TMLS Micro-Summit #1

ML Engineering, Infrastructure & Productization

 

Reserve you spot today

RESERVE YOUR SPOTLEARN MORE

Following up on the sold-out November TMLS Conference, the team has organized a specially focused event on Machine Learning Engineering, Infrastructure & Productization.

This unique micro-summit will host 2 keynotes exploring the productization of AI and organizing ML Projects followed by 6 lightning case study talks on ML Architecture, ML Team Organization, ML Ops. Through 6 vendor showcase/demos, attendees will explore tools and platforms directly from Google Cloud, AWS, Microsoft Azure and startups from Toronto.

Attendees will have the opportunity to interact with the lightning case study presenter in a follow-up open discussion.

#1 ML Engineering, Infrastructure & Productization

 Feb 12th, 2020

5:45 PM to 8:45 PM (EST)

#2 ML in Retail

 Apr. 12th, 2020

5:45 PM to 8:45 PM (EST)

Reservation TBD

 

#3 ML Ottawa

 May, 2020

Location: Ottawa

Reservation TBD

 

#4 ML Computer Vision/ Autonomous Vehicles

 May 14th, 2020

5:45 PM to 8:45 PM (EST)

Reservation TBD

 

#5 ML in Finance

 June 17th, 2020

5:45 PM to 8:45 PM (EST)

Reservation TBD 

 

#6 ML Engineering, Infrastructure & Productization

 June 19th 2020

9:00 AM to 6:00 PM (EST)

Reservation TBD 

#7 ML Vancouver

 Sept, 2020

Location: Vancouver

Reservation TBD

 

#8 ML Techniques & Applications

 Sept 30th 2020

5:45 PM to 8:45 PM (EST)

Reservation TBD

 

Hear about TMLS November and future events! 

Bio

Bio

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Nunc ornare enim in tempor luctus. Nam mollis nisi sapien, vitae lacinia magna consectetur ut. Ut rhoncus finibus tempor. Pellentesque feugiat, elit at posuere congue, tellus sem sodales ante, ut sagittis orci augue in tortor. Donec hendrerit vel purus vitae ullamcorper. Pellentesque ultricies augue et finibus faucibus. Sed blandit volutpat commodo. Sed nec ligula et enim pulvinar accumsan. Ut maximus arcu eget massa facilisis euismod at scelerisque magna. Nulla interdum nulla lacus, eget placerat nisi vulputate at. Duis non felis viverra, ornare metus id, vehicula ligula. Pellentesque cursus ante vel erat convallis commodo.

Talks

Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Varius sit amet mattis vulputate enim nulla aliquet. Sed nisi lacus sed viverra.

You have Successfully Subscribed!

Chris Wiggins

Chief Data Scientist, NY Times

Chris Wiggins is an associate professor of applied mathematics at Columbia University and the Chief Data Scientist at The New York Times. 

At Columbia, he is a founding member of the executive committee of the Data Science Institute, and of the Department of Applied Physics and Applied Mathematics as well as the Department of Systems Biology, and is affiliated faculty in Statistics.

He is a co-founder and co-organizer of hackNY (http://hackNY.org), a nonprofit which since 2010 has organized once a semester student hackathons and the hackNY Fellows Program, a structured summer internship at NYC startups.

Prior to joining the faculty at Columbia he was a Courant Instructor at NYU (1998-2001) and earned his Ph.D. at Princeton University (1993-1998) in theoretical physics. He is a Fellow of the American Physical Society and is a recipient of Columbia's Avanessians Diversity Award.

Talk: Data Science at The New York Times

TBD

You have Successfully Subscribed!

Nicholas Frosst

rSWE, Google Brain

Nick Frosst is a research engineer working at Google Brain in Geoffrey Hinton's Lab. He received his undergraduate from the University of Toronto in computer and cognitive science. He focuses on capsules networks, adversarial examples and understanding representation space.

Talk: TBD

TBD

You have Successfully Subscribed!

Kathryn Hume

Director of Product & BD, Borealis AI

Kathryn Hume leads product and business development for Borealis AI, the machine learning research lab for the Royal Bank of Canada. Prior to joining Borealis AI, Kathryn held leadership positions at integrate.ai and Fast Forward Labs (acquired by Cloudera). She has helped over 50 Fortune 500 companies develop machine learning applications and is a recognized expert in the ethical and responsible deployment of AI.

