Abstract: Recently, advances in computer vision have been dominated by deep learning models and con- volutional neural networks (CNN's). The well-known problem of image classification has been more or less solved by CNN's such as ResNet, and although challenges remain, the blueprint for training models that surpass human performance is well-defined.

Of greater interest is the problem of localization; finding the location of an object within an im- age in addition to determining what that object is. Models that are capable of performing this task are useful in a variety of scenarios, from object detection for self-driving cars to identifying cancerous material in CT scans. However, unlike image classification, there is no general consen- sus on the best practices for building such models, and this remains a field of active research.

In general, localization tasks fall into one of two categories: object detection and semantic seg- mentation. Object detection typically involves locating the general region or bounding box of a potential object, and then classifying said region. Semantic segmentation can be described as pixel-level classification, where the goal is to predict a pixel-by-pixel mask of each class of object.

In this workshop, we will introduce the tasks of object detection and semantic segmentation, as we as discuss how each problem is framed and represented in terms of data. Next, we will review the recent progress and advancements in both fields including: R-CNN, Fast R-CNN, Faster R- CNN, and SSD for object detection; FCN, U-Net, and E-Net for semantic segmentation. Finally, we'll take a deep dive into SageMaker, demonstrating how the platform simplifies training these models for applicable use-cases.

In this workshop, participants will learn about the following two localization techniques:

Object detection

Semantic segmentation

For each technique, participants will go through the following:

Rigorous definition of the task

An overview of recent progress / trajectory in the field

A walk-though for training a model on SageMaker

Prerequisite:This talk is geared to a level 300-400 crowd. At a minimum, participants should be comfort- able with Convolutional Neural Networks (CNN's), which assumes strong understanding of Deep Learning fundamentals. Computer vision techniques such as image processing are also desired.

Should have/create an AWS account with GPU access.

           

Abstract: In this hands-on workshop, developers will build a chatbot using Watson AI. Developers will learn how to use the Watson Assistant and Watson Tone Analyzer services. Developers will test whether the chatbot exhibit intelligent behavior (Turing test). This is a fun, hands-on session you don't want to miss!

           

Abstract:Keras is powerful and simple to use python library for Deep Learning. It integrates well with tensorflow and it is great for learning, prototyping and rapid development. In this workshop we introduce deep learning with Keras and cover fully connected, convolutional and recurrent neural networks. We apply them to a variety of common problems in industry including:

churn prediction.

image classification.

time series forecasting.

The workshop is hands on.

Instructor:Francesco Mosconi is founder and data scientist at Catalit and Data Weekends. He was formerly co-founder and Chief Data Officer at Spire, a company backed by Y-Combinator, that invented the first consumer wearable device capable of continuously tracking respiration and physical activity. Machine Learning and python expert, he also served as Data Science lead instructor at General Assembly and at The Data Incubator. In the past he worked with startups, large companies and investment groups helping them with their data needs and unlocking value from their data.

           

Abstract:Recurrent Neural Networks power deep learning systems that deal with sequences of data: weather data series, sales, sensor measurements, market data, and also natural language. They are a powerful tool you can add to your ML engineering toolbox and this session will help you do so. We will cover the basic principles behind recurrent neural networks and then apply them to various datasets. Along the way, we will share engineering best practices and practical Tensorflow tips and tricks for building and optimizing RNNs so that you can use them in your own projects. Machine learning beginners are welcome although it is recommended to follow a neural network crash course beforehand (you can start here: https://youtu.be/u4alGiomYP4, and then continue to RNNs: https://youtu.be/fTUwdXUFfI8)

Installation:Install Python3 then pip(3)-install jupyter, tensorflow and matplotlib; (on Windows, Anaconda is recommended).

           

Abstract: A deep dive into deep reinforcement learning and get hands on experience on applying deep RL to games.

Agenda:

The basics of reinfrocement learning.

Q-learning

Introduction to deep reinforcement learning

Policy gradients.

Actor-critic algorithms.

Introduction of an extensive, lightweight and flexible platform for deep RL & designing AI agents ( http://github.com/facebookresearch/ELF)

           

Keynote (grand ballroom) Computer Vision (Room 201)

Speech/NLP (Room 206)

Machine Learning/Deep Learning(Room 207)

Data Science/Analytics(Room 212)

Abstract: For the past six years, the Google Brain team (g.co/brain) has conducted research on difficult problems in artificial intelligence, on building large-scale computer systems for machine learning research, and, in collaboration with many teams at Google, on applying our research and systems to dozens of Google products. Our group has open-sourced the TensorFlow system (tensorflow.org), a widely popular system designed to easily express machine learning ideas, and to quickly train, evaluate and deploy machine learning systems. In this talk, I'll highlight some of the research and computer systems work we've done and the progress we’ve made in tackling problems in a wide variety of areas. This talk describes joint work with many people at Google..

