Toronto Machine Learning Summit Day 1  [1]      
Time  Business track Applied ML and Case Studies track  Advanced Technical & Research track
8:45 AM   Attendee Registration and Sponsor Booths Open  
9:45 AM   Opening Welcome: David Scharbach and Chairman Frank Rudzicz  
9:55 AM   Keynote Address: Tomi Poutanen Co-Founder of Vector Institute , CEO of Layer-6 AI   
10:10 AM   Keynote Address:  Kerry Liu, CEO at Rubikloud  
10:25 AM   Keynote Address 2:  Surprise Guest  
    Break  
11:10 AM Jennifer Prendki, VP of Machine Learning at Figure Eight Lubna Khader , Lead Data Scientist at Addictive Mobility Professor Hui Jiang York University 
  Agile for Data Science Teams

Who is this presentation for?
Chief data officers, data executives, and data science managers
Prerequisite knowledge
Familiarity with Agile methodologies and data science management (useful but not required)
What you'll learn
Understand why managing data teams is different from managing engineering teams
Learn how to adapt the planning methods and techniques that work in software engineering for data science
Description
Since the publication of the Manifesto for Agile Software Development in 2001, Agile methodologies have been adopted by a majority of tech companies and have unquestionably revolutionized the tech industry and its culture. Agile’s huge success is hardly a surprise: Agile development came as a breath of fresh air at a time when the tech industry was crippled by the many inefficiencies caused by its own success. Back then, the Agile mindset was a panacea for tech’s growing pains.

However, the tech industry is now facing a new revolution: big data, machine learning, and artificial intelligence. The methodologies that were so beneficial to the field of software development seem inappropriate for data science teams, because data science is part engineering, part research.

Jennifer Prendki demonstrates how, with a minimum amount of tweaking, data science managers can adapt Agile techniques and establish best practices to make their teams more efficient. Jennifer starts by discussing the Agile Manifesto in detail and reviewing the reasons for its major success in software engineering. She then outlines the different ways that organizations set up their data science initiatives and explains in which ways these teams differ or are similar to software engineering teams. Jennifer concludes by detailing how to adapt traditional Agile methodologies to create a powerful framework for data science managers and shares tips on how to allocate resources, improve best practices, and tweak the usage of planning and organization tools for the benefit of data teams.
#1 Challenges with applying ML in a high-transaction Big Data environment

Addictive mobility is a leading mobile advertising company in the canadian market. We have built a large ad serving platform that we use to enable our clients to advertise the right service/product to the right person. Our platform handles up to 100K ad auctions per second, and tries to make the best bidding decision for each of these to meet many business constraints and optimize many metrics, and in this process it generates ~6TB of data a day. We heavily rely on Machine Learning to guide our decisions in real-time, based on knowledge derived from the huge datasets that our bidding processes generate. When universities teach Machine Learning, they usually cover the most popular algorithms, the math and logic behind them, when they are best used etc. and they hand students packaged and prepared datasets to apply the algorithms on. All of this is great, but after working for 4.5 years on Machine learning/ Big Data projects, I found there is so much more to the application of ML than what I was taught. Finding an algorithm that works well for a prepared sample set takes a small fraction of the time needed to solve the real problem in a production environment. Many popular algorithms break when the data grows to the scale we are dealing with at Addictive Mobility, at that point, we had to explore different approaches to tweak the algorithm, or switch the algorithm completely sacrificing performance in favour of a simpler and lighter one. In other cases, the algorithm itself wasn’t the bottleneck, but rather acquiring the necessary training data, or storage of the results or intermediary data. Big Data has introduced various challenges to how traditional systems work, and it is not trivial to use ML with Big Data. In my talk, I will discuss the challenges my team and I have faced with the application of ML at Addictive Mobility, will go through some techniques we’ve used to go around these challenges, and will emphasize the tradeoffs that need to be made when managing an ML project. This information will help data science practitioners shift their perspectives from focussing on the most trending ML algorithms to rather paying attention to the applicability of suitable algorithms and manipulating them to make them work in a real life scenario.


A New Universal Deep Learning Approach for Natural Language Processing

Most NLP tasks rely on modelling variable-length sequences of words, not just isolated words. The conventional approach is to formulate these NLP tasks as sequence labelling problems and apply conditional random fields (CRF), convolutional neural networks (CNN) and recurrent neural networks (RNN). In this talk, I will introduce a new, universal deep learning approach applicable to almost all NLP tasks, not limited to sequence labelling problems. The proposed method is built upon a simple but theoretically-guaranteed lossless encoding method, named fixed-size ordinally-forgetting encoding (FOFE), which can almost uniquely encode any variable-length word sequence into fixed-size representation. Next, simple feedforward neural networks are used as universal function approximators to map fixed-size FOFE codes to various NLP targets. This framework is appealing since it is elegant and well-founded in theory and meanwhile fairly easy and fast to train in practice. It is totally data-driven without any feature engineering, and equally applicable to a wide range of NLP tasks. In this talk, I will introduce our recent work to apply this approach to several important NLP tasks, such as word embedding, language modelling, named entity recognition (NER) and mention detection, coreference resolution, Question Answering (QA) and text categorization. Experiments have shown that the proposed approach yields strong performance in all examined tasks, including Google 1-billion-word language modelling, KBP EDL contests, Pronoun Disambiguation Problem (PDP) in Winograd Schema Challenge, factoid knowledge-base QA, word sense disambiguation (WSD). 
11:40 AM Futute of Healthcare in AI (?) Panel Gary Saarenvirta-  CEO, Daisy Intelligence
Roger Grosse , Assistant Professor, Computer Science UofT, founding member, Vector Institute
  Moderator?        Wanda Peteanu                Director of Information Management, MHSc,CHE | AI + Healthcare Education,  UHN         Healthcare AI For Leaders - Managing Risks, Enabling the Future -   What is the present and future of AI in healthcare and clinical education?  What are the risks and which skills are required for the leaders managing ML technical, business and clinical teams? The talk will commence with an overview of the current state of the art in healthcare AI, based on published research and applications released in clinical practice.  We will also cover AI's implications for the clinical environment and education, particularly the newly introduced risks. We will then highlight risk mitigation strategies employed to ensure clinicians and educators have the appropriate information to make decisions.  We will conclude by discussing the role of leaders of AI-enabled projects, how these projects are different than typical software projects and how leaders can support and coach teams to ensure a successful adoption.   

