By Erik Gfesser — Jun 22, 2019

Amazon re:MARS 2019 Day 2-4 (June 5-7, 2019)

Amazon re:MARS 2019
Aria Resort & Casino
Day 2-4 (June 5-7, 2019)

From the promotional materials:

KEY2: Opening remarks & Innovation Spotlights (Day 2 / 9:00am – 11:30am)

Amazon keynote led by Jeff Wilke, CEO of Worldwide Consumer. Keynote to include Innovation Spotlight talks by: 1) Tom Soderstrom, IT Chief Technology and Innovation Officer, Jet Propulsion Laboratory, 2) Daphne Koller, CEO and Founder, insitro & Rajeev Motwani Professor in the Computer Science Department at Stanford University, 3) Kate Darling, Researcher, MIT Media Lab, and 4) Colin Angle, Chairman, CEO and Founder, iRobot.

My personal notes:

Jeff Wilke, CEO of Worldwide Consumer
some have been working on AI for decades, but we're only just beginning
Wilke discussed operations and how only one individual named Skip took care of it, and later operations developed algorithms to essentially do an "auto-Skip"
the first classical ML method used was item-item collaborative filtering, considered revolutionary at the time, e.g. customers who watched such and such movie watched such and such movie
graph clustering techniques came later, as did deep learning
state of the art methods ended up performing worse than simple methods
the first commercial deep learning model was simplistic but performed better and scaled better
performed 2x better than item-item collaborative filtering
emphasis on starting with customers first and then working backward
the expected customer shopping experience is much different than in the past
a product called StyleSnap, which permits providing photos that help match looks with what Amazon sells
Dilip Kumar, VP Amazon Go
there were many tech options early on in terms of how they would implement, but they ended up settling on computer vision
Kumar commented about some of the challenges they needed to surpass
one example was the constant influx of new products being available in the store, and they needed to handle this with little data
the hardest part was making the technology invisible to the customer
the shopping experience is the same except for one thing: you can just walk out
Jenny Freshwater
forecasting, buying, placement, customer promise
need to account for price elasticity, products that won't sell, and regional demand
two of the hardest challenges are new product demand and seasonal product demand
SQRF was the first algorithm they used, based on the random forest algorithm: random forest + map reduce
they started evolving their algorithm to include deep learning, leading to a 15x improvement in accuracy
in 2016, they built a feed forward neural network (FNN)
but FNN also had its limitations: modeling and training would take months each time they entered new markets etc
convolutional neural network (CNN): demand history for prior products can help predict demand for new products
unfortunately, these black box models cannot be fully explained like before
neural networks make impossible forecasting possible
Brad Porter, VP and Distinguished Engineer, Robotics
advancements in robotics
Kiva was purchased in 2012 and became Amazon Robotics
robotic palletizer: how to get the right inventory to the right placement
customer expectations keep increasing
middle model network: connects delivery network with last mile providers such as USPS
Pegasus system introduced today
due to their investment in simulation, they were able to get Pegasus ready quickly
this gave them a chance to rethink robotic drives
canvas technology will enable these robots to interact with staff beyond what they have currently been thinking
Rohit Prasad, VP and Head Scientist, Alexa Artificial Intelligence
AI pillars to improve customer experience:
(1) trust
(2) smarter via self-learning
(3) learning directly from customers
(4) proactive
(5) natural
(6) transactions to conversations
Alexa Conversations introduced today: less effort, less code, less training data
dialogue flow is now automated rather than manual
anticipating customer's latent goals
cross-skill action predictor: this will be available to customers soon
Jeff Wilke, CEO of Worldwide Consumer
Jeff came back to present
the highlight of all of these innovations is ML
democratizing AI will transform society for the better
broad acceptance of tech only comes with practical application

From the promotional materials:

M10: Infor Coleman: Streamlining Enterprise ML Complexities (Day 2 / 1:00pm – 2:00pm)

AI and machine learning can help solve some of the toughest technical problems today. In this session, you’ll learn how AWS capabilities from AI services to Amazon SageMaker can help developers build intelligent applications and systems.

Rick Rider – Sr. Director, Product Management, Infor
Massimo Capoccia – SVP Infor OS, Technology, Infor
Urvashi Chowdhary – Sr. Product Manager, AI Platforms, Amazon Web Services

My personal notes:

TODO: add notes

From the promotional materials:

M31-L: Intended Consequences: The Power of Intentional Technology Implementations (Day 2 / 2:15pm – 2:45pm)

Bringing compute power to our palms has been one of the most powerful transitions in recent history. We can be anywhere in the world and stay connected to what’s important in our lives. However, it has also brought unintended consequences, like children losing social skills by having their faces buried in their phones. Artificial intelligence has the opportunity to make an even bigger positive impact on society. Naveen Rao, neuroscientist, corporate vice president, and head of Intel’s AI efforts, will discuss how to think about AI in a way that helps bring about the positive aspects of the technology’s future, and try to avoid the pitfalls.

