top clients in Canada. Her focus was on FinTech and building analytics solutions for some of the biggest Financial Services companies in her previous role as the first Data Scientist at Microsoft Canada. She has been working with the VP Analytics, CTOs, Director of Analytics and Data Scientists directly to help and advise them building advanced analytics solutions to increase their customer satisfaction, revenue, efficiency and optimize their operations to stay ahead of the curve in today’s competitive market.

Ozge truly believes the words that she first heard from Grace Hopper “ The most dangerous phrase in the language is, ‘We’ve always done it this way.’ ”. She has the entrepreneurial experience with an academic background. Prior to joining Microsoft, she was the co-founder and CEO of topLog, a start up focused on predictive data analysis based on unstructured log data. She started topLog in 2013 while she was working towards her PhD degree in Computer Science at Dalhousie University. Ozge is a life-long learner and problem solver. Her experience with switching from academia to entrepreneurship has taught her how to solve real world problems by applying her deep academic knowledge.