Panalist # 1         David Qixiang Chen        ?        Co-founder & CTO, BenchSci        The quest for the ideal antibody, revolutionize product search in science        "Problem: The antibody, an important part of the immune system, is a widely used reagent of biomedical experiments. Misuse of antibodies, often due to insufficient data, are responsible for up to 50% of failed experiments, and incur enormous cost in time and money for drug discovery. The best evidence of antibody use, and other scientific products,  are found in scientific publications. Existing publication search tools (pubmed, google scholar) are not meant for products. We decoded antibody experimental contexts from open and close-source publications with a combination of text mining, bioinformatics, and machine learning.  Model: At the end of the day, scientists prefer to judge experimental outcome by inspecting the publication images. We linked antibody contexts to its figure image, by identifying the correct product from amongst 4M antibodies, within 9M publications, across 300K contexts, and associate them with over 37M protein aliases. This complex task was computed using Spark and the search served on Elasticsearch. Deep neural nets were used to judge product/context usage relationship (embeddings, LSTM with attention), and to identify technique subpanel (CNN) in figures to fine-tune data accuracy. Results: Our platform is well received by academic and industry scientists. Some of the largest pharmaceuticals in the world are our customers, where their R&D scientists use Benchsci daily. Scientists told us that we have reduced their search time from weeks to minutes, and Benchsci has proven to be a game-changer in scientific product search. Discussions: The mission for BenchSci is to close the gap between idea to outcome in science. We accelerate the pace of discoveries by removing roadblocks in the scientific iteration cycle. ML has proven to be indispensable, where the scaling of data processing with a small team could only have been achieved through the use of deep learning. "

Panalist # 2         Mason Victors        ?        CTO at Recursion Pharmaceuticals        Using ML to Massively Parallelize Drug Discovery        The drug discovery and development process is largely broken, taking a one-by-one hypothesis-driven approach. This traditional approach is too slow and costly to effectively develop the drugs necessary to treat the thousands of diseases and disorders that are still untreatable. Recursion Pharmaceuticals has adopted a phenotypic screening approach to drug discovery that combines the best of biology, high throughput screening and automation, and machine learning in order to transform this into a scalable process, wherein we can find candidate drugs for dozens of diseases in a matter of weeks instead of years. Each week we currently run over 100,000 biological experiments, generating approximately 20TB of fluorescent microscopy imaging data, and use a combination of human engineered features coupled with deep convolutional neural networks to learn a latent representation space of cellular biology. From this latent representation space, we can rapidly assess which compounds are able to effectively reverse the effects of a given disease, and apply this approach to hundreds of diseases spanning multiple therapeutic areas. As a result, Recursion has been able to generate validated compound hits for many diseases in a short period of time, demonstrating yet another application of ML and data science to the field of biology and drug discovery. In this talk, we will present some of the existing problems of drug discovery and development, demonstrate how the use of ML on phenotypic imaging data can overcome some of these challenges, show the results of this process over the past few years, and discuss some of the remaining challenges in applying machine learning to drug discovery and development.

Panalist # 3         Linda Kaleis         ?        Data Scientist MEMOTEXT        Commercializing Machine Learning in Digital Health         "MEMOTEXT, a leader in digital health and patient engagement, will deliver a presentation on the intricacies of commercializing machine learning in digital health with a focus on a methodology that works best when working with non-technical business stakeholders.

The proposed methodology will touch on: defining the business problem; KPI definition and validation; data collection and cleaning; data exploration; modelling analysis; data reviews and results communication; and lastly, actionable recommendations. Specifically, we will discuss how machine learning algorithms can be used to unlock invaluable insights from disparate sources of data in effort to target patients' individual barriers to disease treatment and prevention. Practical implications of embedding machine learning into the design and solution of digital health programs will be addressed - namely, the effect it has on personalization, value for the end user as well as value to the business stakeholder. We will also discuss different types of health data, data science methodologies and credible MEMOTEXT use cases that have been successful and delivered value to stakeholders in the past - both to the patient themselves and to the provider/payor/pharmacy. Examples will include MEMOTEXT's previous work on medication adherence trajectory and outcomes prediction, pattern analysis using process and sequential rule mining, and mixed data clustering for determining distinct patient types.

Specific learnings taken away from working with non-technical audiences will also be presented to highlight key tips and strategies for analysis design, results communication, and ultimately meeting client expectations. There will be a strong focus on the ROI of machine learning to business stakeholders, particularly in the digital health world."

Panalist # 4                                         
#1 Applied Reinforcement Learning for Retail: An Introduction to the Autonomous Enterprise

Reinforcement learning is a branch of machine learning or artificial intelligence that is loosely based on how humans learn through interaction with their environment. As children, we learned through random trial and error and observed its cause and effect. We remember these experiences and through repetition, reinforced the cause and effect learning resulting in memories allowing us to recall what actions to take in similar situations. We continue to learn in this manner throughout our lives.