           

Abstract: to be update

           

Abstract: to be update

           

Abstract: The research achievements in object classification, detection, segmentation are accompanied by a steep rise of Computer Vision adoption in industry. These progress open the doors to a whole new world of possibilities. As industrial applications mature, the challenges slowly shift towards challenges in limitation of memory, compute and speed. These new important metrics raise from the needs of running computer vision algorithms on mobile or embedded devices and serving large scale audience on backend services. This talk will discuss efficient neural networks with high performance model architectures.

           

Abstract: Neural networks for sequence generation have a long history, from simple recurrent neural networks (RNNs) used for language modeling, through LSTMs applied to translation to many newer models that utilize convolutions and attention mechanisms and are widely applied in natural language processing. We give an overview of these developments with focus on recent methods. We show how attention and convolutions brought training time down by orders of magnitude while allowing to model sequences with dozens of thousands of elements, in contrast to just dozens for early RNNs. These developments allow to apply autoregressive models not just to sentences but also to whole articles and images. We present recent results on generating entire Wikipedia articles and on generating images, where we outperform all current GAN architectures. We highlight a single model that works for text, audio and images and brings us closer to the goal of generating anything. We also talk about our efforts to open-source these models to make generative deep learning broadly accessible.

           

Abstract: Stitch Fix systems, composed of machines and human experts, need to recommend the maternity line when she says she’s in her ‘third trimester’, identify a medical professional when she writes that she ‘used to wear scrubs to work’, and distill ‘taking a trip’ into a Fix for vacation clothing. This talk is about how to form bridges between machine learning models and human experts by post-processing your model into something explainable (via t-SNE, the k-SVD and lda2vec) and by building models that can tell you what they *don't* know via Bayesian variational methods. Throughout, I'll share simple examples that extend word2vec, factorization machines and t-SNE to make Bayesian and interpretable versions. At the end, we'll extend t-SNE to make a hybrid "Poincaré SNE" that allows us to model hierarchical relationships for the first time

           

Abstract: Whether automating digital customer interactions using bots, predicting the best sales and marketing targets, or reducing waste in logistics and manufacturing - Artificial Intelligence improves business operations once deployed. Companies need to redefine themselves by building models and learning from data.

Companies often start with building single models, and quickly realize that this is only the first step. Making predictions available at the right time, in the right context to drive the action requires significant effort: connecting data streams, extracting features, training models and sending predictions to a front-end application. Beyond all the engineering, it means infrastructure and alerting need to be in place, along with an ability to experiment and iterate. Once companies deploy one such model, replicating this success is even more difficult. Generalizing the methods and avoiding duplicated effort is necessary if the desire is to go beyond a handful of additional models. Even so, without planning this means taking one-off approaches to painstakingly handle increased data volumes and variety, new modeling approaches, different applications, etc. Scaling to 100s becomes improbable.

At Salesforce we need to surpass hundreds of thousands. For this we built the Einstein Platform. With its automation of Artificial Intelligence and services built for handling 1000s of customers, each with multiple models. From data ingestion, automated machine learning, experimentation frameworks, and instrumentation and intelligent monitoring and alerting make it possible to serve the varied needs of many different businesses. In this talk we will cover the nuts and bolts of the system, and share how we learned to solve for scale and variability with a fully operational Machine Learning platform.

           

Abstract: State-of-the-art algorithms for applications like face recognition, object identification, and tracking utilize deep learning-based models for inference. Edge-based systems like security cameras and self-driving cars necessarily need to make use of deep learning in order to go beyond the minimum viable product. However, the core deciding factors for such edge-based systems are power, performance, and cost, as these devices possess limited bandwidth, have zero latency tolerance, and are constrained by intense privacy issues. The situation is further exacerbated by the fact that deep learning algorithms require computation of the order of teraops for a single inference at test time, translating to a few seconds per inference for some of the more complex networks. Such high latencies are not practical for edge devices, which typically need real-time response with zero latency. Additionally, deep learning solutions are extremely compute intensive, resulting in edge devices not being able to afford deep learning inference.

Deep learning is necessary to bring intelligence and autonomy to the edge. Siddha Ganju offers an overview of Deep Vision’s solution, which optimizes both the hardware and the software, and discusses the Deep Vision embedded processor, which is optimized for deep learning and computer vision and offers 50x higher performance per watt than existing embedded GPUs without sacrificing programmability. Siddha demonstrates how Deep Vision’s solutions offer better performance and higher accuracy and shares a classified secret to achieving higher accuracy with a smaller network, as well as how to optimize for information density.