Kathryn speaks and writes frequently on AI, with work featured at TED, the Globe & Mail, and the Harvard Business Review. She holds a Ph.D. in Comparative Literature from Stanford University, speaks seven languages, and has given lectures on AI and professional ethics at the Harvard Business School, MIT, Stanford, and the University of Calgary Faculty of Law.

Talk: TBD

Making AI work in the enterprise requires more than talented scientists and large data sets. AI is as powerful and popular as it is because it can automate tasks for which precise rules are too hard to describe. But loosening this constraint also introduces new trade-offs executives need to understand to manage AI projects successfully.



This talk introduces the trade-offs around accuracy, error, explainability, privacy and fairness executives need to understand so they can guide technology teams and manage the risks of applying AI in real-world applications.

You have Successfully Subscribed!

Tomi Poutanen

Co-founder, Layer6

Short bio

Talk: TBD

TBD

You have Successfully Subscribed!

William Falcon

Researcher, Facebook AI

William is the creator of PyTorch Lightning and an AI PhD student at Facebook AI Research and NYU CILVR lab. Before his PhD, he Co-founded AI startup NextGenVest (acquired by Commonbond), and spent time at Goldman Sachs, Bonobos and US Navy. He received his BA in Stats/CS/Math from Columbia University.

Talk: PyTorch Lightning - Squeezing maximum performance from PyTorch models

In recent years, techniques such as 16-bit precision, distributed training, and accumulated gradients have allowed models to train in record times. In this tutorial, I'll introduce PyTorch Lightning, a lightweight PyTorch wrapper around PyTorch, which allows researchers to add these features to their already existing code, and explain how they work. By the end of the tutorial, attendees will know how to quickly prototype and scale new ideas using PyTorch Lightning.

You have Successfully Subscribed!

David Duvenaud

Assistant Professor, UofT

David Duvenaud is an assistant professor in computer science and statistics at the University of Toronto. He holds a Canada Research Chair in generative models. His postdoctoral research was done at Harvard University, where he worked on hyperparameter optimization, variational inference, and chemical design. He did his Ph.D. at the University of Cambridge, studying Bayesian nonparametric with Zoubin Ghahramani and Carl Rasmussen.

David spent two summers in the machine vision team at Google Research, and also co-founded Invenia, an energy forecasting, and trading company. David is a founding member of the Vector Institute and a Faculty Fellow at ElementAI.

Talk: Neural Stochastic Differential Equations

Time series in finance, population genetics, and physics are often naturally modeled by stochastic differential equations (SDEs). We'll show how SDEs can be fit by backpropagation in a scalable way, allowing one to fit large models quickly. We'll give examples of how to build models in this framework, and discuss the pros and cons of this and other time-series modeling approaches.

You have Successfully Subscribed!

Marzyeh Ghassemi

Researcher, Vector Institute

Short Bio

Talk: TDB

TBD

You have Successfully Subscribed!

Mladen Kovacevic

Senior Solutions Architect, Databricks

Mladen Kovacevic is a Senior Solutions Architect at Databricks that has helped dozens of clients spanning data engineers, data scientists and data analysts fully realize the potential of Apache Spark, MLflow and Delta Lake on the cloud by delivering robust engineering and AI solutions. Mladen has been building solutions using Apache Spark since 2014 and has been a contributor to several open-source Apache projects in the Big Data space.

He is a published O'Reilly author who speaks at various events and throughout his career has worked as a software developer, performance analyst, consultant and solutions architect.

Talk: TBD

TBD

You have Successfully Subscribed!

Razvan Peteanu

Lead Architect, ML, TD Securities

Razvan Peteanu's current role is Lead Architect - Machine Learning at TD Securities. He has 25 years of experience in software development, mostly in the financial industry. His focus over the last several years has been on building scalable machine learning solutions, in the cloud or on premise.

Talk: Scaling Machine Learning

Between the laptop carried by the prototypical data scientist and large clusters in the cloud, today's technology offers several options for scaling ML. We'll review them and discuss how to choose between them.

You have Successfully Subscribed!