This tutorial will define the differences between supervised learning and reinforcement learning. Gary will illustrate the mathematics behind reinforcement learning using games/toy examples and discuss the different types of reinforcement learning like model-based learning and types of model free learning. The tutorial will also demonstrate how Daisy uses simulation-based reinforcement learning to assist retailers in making smarter merchandising decisions, including what products to promote each week, what prices to charge for each product at each store and how much inventory of each product to allocate to each store and distribution centre. Daisy’s methods are presented through real-world examples and financial results are shared. The tutorial ends with a future vision of the autonomous enterprise and how the engineering profession will lead the development of practical artificial intelligence applications.
     
Natural Gradient and K-FAC

TBD
12:10 Lunch 
1:30 PM Duncan Stewart,  Director of Research, Deloitte  OPTIONS:   Xiaozhou Wang, Quartic.ai Chief Data Scientist"  OR Kiri Nichol ,Machine Learning Engineer/Researcher at Adeptmind Nathan Killoran Machine Learning & Software Lead, Xanadu
  Next Big Things in Technology, Media & Telecommunications, & The Intersection with Trends in Machine Learning

TBD
#1 Learning Data Science By through Kaggle competitions

Kaggle is the home of Data Science competitions. Successful Kagglers get tickets to jobs in great companies, as well as fame in the community. In this talk, we will walk through the common practice of doing Kaggle competitions, as well as a case study of what it is like to win a Kaggle competition.

Automatic Differentiation Through Quantum Computers

Quantum computers are expected to provide computational advantages in many areas, including machine learning. Yet they are still in their infancy, and remain largely obscure to non-specialists. At the same time, modern machine learning techniques -- such as automatic differentiation and backpropagation -- have evolved to a level which enables even beginners to train powerful models. To spur deeper exploration, adoption, and innovation in quantum computing, it is important to follow the lead of the machine learning community and make software and algorithms which are more powerful and accessible. In this talk, I present methods which allow the end-to-end differentiation of quantum computations. These methods are based on the idea that we can use a quantum computer not just to compute functions, but also to evaluate gradients of those same functions. These techniques will allow users to build and train new types of machine learning models which incorporate -- in whole or in part -- a quantum computer. I also introduce an open-source software library which implements these new ideas. The library supports computations involving both classical simulators and currently available cloud quantum hardware backends.
2:05 PM Fernando Moreira - SVP, Global Insurance at Scotiabank OPTIONS: Dr. Rahmatullah Hafiz   - Lead Researcher, Cognitive Computing, Exiger David Duvenaud Assistant Professor UofT, Founding Member of the Vector Institute
  AI and Exponential Growth - The Problem & The Opportunity

#1 Automated Regulatory Compliance with Cognitive Computing

Successful enforcement of regulatory compliance requires a high precision in the detecting financial crimes including bribery, corruption, money laundering, terrorism financing, fraud, etc. Historically, infrastructures developed by regulatory bodies primarily consist of manual investigation and due diligence. Existing systems also include hard-coded regulatory rules that are painstakingly maintained and curated by manual efforts. These manual research and rule-maintenance are not only very expensive, time consuming, and hard to maintain; but also introduce erroneous amount of costly false positives.

At Exiger, a global authority in regulatory compliance, we developed a set of machine learning (ML) and natural language processing (NLP) powered technologies for automated due-diligence, financial-crime detection, on-boarding screening and transaction monitoring. Our cognitive computing engine is capable of detecting and classifying risk information from completely unstructured text in web by reliably identifying relevant text for the subject, and carefully understanding the context of the text. The contextual understanding helps with extracting valuable information about subjects, which assists us automatically link suspicious subjects in monitory transactions via establishing common adverse incidents.

Our cognitive pipeline employs ML models ranging from decision trees, random forests, CRFs, and RNNs - at various stages. Years of data had been labeled to identify features in the financial crime domain for training models to match and rank relevant structured and unstructured data with respectable F1 scores. CRF-based models are used to perform NLP annotations tasks. Variations of off-the-self neural nets power meaningful information extraction about subjects with very high precision. Prototypical RNNs are trained with contextual semantic information to classify risk with high accuracy. We are fortunate to work closely with our compliance colleagues with nuanced subject matter expertise to train our system. This multi-disciplinary collaboration has helped us pioneering the automaton of the regulatory compliance technology.
Neural Ordinary Differential Equations

We introduce a new family of deep neural network models. Instead of specifying a discrete sequence of hidden layers, we parameterize the derivative of the hidden state using a neural network. The output of the network is computed using a black-box differential equation solver. These continuous-depth models have constant memory cost, adapt their evaluation strategy to each input, and can explicitly trade numerical precision for speed. We demonstrate these properties in continuous-depth residual networks and continuous-time latent variable models. We also construct continuous normalizing flows, a generative model that can train by maximum likelihood, without partitioning or ordering the data dimensions. For training, we show how to scalably backpropagate through any ODE solver, without access to its internal operations. This allows end-to-end training of ODEs within larger models.
14:45 Afternoon Break 
3:15 PM Geordie Rose,  Founder and CEO at Sanctuary.ai, D-Wave and Kindred.ai Prakhar Mehrotra, Walmart Labs Senior Director, Machine Learning - Retail Ihab Ilyas, Professor of Computer Science, Cofounder, University of Waterloo, Tamr
  How Human-Like Robots Will Impact Work, Life and Society

Throughout the entirety of human history we have always dreamed about, and been captivated by, the idea of creating machines that are like us. Human-like robots have appeared in endless science fiction movies, in both utopian and – more often – dystopian narratives. I will talk about advancements in robotics and AI that are enabling robots to become increasingly human-like.
Walmart's Retail Journey: From Business Intelligence to Artificial Intelligence

The objective of this keynote/talk is to take the audience through the AI transformation journey of the world's biggest retailer.  Currently, we are tackling the two most important problems in brick & mortar retail: intelligent pricing, and assortment optimization. Causal nature of these problems demand need for causal inference paradigm. The human decision making requires highly interpretable models. I will discuss how the machine learning group is trying to balance interpretability with accuracy.  
Scalable Machine Learning for Data Cleaning  

Who is this presentation for?:
CIOs, CDOs, VPs of data management, and any other senior IT leaders
Prerequisite knowledge:
A basic understanding of data management and data management technologies
What you'll learn:
Learn how to curate data at scale to enable transformational analytics and business outcomes
Description:
Machine learning tools promise to help solve data curation problems. While the principles are well understood, the engineering details in configuring and deploying ML techniques are the biggest hurdle. Ihab Ilyas explains why leveraging data semantics and domain-specific knowledge is key in delivering the optimizations necessary for truly scalable ML curation solutions.