           

Abstract: Data science and machine learning are critical enabling factors for data-driven organizations. There has been an exponential rise of expectations put on engineering organizations to meet the demand to develop and scale machine learning capabilities. A machine learning platform is not just the sum of it’s parts; the key is how it supports the model life-cycle end-to-end. This includes: data discovery, feature engineering, iterative model development, model training, model scoring (batch and online). The management of artifacts, their associations and deployment across various platform components is vital.

· While there are a number of mature technologies that support each phase of these life-cycle capabilities, there are limited solutions available that tie these components together into a cohesive Machine Learning platform. To support the life-cycle of a model, we need to be able to manage the various ML related artifacts, their associations and automate deployment. A life-cycle management service built for this purpose should be leveraged for storage, versioning, visualizing(including associations) and deployment of artifacts. Below are several requirements for this service: The platform should support model development in different programming languages, language and package versions should be configured specific to a model.

· Having the model environment follow the model through the life-cycle is important, for the environment should be externalized, associated and deployed together with a model.

· Various artifacts and platforms(Connection among these components is required):

Data and datasets: including source data and feature data, training datasets and scoring result-sets

Code: including notebook code, model code, deployment code

Model specific environments: should be externalized, associated and deployed with a model

Platforms: including developing and training platform, batch and online scoring platform

Hope Wang describes how her team at Intuit is managing the machine learning life-cycle, how different components associate and interact with each other, and how to execute in a production environment. Hope takes an example of how an integrated process was developed for data engineers and data scientists to manage the entire life-cycle of a model from ideation through development, training and ultimately scoring.

           

Abstract: to be updated.

           

Abstract: The road to take Machine Learning and Deep Learning applications to production and scaling them is hard and time consuming. In this session we’ll explore how Amazon SageMaker simplifies the end to end machine learning process. Building and debugging Deep Learning models has been cumbersome see how Gluon an imperative interface for Apache MXNet allows to iterate faster. In addition we’ll dive in to the architecture of Amazon SageMaker and walk through examples of SageMaker training and deployment with Apache MXNet and Gluon.

           

Abstract: I will summarize some recent work from my group at Google related to visual scene understanding and "grounded" language understanding, including:

Our DeepLab system for semantic segmentation (PAMI'17, 2nd place COCO-Stuff 2017) [1].

Our object detection system (CVPR ‘17, 1st place COCO-Detection 2016) [2].

Our person detection/pose estimation system [3] (CVPR'17, 4th place COCO-Keypoints 2017)

Visually grounded referring expressions (CVPR'16) [4].

Discriminative image captioning (CVPR'17) [5].

Optimizing semantic metrics for image captioning using RL (ICCV'17, 2nd place on COCO-captions 2017) [6]

Generative models of images and text (ICLR'18). [7]

I will explain how each of these pieces can be combined to develop systems that can better understand images and words

Reference:

[1] https://arxiv.org/abs/1606.00915

[2] https://arxiv.org/abs/1611.10012

[3] https://arxiv.org/abs/1701.01779

[4] https://arxiv.org/abs/1511.02283

[5] https://arxiv.org/abs/1701.02870

[6] https://arxiv.org/abs/1612.00370

[7] https://arxiv.org/abs/1705.10762

           

Abstract: Airbnb has had over 260 million guest arrivals over all time and is growing rapidly. Naturally, this exponential growth is a challenge to deal with from the perspective of the sheer amount of user-generated text created. Every listing contains text written by a host (listing descriptions) and text written by guests (reviews). Customer service issues come in over chat and email, and guests and hosts exchange hundreds of thousands of messages every day. This talk will present an overview of some of the work we have done in this area and will also preview a lot of the work that still needs to be done.

           

Abstract: Over the past few years, deep learning training has matured significantly, with well-known and standard solutions being applied to numerous scenarios. In contrast, deep learning inference is still largely implemented in ad-hoc ways, where existing solutions do not yet offer the required performance, efficiency, scalability and extendibility across hardware, frameworks, and usage scenarios for large-scale services. Here at Microsoft, we present the Deep Learning Inference Service, a cloud offering for ultra-fast and cost-efficient inference at scale. This distributed service handles several million production requests per second with single-digit millisecond latencies. We will go over the anatomy of the service and discuss topics such as:

Model acceleration via modern CPU features

Framework and hardware extensibility

Resource isolation in multi-tenant environments

Online/offline workload differentiation

Distributed batching and routing

Further, we will discuss how this service fits various needs of machine learning scientists, and performance requirements of site engineers. We will also dive into how it greatly boosts relevance for Microsoft Bing Search.