Helen Ngo

ML Engineer, Dessa

Helen Ngo is a machine learning engineer at Dessa, a Toronto-based artificial intelligence company, and a 2019 Fellow at the Recurse Center in New York City. She co-organizes the Toronto Women’s Data Group and was named a Sidewalk Toronto Fellow as part of the Sidewalk Labs and Waterfront Toronto joint initiative. Previously, she worked as a data scientist in the telecommunications industry and was part of the editorial staff at Towards Data Science.

Talk: Panel

TBD

You have Successfully Subscribed!

Simona Gandrabur

Sr. Director, AI Lead of the Wealth Division, National Bank of Canada

Simona Gandrabur has been working in the general field of AI for close to 20 years, most notably in areas related to the processing of human languages – such as automatic speech recognition, natural language understanding, machine translation, and conversational reasoning. Her experience ranges from many years in research, in the development of smart assistant applications, to defining strategy of AI-based offers. She is currently the head of AI strategy within the Wealth Management division of the National Bank of Canada. 

Talk: TBD

TBD

You have Successfully Subscribed!

Raj Verma

Senior Staff Engineer at Uber

Raj Verma is a Senior Staff Engineer at Uber and head of Uber's FinTech team in Toronto where he leads the development of financial forecasting and risk management solutions using machine learning and data science models.   Before joining Uber Raj was the founder and president of risk analytics firm RiskGrid Technologies which provides quantitative risk management solutions to asset management firms and wealth managers. Prior to RiskGrid Raj was a founding member and director of financial engineering at Algorithmics Inc which was acquired by IBM 2011.  Raj holds an MSc in computer science from University of Toronto where he did research in machine learning and pattern recognition co-supervised by Geoffrey Hinton.

Talk: TBD

TBD

You have Successfully Subscribed!

Darin Graham

Director, R&D Strategy & Operations, Toronto AI Lab at LG

Short Bio

Talk: TBD

TBD

You have Successfully Subscribed!

Daniel Weimer

Head of AI, Group Volkswagen of America

Short Bio

Talk: TBD

TBD

You have Successfully Subscribed!

Ofer Shai

Chief AI Officer, Deloitte Canada

Ofer is Chief AI Officer in Omnia AI, Deloitte’s AI practice, where he is leading Deloitte’s AI and ML initiatives through Data Science and the AI Factory. Ofer uses the latest in deep learning, machine learning, and AI to build products and solutions for Deloitte’s clients, and help our clients define and implement their AI strategy.

Ofer has over 15 years of experience in natural language processing, predictive and advanced analytics, recommendation systems, information retrieval, and computational biology. Ofer holds a Ph.D. from the University of Toronto Machine Learning Group and has held leadership positions at Upsight, Meta, and Chan-Zuckerberg Initiative

Talk: TBD

TBD

You have Successfully Subscribed!

Jaya Kawale

Senior Research Scientist, Netflix

 

Short Bio

TBD

TBD

You have Successfully Subscribed!

Jian Chan

Staff Algorithm Expert, Alibaba Group

 

Short Bio

Talk: TBD

TBD

You have Successfully Subscribed!

Trishala Pillai

Partnerships Lead,Myplanet

Trishala is Partnerships Lead of Voice + AI Channel Partners @ Myplanet, where she is working with big and emerging technology companies to help the world's most influential organizations take a business problem and understand where AI (specifically conversational user interfaces and machine learning models) can be useful. Outside of her work with Myplanet, she is also a member of St. Paul's Greenhouse at the University of Waterloo's Advisory Board (a nationally recognized social innovation discovery lab) and Partnerships Chair of B Lab's Certified Benefit Corporation movement for East-Central Canada.

Talk: TBD

TBD

You have Successfully Subscribed!

Chatterjee Chanchal

AI Leader, Google Cloud

Chanchal Chatterjee, Ph.D., held several leadership roles in machine learning, deep learning, and real-time analytics. He is currently leading Machine Learning and Artificial Intelligence at Google Cloud Platform with a focus on Financial Services. Previously, he was the Chief Architect of EMC CTO Office where he led end-to-end deep learning and machine learning solutions for data centers, smart buildings, and smart manufacturing for leading customers. He was instrumental in Industrial Internet Consortium, where he published an AI framework for large enterprises. Chanchal received several awards including an Outstanding paper award from IEEE Neural Network Council for adaptive learning algorithms recommended by MIT professor Marvin Minsky. Chanchal founded two tech startups between 2008-2013. Chanchal has 29 granted or pending patents, and over 30 publications. Chanchal received M.S. and Ph.D. degrees in Electrical and Computer Engineering from Purdue University.