Ihab focuses on two main problems: entity consolidation, which is arguably the most difficult data curation challenge because it is notoriously complex and hard to scale, and using probabilistic inference to enrich data and suggest data repair for identified errors and anomalies. The problem statement in both cases sounds deceptively simple: find all the records from a collection of multiple data sources that refer to the same real-world entity or use trusted data sources to suggest how to correct errors. However, both problems have been challenging researchers and practitioners for decades due to the fundamentally combinatorial explosion in the space of solutions and the lack of ground truth.

There’s a large body of work on this problem by both academia and industry. Techniques have included human curation, rules-based systems, and automatic discovery of clusters using predefined thresholds on record similarity Unfortunately, none of these techniques alone has been able to provide sufficient accuracy and scalability. Ihab provides deeper insight into the entity consolidation and data repair problems and discusses how machine learning, human expertise, and problem semantics collectively can deliver a scalable, high-accuracy solution.
3:45 PM PANEL  Carolina Bessega - Carolina Bessega Chief Scientific Officer, co-founder at Stradigi AI David Madras- University of Toronto, Vector Institute
  The Realities of Production-izing ML Models:                                                                                                                                                                                                                
Moderator?                                                                                                                                                                                                             

Panalist # 1         Sandy Ward                Technical Team Lead, Integrate AI         The Realities of Productionizing ML Models        "Many developers and business think creating a model is the difficult part but in reality integrating into an existing business's workflow to use the prediction is often much  a lot more work. At integrate.ai we have launched models into production with many companies (including: Scotiabank, Corus, Kanetix, and more) and I want to share these stories so it is easier to us ML! I have presented this framework to the Toronto CTO group -- and got feedback from many companies -- and had a great response to how it was 'real'.

The discussion would be best if it was during the talk, there are a number of areas where the audience could chime in with their experience. "                                                                                                                                                                       
Panalist # 2         Rebecca Tessier                 Data Science Lead, at Shopify        Building Data Foundations for Machine Learning         This panel will focus on best practices for building data assets and infrastructure for implementing quality machine learning models at scale. While there is often a lot of discussion around algorithmic approaches to machine learning, the quality and investment of time in creating solid data assets that are resilient to data changes over time is equally important for generating good predictions. Some topics for discussion may include: best practices for feature engineering, how to develop ETLs for training data sets, champion/challenger frameworks for re-training production models, etc.
                                                                                                                                                                       
  
                                                                                                                                                                     
Panalist # 4         Diane Reynolds        ?        Data Scientist, IBM        Client Insight Analytics in the Cloud        I've worked in financial services sector analytics for over 20 years, but one of my most interesting projects began 2 years ago when a small team set out to transform a client insight analysis tool based on traditional software tools into a fully cloud-based solution.  In this talk, I'll share with you our experience in moving from relational databases to big data, how we achieved multi-tenancy despite the prevalence of sensitive data, and how we're leveraging a variety of machine learning platforms and algorithms.  You'll walk away with insights into how to 'productize' your machine learning algorithms on the cloud.                                                                                                                                                                       


# 1 How to Leverage your Data to get the Best Recommendations

Machine learning based recommender systems are an important component in different industries, from retail-banking to movie recommendations, they can be key to business success. During this talk, we will discuss what is really important when it comes to designing a recommender system. What data should you take into account? What are the most common problems you can expect? From the well known cold start problem to different attacks that a recommender system can suffer, we will dive into solutions for these issues and give practical advice on how to design, deploy and test a recommender system.

We will describe a technique to improve accuracy and coverage, while keeping recommendations real-time. Our technique significantly improves the accuracy of the recommendations by mitigating the sparsity of the dataset providing an original solution to the cold-start problem. We also improve the coverage at no accuracy cost by favouring less popular items by applying the similarity translation mapping.

Finally, by modelling item–item rather than user–item correlations, we are able to update the recommendations for a given user in real-time, without re-training, as the user’s history receives new entries.
Predict Responsibly: Improving Fairness and Accuracy by Learning to Defer

In many machine learning applications, there are multiple decision-makers involved, both automated and human. The interaction between these agents often goes unaddressed in algorithmic development. In this talk, we explore how to model this interaction in learning, and how we can use information about this interaction to improve the performance of systems containing machine learning models.

We discuss a simple two-stage framework containing an automated model and an external decision-maker. The model can choose to say "Pass", and pass the decision downstream. This model reflects many real-world situations where machine learning is applied for decision-making.

We propose a method for optimizing a model within an interactive system: “learning to defer”, which generalizes previous methods by considering the effect of other agents in the decision-making process. We discuss how learning to defer can help an algorithm account for potential biases held by external decision-makers. In experiments, we consider how a model would interact with different types of users, who may have access to external information, but may be irrational: biased or inconsistent.

Our experiments demonstrate that learning to defer can greatly improve the accuracy and decrease the bias of an entire system. We conclude that when considering the “fairness” of a machine learning model, it is important to understand the context in which the model is used, and the nature of its interaction with those who use it.
     