           

Abstract: Data products are everywhere nowadays. It seems like even the most rudimentary products can be made glamorous by simply injecting some Machine Learning magic into the mix. This trend unsurprisingly correlates with the explosion of content that we consumers have to deal with on a daily basis. The huge amount of information that submerges us causes simultaneously a problem that we need to solve (retrieving relevant information efficiently), and the opportunity to solve this problem.

At Atlassian, we take this deluge of content that our users face very seriously. Indeed, if a Confluence page or a Jira ticket that isn’t discoverable, the information that it contains is lost forever. And this is a problem that needs to be solved for every single one of our products. How does a single team supports all the Machine Learning needs of many different product teams? This is precisely the purpose of Data Science as a Service. In this talk, I will discuss how the Search and Smarts Engineering team at Atlassian is building models to serve the majority of internal customers, and talk about the challenges of building a Data Science function as a Platform team

           

Abstract: The practice of medicine involves diagnosis, treatment, and prevention of diseases. Recent technological breakthroughs have made little dent to the centuries-old system of practicing medicine: complex diagnostic decisions are still mostly dependent on “educated” work-ups of the doctors, and rely on somewhat outdated tools and incomplete data. In the meantime, AI is emerging as a truly world-changing revolution and it is significantly impacting every industry and scientific field.

With a growing research community as well as tech companies working on applying AI advances to medicine, the hope for healthcare renaissance is definitely not lost. In this talk, we will review some of the recent AI advances, with an emphasis on ML-driven medicine. We will discuss using AI for aiding medical decision including language understanding, medical knowledge base construction and diagnosis systems. We will discuss the importance of personalized medicine that takes into account not only the user, but also the context, and other metadata. We will also highlight challenges in designing ML-based medical systems that are accurate, but at the same time engaging and trustworthy for the user.

           

Abstract: Any large-scale Machine Learning (ML) application requires substantial infrastructure and systems engineering elements that are critical to its success. At Netflix, we conduct research in Recommendations and Personalization to optimize for member joy as our increasingly global membership discovers engaging content created by storytellers all over the world. To handle this challenge, our systems have scaled to the diverse personal preferences of the 117+ million Netflix members. Over the years, as the sophistication and breadth of ML algorithms and applications in the Personalization space has grown, so too has our need for robust, flexible and extensible infrastructure. In this talk, we will explore some of the design abstractions we use and lessons learned from building a paved path for ML systems in the Personalization domain. We will discuss details of a handful of generic libraries and systems we have built for various elements of ML infrastructure such as preparing training data, adhoc exploration, model development, and disciplined online deployments. We will also touch upon future directions to meet the evolving infrastructure needs of our use cases.

           

Abstract: Advanced RNN architectures are powering real-world applications such as comment moderation or language translation. This session is a deep dive into the most recent advances in the field, such as attention mechanisms, and gives you tips, engineering best practices and pointers to apply in your own projects. No PhD required.

           

Keynote (grand ballroom) Computer Vision (Room 201)

Speech/NLP (Room 206)

Machine Learning/Deep Learning(Room 207)

Data Science/Analytics(Room 212)

Abstract: Artificial Intelligence (AI) is behind practically every product experience at LinkedIn. From ranking the member’s feed to recommending new jobs, AI is used to fulfill our mission to connect the world’s professionals to make them more productive and successful. In this talk, I will provide an overview of the lessons learned and approaches we have developed to address this challenging problem, including scaling to large problem sizes, handling multiple conflicting objectives functions, and our progress toward using machine learning to optimizing the LinkedIn product ecosystem more holistically

           

Abstract: Deep Reinforcement Learning (DRL) has made strong progress in many tasks that are traditionally considered to be difficult, such as complete information games, navigation, architecture search, etc. Although the basic principle of DRL is quite simple and straightforward, to make it work often requires substantially more samples with more computational resource, compared to traditional supervised training. This task presents our recent open-sourced works: efficient, lightweight and flexible frameworks and diverse 3D environments, to facilitate DRL research. Based on these, we train agents in Real-time Strategy games and navigations, and show interesting behaviors using only a small amount of resource

           

Abstract: Over the last few years TensorFlow has enabled significant advances in deep learning research. In this talk I will discuss how the flexibility of TensorFlow makes it a great tool for research from workstations to supercomputer scale. The impact of TensorFlow and deep learning has come not just from the exciting research, but also the design of providing a library that can easily be deployed in products across a wide variety of platforms - from large scale data centers to mobile phones and edge devices. The talk will also cover some of the recent changes in TensorFlow and provide a glimpse of where it is headed

           

Abstract: to be update

           

Abstract: This morning I had the conversation with my vacuum cleaning robot" - in this session, Marek Sadowski, an entrepreneur and a developer advocate, will demonstrate the power of voice UI, mobile and AI services from the cloud platform to securely connect and manage devices, analyze social network data, and apply AI to add human-like interaction to lightweight IoT products. Marek will show how you can quickly and securely turn a simple idea into reality by transforming a regular robot vacuum into a AI-enabled smart and chatty device, that would analyze and react to pictures, and could be turned into the media information center with an opinion on virtually any subject based on the social network sentiment digesting with cognitive services.