Talk: Google’s Journey to AI-First. “Hey Google, What is AI First?”

Describe Google's journey from data to mobile to AI company. Discuss how AI is done at Google. Some advanced techniques of AI will be discussed followed by industry applications and finally a vision of AI in the future.

You have Successfully Subscribed!

Niki Athanasiadou

Data scientist, H2O.ai

Short Bio

Talk: TBD

TBD

You have Successfully Subscribed!

What You'll Learn

 

Hundreds of petabytes of time series data are generated each day in many enterprises. It’s challenging to query this rapidly growing data in a timely manner. TSDB is short for "time series database", which can be used the backbone service for hosting all this data to enable high-concurrency storage and low-latency query.

An AI engine on TSDB provides intelligent advanced analysis capabilities and end-to-end business intelligence solutions and empowers companies across various industries to better understand data trends, discover anomalies, manage risks, and boost efficiency. We the design of the AI engine to enable fast and complex analytics of large-scale time series data in many business domains.

Along the way, they highlight solutions to the major technical challenges in data storage, processing, feature engineering, and machine learning algorithm design.

 

You have Successfully Subscribed!

What You'll Learn

The vast majority of successful deep neural networks are trained using variants of stochastic gradient descent (SGD) algorithms. Recent attempts to improve SGD can be broadly categorized into two approaches: (1) adaptive learning rate schemes, such as AdaGrad and Adam, and (2) accelerated schemes, such as heavy-ball and Nesterov momentum.

In this paper, we propose a new optimization algorithm, Lookahead, that is orthogonal to these previous approaches and iteratively updates two sets of weights. Intuitively, the algorithm chooses a search direction by looking ahead at the sequence of "fast weights" generated by another optimizer. I will discuss how neural network algorithms can be analyzed and show that Lookahead improves the learning stability and lowers the variance of its inner optimizer with negligible computation and memory cost.

I will then present empirical results demonstrating Lookahead can significantly improve the performance of SGD and Adam, even with their default hyperparameter settings on ImageNet, CIFAR-10/100, neural machine translation, and Penn Treebank.

You have Successfully Subscribed!

What You'll Learn

Companies are increasing investing in infrastructure and integration projects to amass large-scale datasets for data projects. Against a back-drop of emerging global privacy regulation that seemingly seeks to reign in use of those datasets (e.g. impending changes to the federal privacy legislation, PIPEDA; the soon to be enacted California Consumer Protection Act, etc.), the returns from such investment may seem uncertain. However, those returns can be maximized through adopting a sound data governance strategy.

First, by defining clear principles for data validation and protection and a process for adhering to the same, a company can minimize regulatory and reputational risk. Second, by articulating its data vision, a company can ensure that it has the necessary rights to the datasets to maximize use cases and the related business opportunities. This talk will also describe how companies are effectively operationalizing data governance.

You have Successfully Subscribed!

What You'll Learn

 

As data processing and storage is becoming cheaper, the main barrier to entry for AI adoption is often data availability. This couldn’t be better exemplified than in medicine, where advancements in data capturing and storage are creating the necessary conditions for efficient AI support systems.

 

AI-enabled clinical decision support includes diagnosis and prognosis, and involves classification or regression algorithms that can predict the probability of a medical outcome or the risk for a certain disease. Several image classification algorithms using medical images have been approved by the FDA as diagnostic tools in the last two years, and more are certain to follow.

 

Similarly, FDA approval has already been given to wearable devices that monitor vital signs to capture irregularities.  These early examples demonstrate the huge potential of AI applications in medicine, as the volume and variety of medical data that get captured increases.

 

More than 80-90% of US hospitals and physician offices are implementing some form of an EHR, and similar or even higher adoption rates are seen globally.  