     
     
Day 2      
8:45 AM   Attendee Registration and Sponsor Booths Open  
9:45 AM   Opening Welcome: David Scharbach  
9:55 AM   Keynote Address: Philippe Beaudoin  SVP Research Group and Co-Founder at Element AI  
10:10 AM   Keynote Address 2:  Surprise Guest  
10:25 AM   Keynote Address 3:  Surprise Guest  
Break
11:10 AM Diane Reynolds, Chief Data Scientist, Client Insights, Financial Services Sector IBM   Isar        Nejadgholi        IMRSV Data labs        Head of Machine Learning Research    Anne Martel Professor, Medical Biophysics, UofT, Senior Scientist at Sunnybrook Research
  Client Insight Analytics in the Cloud

I've worked in financial services sector analytics for over 20 years, but one of my most interesting projects began 2 years ago when a small team set out to transform a client insight analysis tool based on traditional software tools into a fully cloud-based solution.  In this talk, I'll share with you our experience in moving from relational databases to big data, how we achieved multi-tenancy despite the prevalence of sensitive data, and how we're leveraging a variety of machine learning platforms and algorithms.  You'll walk away with insights into how to 'productize' your machine learning algorithms on the cloud.
#1 How can natural language processing be applied to identify toxic online conversions?        "With the sheer volume of online content, we are plagued by our current inability to effectively monitor its contents. Social media platforms are ridden with verbal abuse, giving way to an increasingly unsafe and highly offensive online environment. With the threat of sanctions and user turnover, governments and social media platforms currently have huge incentives to create systems that accurately detect and remove abusive content.
When considering possible solutions, the binary classification of online data, as simply toxic and non-toxic content, can be very problematic. Even with very low error rates of misclassification, the removal of said flagged conversations can impact a user's reputation or freedom of speech.  Developing classifiers that can flag the type and likelihood of toxic content is a far better approach. It empowers users and online platforms to control their content based on provided metrics and calculated thresholds. While a multi-label classifier would yield a more powerful application, it's also a considerably more challenging natural language processing problem. Online conversational text contains shortenings, abbreviations, spelling mistakes, and ever evolving slang. Huge annotated datasets are needed so that the models can learn all this variability across communities and online platforms.
In our work, we used the Wikimedia Toxicity dataset to train models that can flag toxicity types such as insult, identity hate and thread. We considered stacking of multiple neural network models that can learn sentence labels through training recurrent and attention layers and reached 0.9862 ROC AUC score.  We also pruned the stacked model to be efficiently deployed in real time and studied how the model performs across subgroups of data and other publicly available datasets that contain online content. Our results shed light on the discussion around how automatic labeling of online conversations can be used to make social media safer and more inclusive environments. "
Machine Learning in Medical Imaging

Recent advances in machine learning in general, and deep learning in particular, have transformed the field of medical image analysis. In applications ranging from image reconstruction, to the detection and diagnosis of disease, neural networks are outperforming more traditional methods of analysis. As well as having profound implications for clinicians and researchers, this has led to an explosion of new companies who are developing medical applications built on imaging data. This talk will provide a brief overview of the field and will provide some case studies in computer aided diagnosis and survival prediction in breast cancer with MRI and digital pathology.
11:40 AM Ethics & AI: Pursuing Responsible Innovation: Panel   Irina Kezele, Director of AI, ModiFace Alán Aspuru-Guzik Professor, Faculty Member, Vector Institute
  "As companies develop and seek to commercialize novel applications of AI in areas such as recidivism, hiring, and public policy, there is increasing concern that these automated decision-making systems may unconsciously duplicate social inequalities and biases, with unintended societal consequences. This panel will explore how companies can counteract such prejudices by adopting an ethics-based approach to innovation. Drawing on their experiences working for, or with companies, engaging in bleeding edge AI innovation, the panelists will discuss:
        The key principles for responsible innovation
        Integrating ethical training and awareness within organizations
        The importance of identifying data bias
        Designing algorithms to address bias and, potentially, reduce it in some cases
        Algorithmic accountability"

Moderator:        Ozge Yeloglu                 Chief Data Scientist, Microsoft Canada      

Panalist # 1         Laila Paszti                Attorney | AI Software Engineer        Investment in AI: Due Diligence Considerations        "The session will cover the areas of legal due diligence specific to AI in merger and acquisition (M&A) or in other transactions where a company’s AI software or services are a key consideration. An understanding of the AI-related issues/risks described below are key to a successful transaction. This session will describe ways for buyers/investors to mitigate such issues/risks through both software and legal remediation and the negotiating of favourable deal terms (e.g. representations and warranties, closing conditions, etc.). This session will also describe best practices for future sellers/investees to implement when designing/maintaining their AI solutions to minimize such concerns for buyers/investors.
        Licensing and Compliance Risk: legal compliance with licenses of third party datasets/software utilized in AI solutions, including public datasets and open source software
        Privacy Risks: Compliance with Canadian/US/EU data protection regulation
        Cybersecurity Risks"

Panalist # 2         Karen Bennet                 VP Engineering, Cerebri AI        Building AI for Ethics and Privacy        There are many AI deployments that are failing on the ethics and privacy front. if you don't build in design, deployment and validation guidelines, it's not easy to find these issues once deployed.  This talk will explain how to  design,  build and validate to prevent failures like fake news, diversity bias, security  breaches,  and chat bots which learn negative gender stereotypes. Designing and validating models will help prevent assumptions being made by the individuals designing the models.   The building of models has been running fast and free with limited ethical and privacy guidelines.  The GDPR is the start of these guidelines being put in place and industry is talking more and more about  ethics and privacy concerns  but putting in general practice is where the issue is.   This isn't surprising based on industry's lack of diversity in depth and without guidelines they aren't as attuned to ethical and privacy concerns.  Come learn about guiding practices to address these concerns.  The future of AI needs to embrace ethics and privacy by design, deployment  and validation.  to reach it's full potential.