           

Abstract: to be update

           

Abstract: GoPro’s camera, drone, mobile devices as well as web, desktop applications are generating billions of event logs. The analytics metrics and insights that inform product, engineering, and marketing team decisions need to be distributed quickly and efficiently. We need to visualize the metrics to find the trends or anomalies.

While building up the features store for machine, we need to visualize the features, Google Facets is an excellent project for visualizing features. But can we visualize larger feature dataset?

These are issues we encounter at GoPro as part of the Data Platform evolution. In this talk, we will discuss few of the progress we made at GoPro. We will talk about how to use Slack + Plot.ly to delivery analytics metrics and visualization. And we will also discuss our work to visualize large feature set using Google Facets with Apache Spark.

           

Abstract: to be update

           

Abstract: I will talk about how sequence modeling can be applied in many different areas, more specifically I will talk about speech recognition, machine translation and speech synthesis by using sequence modeling

           

Abstract: Machine Learning (ML) software differs from traditional software in the sense that outcomes are not based on a set of hand-coded rules and hence not easily predictable. The behavior of such software changes over time based on data and feedback loops. At Salesforce Einstein, we care deeply about building trust and confidence in such intelligent software programs. Why does a particular email have a higher likelihood of being opened than another? What are the shapes and patterns in the dataset, which lead to certain predictions? And can such insights be actionable?

As machine learning pervades every software vertical, and is increasingly used to automate decisions, model interpretability becomes an integral part of the ML pipeline, and can no longer be an afterthought. In the real world, the demand for being able to explain a model is rapidly gaining on model accuracy and other model evaluation metrics.

This talk will discuss the steps taken at Salesforce Einstein towards making machine learning transparent and less of a black box. We will explain how interpretability fits into the ML data pipeline, what we learned trying different approaches and how it has helped drive wider adoption of ML software.

           

Abstract: As the growth in the e-commerce sector accelerates, it has created a fertile environment for the Dark Web to develop an increasingly sophisticated and matured ecosystem for fraud. In this talk, we will expose the techniques and methods for the leading e-commerce fraud tools in the Chinese market. These tools have a variety of functionalities, including:

1. bot registration

2. bot ordering

3. captcha solving by AI and humans

4. real-time price monitoring

5. unblacklist an account

           

Abstract: Applying my Netflix experience to a real-world problem in the ML and AI world, I will demonstrate a full-featured, open-source, end-to-end TensorFlow Model Training and Deployment System using the latest advancements with TensorFlow, Kubernetes, OpenFaaS, GPUs, and PipelineAI.

In addition to training and hyper-parameter tuning, our model deployment pipeline will include continuous canary deployments of our TensorFlow Models into a live, hybrid-cloud production environment.

This is the holy grail of data science - rapid and safe experiments of ML / AI models directly in production.

Following the famous Netflix Culture that encourages "Freedom and Responsibility", I use this talk to demonstrate how Data Scientists can use PipelineAI to safely deploy their ML / AI pipelines into production using live data.

Offline, batch training and validation is for the slow and weak. Online, real-time training and validation on live production data is for the fast and strong.

Learn to be fast and strong by attending this talk!

           

Abstract: This talk offers a quick deep dive into the technical complexities of achieving personable conversation in technology. It kicks off with a high-level overview of the history of design in technology, highlighting what we have learned over the years in developing for a screen. Then, we will establish new best practices for a voice-first design using the Alexa Skills Kit, contrasting with GUI design principles. We dive into some of the complexities of accepting conversational nuances and interpreting them in the ASK SDK. In the end, we hope to offer a new perspective for technical design and implementation, as well as spark interest in joining the Alexa community to help revolutionize voice technology.

           

Abstract: Analytic applications rely on data and most of time and resources of machine learning project were spent on bringing together existing data sources and when appropriate, transform , enrich them with other data sets. I will explore the architecture and design patterns for a modern data pipeline framework and share the observations of technology assessment between Apache Nifi and StreamSets.