 

Despite persistent outstanding issues, the lack of interoperability between EHR systems or patient history continuity, past barriers to adoption relating to data usability and availability are being overcome. Three examples of clinical decision support AI models built on EHR data will be discussed. (1) Accumulation of medical histories from birth alongside linked maternal EHR information in a healthcare facility, enabled the  prediction of high obesity risk children as early as two years after birth, possibly allowing life-altering preventative interventions. (2) The Advanced Alert Monitoring system developed and deployed by Kaiser Permanente uses Intensive Care Unit (ICU) data to predict fatally deteriorating cases and alert staff to the need of life-saving interventions. (3) Last, but not least, clinical decision support systems are often required to provide sufficient explanations of their predictions.

 

Global and local explanations of predictions regarding hospital readmissions demonstrate how interpretability techniques enable such explanations. As EHR information becomes standardized and enriched with eg. genomic information, medicine is poised to leverage AI breakthroughs to improve health outcomes. 

You have Successfully Subscribed!

What You'll Learn

 

Much real-world data is sampled at irregular intervals, but most time series models require regularly-sampled data. Continuous-time latent variables models can handle address this problem, but until now only deterministic models, such as latent ODEs, were efficiently trainable by backprop.

We generalize the adjoint sensitivities method to SDEs, constructing an SDE that runs backwards in time and computes all necessary gradients, along with a general algorithm that allows SDEs to be trained by backpropgation with constant memory cost.

We also give an efficient algorithm for gradient-based stochastic variational inference in function space, all with the use of adaptive black-box SDE solvers. Finally, we'll show initial results of applying latent SDEs to time series data, and discuss prototypes of infinitely-deep Bayesian neural networks.

 

You have Successfully Subscribed!

What You'll Learn

Is AI trustworthy? It can be with the right validation tools. As AI continues to transform our capabilities, it becomes imperative to validate that our models are fair and ethical.

The speakers will show work done at RBC on developing model validation at scale through self-serve capabilities which allows them to reduce validation time, increasing model trust, and establishing guarantees on specific properties of models. 

In this presentation, attendees will learn how to develop ethical model validation, and integrate model validation and testing into the data science pipeline. They will introduce the key packages including fairness, model performance, model interpretation, auto benchmark building and data validation. 

You have Successfully Subscribed!

What You'll Learn

 

Modern AI systems are increasingly capable of tackling real-world problems. Yet the black box nature of some AI systems, giving results without a reason, is hindering the mass adoption of AI. According to an annual survey by PwC, the vast majority (82%) of CEOs agree that for AI-based decisions to be trusted, they must be explainable. As AI becomes an ever more integral part of our modern world, we need to understand why and how it makes predictions and decisions. These questions of why and how are the subject of the field of Explainable AI, or XAI. Like AI itself, XAI isn’t a new domain of research, and recent advances in the theory and applications of AI have put new urgency behind efforts to explain it.

 

In this talk the speaker will present a technical overview of XAI. The presentation will cover the there key questions of XAI: “What is it?”, “Why is it important?”, and “How can it be achieved?”.

 

The what of XAI part takes a deep dive into what it really means to explain AI models in terms of existing definitions, the importance of explanation users’ roles and given application, possible tradeoffs, and explanation studies beyond the AI community. 

 

In the why of XAI part, we explore some of the most important drivers of XAI research such as establishing trust, regulatory compliance, detecting bias, AI model generalization and debug.

Finally, in the how of XAI part we discuss how explainability principles can be applied before, during, and after the modelling stage of AI solution development. In particular, attendees will be introduced to a novel taxonomy of post-modelling explainability methods, which are then leveraged to explore the vast XAI literature work.

 

Who is this presentation for: The first two parts of the talk (the What and Why of XAI) are targeted at a broader audience who are not AI experts but are somewhat familiar with AI.
The last part (the How of XAI) part is intended for AI experts and practitioners who would like to learn about applying XAI in their work.

 

What you’ll learn: The content presented in this talk is unique in two ways. The breadth of XAI literature that is covered is quite vast and, yet, it is highly structured to make the material easier to digest. More importantly, it is intended to provide insights to audience with both technical and business background.

 

 

 

You have Successfully Subscribed!

What You'll Learn

 

Data scientists spend big chunk of their time preparing, cleaning, and transforming raw data before getting the chance to feed this data to their well-crafted models. Despite the efforts to build robust predication and classification models, data errors still the main reason for having low quality results. This massive labor-intensive exercises to clean data remain the main impediment to automatic end-to-end AI pipeline for data science.

