Artificial Intelligence Conference 2019 – New York, New York

About this course

so we wanted to take a few minutes to go over the results of a recent survey as well let's talk about some of the things we're seeing in the AI space so first

off as many of you know last year we gave talks where we excited the number of papers being published in machine learning so for example there's a website called archive and there's about

a hundred papers uploaded there every day but all of that research is also being translated into patent filings so I was fortunate enough to be part of a recent study from the World Intellectual

Patent Office and as you can see the number of patent filings is also growing fast and this basically reinforces the notion that we are entering the implementation phase for AI so notably

computer vision is mentioned in about half of all AI related patents and the number of patent filings in computer vision is growing fast 24 percent since 2011 so in this conference we have many

many sessions on computer vision over the next few days now it's been a extremely productive year for researchers and natural language in particular it feels like every few

months there are new deep learning models are establishing records and many different natural language tasks and benchmarks and much of this research was done in the open fortunately accompanied

by open source code and pre-trained models which is always great for the communities we already are already seeing companies were going to tune these models for the specific domains

and applications there are many sessions on text speech and natural language at this conference including a talk tomorrow by shaming way one of the creators of Bert now we continue to see

improvements and tools for deep learning as well to our survey tional that tentacle implies words remain in most popular libraries but we're also seeing tools from companies like weights and

biases and determined AI designed for companies with growing teams of deep learning engineers and data scientists these tools help optimize compute resources automate various stages of

model building and help users tract and manage their experiments so in our recent survey which drew over 1300 respondents we found about a fifth of the people signaled that they're

beginning to use reinforcement learning so one of the reasons is that we have many more tools so open source proprietary and cloud services for doing RL so over the next few days we're gonna

have many many interesting sessions on RL including a talk by Danny Lang of unity who will describe their platform for conducting RL research as well as sanjay krishnan of the university of

chicago who will share his work on the application of RL in diverse areas including robotics and data management so let's zoom in and one of the tools that is gaining traction so we've been

offering a tutorial on distributed computing framework which is open source called ray so ray supports both stateless and stateful competent computations and fine gained

fine-grained control over scheduling this means that you can implement many many services so you can implement hyper parameter tuning streaming in reinforcement learning

so in particular the most popular library for Ray is our a Lib and our a Lib provides a unified API for different types of RL training this means that both users and researchers are beginning

to benefit from using RL Lib last year we began tracking many starts building specialized hardware for deep learning AI both for training and inference and both of these happen in the edge but

also within data centers now towards the latter part of this year we expect some of these companies to begin releasing hardware that's going to greatly accelerate a training inference while

being much more energy efficient in our recent survey we found that over 60% of companies are planning to invest at least some of their IT budget into AI but the level of investment really

depended on how much experience a company fill I already had with AI technologies and as you can see from this slide those are mature practice practices plan to invest a sizable

portion of their IT budget into AI and there's a strong likelihood as a result that there may be a widening gap between leaders and laggards in this regard so let's take a deep dive in

what's holding back adoption so according to our survey the answer depends on the level of maturity so those who are just getting started still struggle with identifying use

cases for AI technologies or in or even convincing their peers within the company that AI will be useful on the other hand those with mature practice struggle with things like lack of data

and lack of skilled people so let's look at the skill scrap more closely skills requirements really depend again as Ben mentioned on the level maturity and companies are more mature AI practices

tend to have less trouble finding use cases and they also have less need for data scientists however the need for data infrastructure engineers cut across numerous companies right and it's

important to remind ourselves that much of AI today requires large amounts of training data but in order to process that you need large amounts of computer resources to Train very large models as

the use of AI technology is grows within companies we're going to see better tools for machine learning development tools that can automate many stages of a machine learning pipeline at this

conference we have many talks on your architecture speech for including one from Amita Walker of Carnegie Mellon and determined AI we've also got several companies – about Thomas and my

offerings including data a ku Microsoft a robot who are all going to be exhibiting at the expo expo hall so companies are also realizing that ml and AI requires more than just optimizing

business and statistical metrics so we have many speakers over the next few days who will explain various techniques that are being used to develop responsible AI so there are many

important considerations so for example in our survey we found that many companies are already starting to check for fairness and bias there are also a growing number of researchers and

companies looking into machine learning models that are able to protect privacy so in this conference we have several sessions these topics including talks by Yishai

Carmel and Martin doll who who will be giving talks on privacy preserving machine learning tools specifically tools that applied to deep learning and topped by Alina Mathieu kina and Seaway

Lu who will be giving talks between the interplay of machine learning and digital forensics as that gets deployed in mission critical situations like autonomous vehicles or health medical

applications this community needs to develop tools and best practices to ensure safety we have leading researchers like Joanna Bryson of the University of Bath and Alexander

Marjorie of MIT here at the conference to talk about that topic now many AI systems will continue to interact with or even depend on humans and that's why design and transparents

are also going to be important considerations for those who build AI applications these are topics where we have many sessions over the next two days including talks by furrow pour

sabzi of Microsoft Research near my C and Andrews Zaldivar of Google one last word about data security in the age of AI there are many situations where data integrity will be as critical as did a

confidentiality now remember that many add systems are highly depend on data used for training and so building data infrastructure that can keep track of data governments and lineage will be

difficult but important not only for security and quality assurance audits but also for compliance with existing and future regulations the good news is that companies are beginning to build

these foundational tools for data infrastructure that will be critical to success for AI you