Panalist # 3         Hashiam Kadhim                Machine Learning Engineer - Dessa        Evolving AI for your enterprise        In the six years since deep learning first emerged as the state-of-the-art technique for tasks like image recognition, the AI field has seen tremendous growth and increasingly percolated into the business world. Today, over 60% of large enterprises have at least one pilot project in production, but the next big challenge will be scaling AI throughout their organizations. While developing advanced AI solutions that solve real business problems has become easier than ever, there are still many challenges that enterprises must overcome to unlock the technology’s full value. In this talk, Hashiam Kadhim, a Lead Machine Learning Engineer at Dessa, will address how enterprise teams can ensure AI projects result in transformative impact for their business.

Panalist # 4         Kathryn Humne        ?                                                    
Generative Adversarial Networks for Live Makeup Augmentation

# 1 Real-time virtual makeup try-on is becoming an essential component in beauty e-commerce as well as in-store shopping experience. It gives users easy tools to tune relevant attributes of the product (e.g. color and glossiness) according to their personal preferences.

Under the hood, traditional solutions to this problem involve two steps: detecting facial landmarks and using them to overlay augmented makeup. This approach has several limitations. First, errors in facial landmark detection cause incorrect makeup alignment. Second, correctly simulating all the physical properties of the augmented makeup is challenging. Third, blending the augmented make-up and the original image is not trivial. Due to these limitations and others, this solution cannot easily scale to a large amount of products while maintaining realism, and it would therefore be convenient to have an end-to-end model that can learn a conditional makeup space on its own.

We opt for the CycleGAN architecture with a number of modifications to allow for multi-directional image translation, and to ensure greater stability in training. Given an input image without makeup and an encoding of the desired makeup, we conditionally train a generator to produce a realistic image with makeup while preserving all the other input image properties, such as person identity, head pose, and facial expression. We show that the resulting approach does not suffer from the limitations of the aforementioned, standard approach, and is easily extendable to support an arbitrary number of makeup products.
Title TBA

Abstract TBA
  Lunch 
1:30 PM Prajakta Kharkar - Senior Economist (leading AI Partnerships strategy, Data valuation strategy) - TELUS
Xavier Snelgrove - Applied Research Scientist at Element AI Nicholas Frosst, Researcher Google Brain
  #1 Big data - from competition to conglomeration

In the course of my work in Big Data and AI, I've noticed an unmistakable trend towards large data partnerships forming across industries. Such examples are sprouting up everywhere - RBC and WestJet launched a joint loyalty program, Google purchased MasterCard data for millions to be able to track offline purchases, the World Bank published a study where a bank and a telco collaborated to use mobile calling data to accurately predict loans with highest chances of repayment - these are all examples of large custodians of data coming together to merge unrelated, proprietary datasets to uncover unprecedented insights. The future in my opinion belongs to 'data conglomerates'. I'd like my speech to be about this insight. "
Generative Adversarial Networks for Live Makeup Augmentation

As AI algorithms become more common in the world, there is a recurring critique  that they're "black-boxes": we can't evaluate how they make their decisions and that this makes them difficult to trust, debug, or wield as tools. This can be a major obstacle to deploying algorithms, especially in regulated contexts.

The question of understanding "why" a model makes a decision turns out to be deep, as it touches on the problem of the nature of knowledge. In the end an explanation is for a *human*, so explainability research has strong connections to other fields such as design and education, with all of the complexity and ambiguity that comes with this association.

In this talk we'll look at different approaches to AI explainability today, and exciting research frontiers. We'll discuss the appropriateness of different kinds of explanations in different contexts, the dangers of convincing but inaccurate explanations, and how explanations can help in other fields like AI ethics, and human computer interaction.
Title TBA
TBA
2:00 PM Design Thinking and AI First Products  Choosing your technology stack with, and making it AI Capable : Panel  Pascal Poupart Professor, David R. Cheriton School of Computer Science, U of Waterloo
  Moderator?        Ramy Nasser                Director, Retail Innovation Lab, Mattel, Inc.        Why Machine Learning Needs Design Thinking        "Machine Learning is accelerating at a blistering pace, but with that comes a new class of problem - one that algorithms (at least so far) can’t solve - the human problem. At the end of the day, many machine learning tools & technologies will impact people at a personal level, and keeping that in mind when designing solutions will become increasingly critical as we progress. Design Thinking is the glue that contextualizes the technical capabilities of machine learning in order to build meaningful solutions & customer experiences. Working hand­-in-hand, designers, business leaders and data scientists can collaborate to better understand user behaviours, as well as the desires of large populations­ while creating highly personalized products & services.

In this talk I’ll leverage real­ world case studies (both projects I've worked on as well as other industry examples) as well as some theory and best practices for how to incorporate design thinking to take full advantage of machine learning. While most AI presentations focus on technical implementations or over­simplified case studies, I hope that this session presents conference participants with a different perspective and approach. I’d be happy to give this presentation as a talk or to support a planned panel discussion, bringing a unique perspective to the conversation. "

Panalist # 1         Amin Bashi                Product Lead,CrowdRiff        Building AI-first Products         Building an AI-first product is hard, but when it wins, it wins flat out. AI-first products are products that just would not make sense without AI. Our 2016 cohort of customers are getting an 86% average increase in impressions and engagement. Yet, we have an 46% average decrease in product usage among this cohort of our customers. We have minimized human monitoring, analysis and manual optimization. Our product can exist in the background with little or no interaction. Similar to mobile-first approach, you can't just add AI to the top of your product. To build an AI-first product, you need to completely reimagine and rebuild your product. We had to rebuild all of our product's critical events but in less than a year, we managed to build the world's first AI-powered visual content engine that empowers marketers to find top-performing visuals in a matter of seconds. We invested our resources into building an AI-first product because we realized that our product's competitive advantage comes from applying AI to our data and innovating on our business model.