In this talk, the speaker will focus on data prep and cleaning as an inference problem, which can be automated by leveraging modern abstractions in ML. The speaker will describe the HoloClean framework, a scalable prediction engine for structured data. The framework has multiple successful proof of concepts with cleaning census data, market research data, and insurance records. The pilots with multiple commercial enterprises showed a significant boost to the quality of source (training) data before feeding them to downstream analytics.

HoloClean builds two main probabilistic models: a data generation model (describing how data was intended to look like); and a realization model (describing how errors might be introduced to the intended clean data). The framework uses few-shot learning, data augmentation, and self supervision to learn the parameters of these models, and use them to predict both error and their possible repairs.

You have Successfully Subscribed!

What You'll Learn

 

The speaker will describe Google's journey from data to mobile to AI company.

 

Discussing will be focused on how AI is done at Google. Some advanced techniques of AI will be discussed followed by industry applications and finally a vision of AI in the future.

You have Successfully Subscribed!

What You'll Learn

 

Some of the libraries very commonly taught and used in data science have not been designed for large scale machine learning so scaling up computation can be a challenge, particularly that many courses tend to focus on the algorithms and do not cover ML engineering. On the positive side, there are many ways to address this today and choosing the right one for a given project is an important decision as changing architectures can be expensive.

The talk will go through the pros and cons of several approaches to scale up machine learning, including very recent developments.

What you’ll learn: The audience will learn what criteria to use to choose the appropriate approach for their case, as well as the practical pros & cons of each.

 

What languages will be discussed: Python and Java. The speaker will discuss multiple infrastructure options

 

You have Successfully Subscribed!

What You'll Learn

 

In his talk, the speaker will cover image augmentations that are used for object detection and semantic segmentation tasks. They will also talk about novel types of transforms that allow achieving state of the art results in research and in deep learning competitions. They will also discuss their applications for different domains such as self-driving, satellite and medical imagery.

You have Successfully Subscribed!

What You'll Learn

Choosing which AI projects to invest in can be a daunting task. How do Layer6’s capabilities fit within TD bank? How does ML research get rolled into products at National Bank of Canada? How is Hudson’s Bay Company selecting the AI projects to build out? How has advanced research from Toronto’s Machine Learning Group help clients at Deloitte?

You have Successfully Subscribed!

What You'll Learn

 

 

TBA

You have Successfully Subscribed!

What You'll Learn

 

 

TBA

You have Successfully Subscribed!

What You'll Learn

 

 

TBA

You have Successfully Subscribed!

What You'll Learn

This talk will introduce differential privacy and its use cases, discuss the new component of the TensorFlow Privacy library, and offer real-world scenarios for how to apply the tools.

In recent years, the world has become increasingly data-driven and individuals and organizations have developed a stronger awareness and concern for the privacy of their sensitive data. It has been shown that it is impossible to disclose statistical results about a private database without revealing some information. In fact, the entire database could be recovered from a few query results.

Following research on the privacy of sensitive databases, a number of big players such as Google, Apple, and Uber have turned to differential privacy to help guarantee the privacy of sensitive data. That attention from major technology firms has helped bring differential privacy out of research labs and into the realm of software engineering and product development. Differential privacy is now something that smaller firms and software startups are adopting and finding great value in.

Apart from privacy guarantees, advances in differential privacy also allow businesses to unlock more capabilities and increased data utility. One of these capabilities includes the ability to transfer knowledge from existing data through differentially private ensemble models without data privacy concerns. As differential privacy garners recognition in large tech companies, efforts to make current state-of-the-art research more accessible to the general public and small startups are underway.

As a contribution to the broader community, Georgian Partners has provided its differential privacy library to the TensorFlow community. Together, we will make differentially private stochastic gradient descent available in a user-friendly and easy-to-use API that allows users to train private logistic regression

What you’ll learn: This talk will introduce differential privacy and its use cases, discuss the new component of the TensorFlow Privacy library, and offer real-world scenarios for how to apply the tools.

You have Successfully Subscribed!

What You'll Learn

 

Concepts in videos are high level labels given to segments within videos. They can be actions such as "skateboarding", "dancing" or more general entities such as "acoustic guitar", "wedding". This presentation will be an overview of concept recognition and localization task with respect to YouTube-8M dataset. It will cover methods in recent literature and our approach to reach state-of-the-art results on the dataset.

You have Successfully Subscribed!

What You'll Learn

AI technology is redefining almost every industry by enabling transformation of established business models and products. Still, a main challenge is to transfer research findings into actual AI products. This talk will give an overview on current AI challenges within automotive and what is necessary to make a company AI-ready.

You have Successfully Subscribed!

What You'll Learn

 

TBA

You have Successfully Subscribed!

What You'll Learn

In the past decade, computer systems and chips have played a key role in the success of AI. Our vision in Google Brain's ML for Systems team is to use AI  to transform the way systems and chips are designed. Many core  problems in systems and hardware design are combinatorial optimization or decision making tasks with state and actions sizes that are orders of magnitude larger than common AI benchmarks in robotics and games. In this talk, I will go over some of our research on tackling such optimization problems.

First, I talk about our work on deep reinforcement learning models that learn to do resource allocation, a combinatorial optimization problem that repeatedly appears in systems. Our method is end-to-end and abstracts away the complexity of the underlying optimization space; the RL agent learns the implicit tradeoffs between computation and communication of the underlying resources and optimizes the allocation using only the true reward function (e.g., the runtime of the generated allocation).

I will then discuss some of our recent work on deep reinforcement learning methods for sequential decision making tasks with long horizons and large action spaces, built upon imitation learning and tree search in continuous action spaces. Finally, I discuss our work on deep models that learn to find solutions for the classic problem of balanced graph partitioning with minimum edge cuts. We define an unsupervised loss function and use neural graph representations to adaptively learn partitions based on the graph topology. Our method enables the first generalized partitioner, meaning we can train models that produce performant partitions at inference time on new unseen graphs. 

You have Successfully Subscribed!

What You'll Learn

 

TBA

You have Successfully Subscribed!

What You'll Learn

 

The Data Science group at The New York Times develops and deploys machine learning solutions to newsroom and business problems. Re-framing real-world questions as machine learning tasks require not only adapting and extending models and algorithms to new or special cases but also sufficient breadth to know the right method for the right challenge.

The speaker will first outline how unsupervised, supervised, and reinforcement learning methods are increasingly used in human applications for description, prediction, and prescription, respectively. The speaker will then focus on the 'prescriptive' cases, showing how methods from the reinforcement learning and causal inference literatures can be of direct impact in engineering, business, and decision-making more generally.

 

You have Successfully Subscribed!

What You'll Learn

 

TBA

You have Successfully Subscribed!

What You'll Learn

 

TBA

You have Successfully Subscribed!

What You'll Learn

Traditional software development has a roadmap—the Software Development Life Cycle, coalesced around a specific set of tools and processes. In contrast, machine learning development is a tangle of tools, languages, and infrastructures, with almost no standardization at any point in the process. Manual stopgaps and one-off integrations get models into production, but introduce fragility and risk that prevents businesses from trusting them with mission-critical applications. To build and deploy enterprise-ready models that generate real value, businesses need to standardize on a new stack and a new, ML-focused life cycle.

 

This talk will cover:

- Key differences between ML and traditional software development

- Where the SDLC works with ML, and where it breaks down

- An overview of the new ML stack, from training to deployment to production

- The five biggest infrastructure and process mistakes ML teams commit 

- How successful early movers have succeeded, and lessons you can use today

 

Infastructures discussed: Docker, Kubernetes.

Languages discussed: Python, R

DevOps Tools Discussed: Jenkins, Github, potentially others

You have Successfully Subscribed!

What You'll Learn

 

The prevailing progress around AI has created in its wake a heightened interest in Explainable Artificial Intelligence (XAI), whose goal is to produce interpretable decisions made by machine learning algorithms. Of particular interest is the interpretation of how deep neural networks make decisions, given the complexity and 'black box' nature of such networks.