Panalist # 2       Steve        Kludt        "Canvass Analytics        CTO"        AI and the New Industrial Revolution        "The rise of the new industrial revolution, coined Industry 4.0, and Artificial Intelligence has created a hot bed of innovation for the Industrial sector. Today with AI-predictive analytics, industrial manufacturers can optimize processes, identify potential impacts to quality as well as predict asset failure, to name a few. But this is just the tip of the iceberg. AI can bridge the gap between current forms of automation and learning with more advanced forms in the near future.

As techniques and algorithms underlying AI have continued to be developed, achieving industrial automation is just around the corner. Soon, AI-powered systems will not just notify but eventually control them before it impacts the quality of the product, using Reinforcement Learning models that learn exactly how to tweak parameters and inputs to optimize a process. From here, an auto plant can determine the optimal order of operations to increase throughput; a steel manufacturer can reduce scrap by reducing the number of errors in the production process, and energy plants can increase generation by precisely pinpointing what direction its wind turbines should face and the angle to pitch its blades at. Using Reinforcement Learning, a manufacturer can optimize their processes, dramatically scale up its operations with the same amount of manpower and reduce operating costs.

Hear from Steve Kludt, CTO for Canvass Analytics, as he previews the use cases of the Industrial future as AI adoption continues to take hold. "                                 
Panalist # 3   CEO Dessa ?                                      
Panalist # 4                                         
Moderator?        Rupinder Dhillon                 Rupinder Dhillon - Machine Learning and AI at Bell Canada      

Panalist # 1         Terry Hickey                  Chief Analytics Officer, CIBC               

Panalist # 2         Brian Keng        ?        Chief Data Scientist, Rubikloud        Building AI Products for the Enterprise        "Enterprise software projects are notoriously challenging endeavours due to the intricacies of interfacing with large enterprise organizations. These challenges partially arise from technological issues where legacy, on-premise, consultant-driven, custom software projects are the norm. The other major category of challenges arise from company-specific business problems and workflows necessitating the need for highly-customized solutions. This creates a natural tension between building shallow, highly generic enterprise products and one-off deep, highly customized solutions.

At the same time, the success of AI in the consumer space has slowly started to permeate into the enterprise. However, AI adds another layer of complexity to
these types of projects that only magnifies the already existing challenges of building enterprise products.

This talk will address some of the core issues around building enterprise-scale AI products based on our extensive experience deploying to some of the largest retailers around the world. It will cover a wide array of topics from data
integration to team organization to modelling challenges that are applicable to both internal and external practitioners who are working at the intersection of AI and enterprise."

Panalist # 3         Abraham Kang        ?        Senior Director Software at Samsung Research America        Identifying and Remediating Security Vulnerabilities in AI/ML Applications        Adversarial Samples/Patching, Trojan Network/BadNets, and Model and Training Data Extration Techniques.  Come to this talk to better understand how attackers are targeting your AI assistant and ML applications.  Gain detailed knowledge in the techniques used by the bad guys and possible defenses.

Panelist # 4        Deloitte?    
                          
Panalist # 5        Diane Reynolds        ?        Data Scientist, IBM        Client Insight Analytics in the Cloud        I've worked in financial services sector analytics for over 20 years, but one of my most interesting projects began 2 years ago when a small team set out to transform a client insight analysis tool based on traditional software tools into a fully cloud-based solution.  In this talk, I'll share with you our experience in moving from relational databases to big data, how we achieved multi-tenancy despite the prevalence of sensitive data, and how we're leveraging a variety of machine learning platforms and algorithms.  You'll walk away with insights into how to 'productize' your machine learning algorithms on the cloud.
Unsupervised Video Object Segmentation for Deep Reinforcement Learning

Deep reinforcement learning (RL) in visual domains is often sample inefficient since the agent is implicitly learning to extract useful information from raw images while optimizing its policy. Furthermore the resulting policy is often a black box that is difficult to explain. I will present a new technique for deep RL that automatically detects moving objects and uses the relevant information for action selection. The
 detection of moving objects is done in an unsupervised way by exploiting structure from motion. Over time, the agent identifies which objects are critical for decision making and gradually builds a policy based on relevant moving objects. This approach, which
 we call Motion-Oriented REinforcement Learning (MOREL), is demonstrated on a suite of Atari games where the ability to detect moving objects reduces the amount of interaction needed with the environment to obtain a good policy. Furthermore, the resulting policy
 is more interpretable than policies that directly map images to actions or values with a black box neural network. We can gain insight into the policy by inspecting the segmentation and motion of each object detected by the agent. This allows practitioners
 to confirm whether a policy is making decisions based on sensible information. 
14:45 Afternoon Break 
3.15 PM Futute of Healthcare in AI (?) Panel   Ethan Fetaya University of Toronto        Post-doc
  Moderator?        Wanda Peteanu                Director of Information Management, MHSc,CHE | AI + Healthcare Education,  UHN         Healthcare AI For Leaders - Managing Risks, Enabling the Future -   What is the present and future of AI in healthcare and clinical education?  What are the risks and which skills are required for the leaders managing ML technical, business and clinical teams? The talk will commence with an overview of the current state of the art in healthcare AI, based on published research and applications released in clinical practice.  We will also cover AI's implications for the clinical environment and education, particularly the newly introduced risks. We will then highlight risk mitigation strategies employed to ensure clinicians and educators have the appropriate information to make decisions.  We will conclude by discussing the role of leaders of AI-enabled projects, how these projects are different than typical software projects and how leaders can support and coach teams to ensure a successful adoption.   

Panalist # 1         David Qixiang Chen        ?        Co-founder & CTO, BenchSci        The quest for the ideal antibody, revolutionize product search in science        "Problem: The antibody, an important part of the immune system, is a widely used reagent of biomedical experiments. Misuse of antibodies, often due to insufficient data, are responsible for up to 50% of failed experiments, and incur enormous cost in time and money for drug discovery. The best evidence of antibody use, and other scientific products,  are found in scientific publications. Existing publication search tools (pubmed, google scholar) are not meant for products. We decoded antibody experimental contexts from open and close-source publications with a combination of text mining, bioinformatics, and machine learning.  Model: At the end of the day, scientists prefer to judge experimental outcome by inspecting the publication images. We linked antibody contexts to its figure image, by identifying the correct product from amongst 4M antibodies, within 9M publications, across 300K contexts, and associate them with over 37M protein aliases. This complex task was computed using Spark and the search served on Elasticsearch. Deep neural nets were used to judge product/context usage relationship (embeddings, LSTM with attention), and to identify technique subpanel (CNN) in figures to fine-tune data accuracy. Results: Our platform is well received by academic and industry scientists. Some of the largest pharmaceuticals in the world are our customers, where their R&D scientists use Benchsci daily. Scientists told us that we have reduced their search time from weeks to minutes, and Benchsci has proven to be a game-changer in scientific product search. Discussions: The mission for BenchSci is to close the gap between idea to outcome in science. We accelerate the pace of discoveries by removing roadblocks in the scientific iteration cycle. ML has proven to be indispensable, where the scaling of data processing with a small team could only have been achieved through the use of deep learning. "

Panalist # 2         Mason Victors        ?        CTO at Recursion Pharmaceuticals        Using ML to Massively Parallelize Drug Discovery        The drug discovery and development process is largely broken, taking a one-by-one hypothesis-driven approach. This traditional approach is too slow and costly to effectively develop the drugs necessary to treat the thousands of diseases and disorders that are still untreatable. Recursion Pharmaceuticals has adopted a phenotypic screening approach to drug discovery that combines the best of biology, high throughput screening and automation, and machine learning in order to transform this into a scalable process, wherein we can find candidate drugs for dozens of diseases in a matter of weeks instead of years. Each week we currently run over 100,000 biological experiments, generating approximately 20TB of fluorescent microscopy imaging data, and use a combination of human engineered features coupled with deep convolutional neural networks to learn a latent representation space of cellular biology. From this latent representation space, we can rapidly assess which compounds are able to effectively reverse the effects of a given disease, and apply this approach to hundreds of diseases spanning multiple therapeutic areas. As a result, Recursion has been able to generate validated compound hits for many diseases in a short period of time, demonstrating yet another application of ML and data science to the field of biology and drug discovery. In this talk, we will present some of the existing problems of drug discovery and development, demonstrate how the use of ML on phenotypic imaging data can overcome some of these challenges, show the results of this process over the past few years, and discuss some of the remaining challenges in applying machine learning to drug discovery and development.

Panalist # 3         Linda Kaleis         ?        Data Scientist MEMOTEXT        Commercializing Machine Learning in Digital Health         "MEMOTEXT, a leader in digital health and patient engagement, will deliver a presentation on the intricacies of commercializing machine learning in digital health with a focus on a methodology that works best when working with non-technical business stakeholders.

The proposed methodology will touch on: defining the business problem; KPI definition and validation; data collection and cleaning; data exploration; modelling analysis; data reviews and results communication; and lastly, actionable recommendations. Specifically, we will discuss how machine learning algorithms can be used to unlock invaluable insights from disparate sources of data in effort to target patients' individual barriers to disease treatment and prevention. Practical implications of embedding machine learning into the design and solution of digital health programs will be addressed - namely, the effect it has on personalization, value for the end user as well as value to the business stakeholder. We will also discuss different types of health data, data science methodologies and credible MEMOTEXT use cases that have been successful and delivered value to stakeholders in the past - both to the patient themselves and to the provider/payor/pharmacy. Examples will include MEMOTEXT's previous work on medication adherence trajectory and outcomes prediction, pattern analysis using process and sequential rule mining, and mixed data clustering for determining distinct patient types.

Specific learnings taken away from working with non-technical audiences will also be presented to highlight key tips and strategies for analysis design, results communication, and ultimately meeting client expectations. There will be a strong focus on the ROI of machine learning to business stakeholders, particularly in the digital health world."

Panalist # 4                                         
#1 Monica        Holboke        "CryptoNumerics        CEO / Founder"        Learn how to leverage proprietary datasets while protecting privacy and intellectual property        

Data is the key ingredient in any AI strategy. But some of the most valuable datasets are inaccessible because they reside in silos created by privacy concerns, a fear of loss of intellectual property, regulatory requirements and contractual obligations. This is most common in highly regulated industries such as healthcare and financial services. Siloed data hampers organizations ability to accelerate their data-driven innovation.  However, by employing state of the art techniques in cryptography and numerical methods, organizations can collaborate on building statistical and machine learning models with decentralized siloed data to generate superior insights while preserving privacy and IP. 
#1 Learning Deep Neural Networks with Discrete Weights

Deep neural networks are very successful, but can be problematic to run in real time on limited hardware. One possible solution is training neural networks with binary or ternary weights that can run much faster (and with lower memory cost) on dedicated simple hardware. Training such networks can be hard as standard gradient-based methods cannot be used in this discrete setting. I will talk about current leading methods for training such networks, including my work using the local reparametrization trick.