Given the infancy in the field, there has been limited exploration into the assessment of the performance of explainability methods, with most evaluations centered on subjective and visual interpretations of current approaches. In this talk, the speakers introduce two quantitative performance metrics for quantifying the performance of explainability methods on deep neural networks via a novel decision-making impact analysis: 1.) Impact Score, which assesses the percentage of critical factors with either strong confidence reduction impact or decision changing impact; and 2.) Impact Coverage, which assesses the percentage coverage of adversarially impacted factors in the input. We further consider a comprehensive analysis using this approach against numerous state-of-the-art explainability methods.

You have Successfully Subscribed!

What You'll Learn

You have Successfully Subscribed!

What You'll Learn

 

TradeRev uses regression models for predicting the auction price of cars. The early years of ML/development focused entirely on time to market which lead to a successful product but we ended up with a code base that had huge tech debt (spaghetti code, monolithic architecture, manually created infrastructure etc.). Increasing adoption rate of the product exposed the tech debt as scaling the product became a massive bottleneck.

The speaker will discuss how they took the challenge of rearchitecting the entire ML product from both software engineering and data science perspectives.

They will share how they accomplished many milestones as a result of this endeavour:

 

- Improved model accuracy

- Microservices architecture

- Scalable ML solution

- Continuous Integration & Automated Deployments

- Dockerized software solution- 80% + code coverage

- Regression/performance testing

- Enhanced monitoring of evaluation metrics

- Infrastructure as service 

What you’ll learn: Attendees will learn tips, and techniques to embark on rearchitecting a legacy ML system from inception to production.

You have Successfully Subscribed!

What You'll Learn

 

A critical component of data management and enrichment pipelines is connecting large datasets from various sources to form a holistic view; to make connections between entities across data sources. Oftentimes, these entities – such as individuals, organizations, or addresses – may not have a unique identifier that can be used as a key to detect duplicates or to merge datasets on. ThinkData has developed a scalable entity resolution engine to solve these problems.

After experimenting with both deep learning and traditional NLP techniques, the team has found the best balance of accuracy and performance. Specifically, we have achieved near-parity in accuracy compared to Magellan (the leading entity resolution project in research), albeit with much better performance metrics and greater scalability. This talk will discuss the importance of entity resolution, our approach to solving real-world challenges, and the potential in using entity resolution and graph relationships in tandem.

You have Successfully Subscribed!

What You'll Learn

 

TBA

You have Successfully Subscribed!

What You'll Learn

Deep learning models perform poorly on tasks that require commonsense reasoning, which often necessitates some form of world-knowledge or reasoning over information not immediately present in the input.

We collect human explanations for commonsense reasoning in the form of natural language sequences and highlighted annotations in a new dataset called Common Sense Explanations (CoS-E). We use CoS-E to train language models to automatically generate explanations that can be used during training and inference in a novel Commonsense Auto-Generated Explanation (CAGE) framework. CAGE improves the state-of-the-art by 10% on the challenging CommonsenseQA task. We further study commonsense reasoning in DNNs using both human and auto-generated explanations including transfer to out-of-domain tasks. Empirical results indicate that we can effectively leverage language models for commonsense reasoning.

 

What you’ll learn: 

- Human explanations used only during training improves performance on downstream tasks 

- Explanations are a way to incorporate commonsense in neural networks 

- Language Models are powerful enough to generate meaningful commonsense explanations 

- Auto-generated explanation improves accuracy by 10% points on Commonsense Question Answering.

You have Successfully Subscribed!

What You'll Learn

Deep Reinforcement Learning has seen a lot of breakthroughs in the news, from game playing like Go, Atari and Dota to self-driving cars, but applying it to millions of people in production poses a lot of challenges. The promise of Reinforcement Learning is automated user experience optimization. As one of the world’s largest mobile video game companies, Zynga needs automation in order to personalize game experiences for our 70 million monthly active users. This talk discusses how RL can solve many business problems, the challenges of using RL in production and how Zynga’s ML Engineering team overcame those challenges with our Personalization Pipeline.

 

Infrastructure discussed: Spark, TensorFlow, TensorFlow-Agents

 

What you’ll learn: How Reinforcement Learning can be applied to many business problems, the challenges of dealing with RL in production over traditional supervised models, Zynga's solutions for those challenges

You have Successfully Subscribed!

Rupinder Dhillon

Chief Data Officer, SVP Data & AI, Hudson's Bay Company

Short Bio

Talk: TBD

TBD

You have Successfully Subscribed!