Naveen Rao – Corporate Vice President and General Manager of the Artificial Intelligence Products Group, Intel

My personal notes:

TODO: add notes

From the promotional materials:

A19-L: Unlocking the Creativity of Humans (Day 2 / 2:45pm – 3:15pm)

ABB Chief Digital Officer Guido Jouret believes robots can automate jobs that fit one of the 3 D's – dirty, dangerous, or dull – allowing people to focus on more creative, highly skilled tasks. In this session, he will discuss the future of job automation with robotics and the positive impacts this will have on our society and careers.

Guido Jouret – Chief Digital Officer, ABB, Inc

My personal notes:

TODO: add notes

From the promotional materials:

M23: AI from Prototype to Production: Vertical Market Value (Day 2 / 3:30pm – 4:30pm)

Intel will focus not only on AI advances driven by the power of cloud computing, but on the emergence of distributed computing architectures that leverage inference out at the source of the data in the real world, including as hospitals, cities and factory floors.

Jonathan Ballon – Vice President – Internet of Things, Intel, Intel

My personal notes:

TODO: add notes

From the promotional materials:

A26-L: The Artificial Intelligence–Powered Pivot to the Future (Day 2 / 4:45pm – 5:15pm)

AI and automation are fueling a new norm of digital disruption and breeding new forms of competition. To stay ahead, businesses must continuously innovate with speed and agility like never before. Yet, the majority of companies are challenged in embracing new AI-fueled business opportunities and calibrating their investments in innovation. Learn the secrets of “Rotation Masters” and how they are using AI to do things differently and drive double-digit growth.

John Matchette – Managing Director, Accenture Applied Intelligence, Accenture

My personal notes:

we're now able to compete on imagination, which couldn't be claimed 5 years ago
a disrupter in a marketplace is one who unlocks trapped value
if you learn the pattern, you can use disruption as a tool
what do you do if faced with disruption? do things differently and do different things
only about 6% of companies are good at being disrupters
these 6% are referred to as the "rotation masters"
Accenture came up with something called the "wise pivot"
you need to transform the core
you need to generate cash to do those new things reinvested
the next thing is that you need to grow the core, something a lot of businesses miss
you also need to scale the new
it's a new discipline, you can't move too fast or too slow
the first conference presenter I actually heard mention "munging" data
time series and human judgement are 2 ways forecasting used to be done
Matchette commented that he has a lot of past experience in supply chain
humans can probably only handle 2 ways, but these restrictions don't exist with ML
you can use disruption as a competition strategy
"data driven is hard to do"
just because you're a competitive engineering company doesn't mean you're going to be a competitive data science company
Matchette commented that much of this information is outlined at this link, which includes a recently published book by Omar Abbosh

From the promotional materials:

KEY3: Thursday Keynote & Innovation Spotlights (Day 3 / 9:00am – 11:30am)

Amazon Web Services keynote led by Dr. Werner Vogels, CTO of Amazon.com. Keynote to include Innovation Spotlights talk by: 1) Andrew Ng, Founder and CEO, Landing AI | Founder and CEO, deeplearning.ai, 2) Aicha Evans, CEO, Zoox, 3) Andrew Lo, The Charles E. and Susan T. Harris Professor, MIT Sloan School of Management | Director, MIT Laboratory for Financial Engineering, and 4) Ken Goldberg, Engineering Professor, UC Berkeley | Chief Scientist, Ambidextrous Robotics. This keynote also includes a conversation with Jeff Bezos, Founder and CEO, Amazon.

My personal notes:

TODO: add notes
Jeff Bezos, Founder and CEO, Amazon
interview by Jenny Freshwater
builders and dreamers: the dreamers come first, and the builders get inspired by them
the dreamers can dream more as the builders build
the dreamers can't continue to dream until the builders build
you needs to be right a lot: telling someone to be right a lot isn't super helpful
people who are right a lot listen a lot
they also change their mind a lot
changing minds based on data makes sense, but you can also change your mind without data
if you don't change your mind a lot, you're going to be wrong a lot
you need to disconfirm your biases
it's human to be biased
you're going to be right more often if you can disconfirm biases
Bezos would probably have been an extremely happy software engineer if he hadn't done Amazon
he seeks out people who say no, and has had trouble winning support for his ideas
disagree and commit is a leadership principal at amazon
it works in both directions, it's not just the boss who says to agree with them
you're not going to have the same ground truth as someone else
it's not a matter of being right sometimes, but a willingness to gamble with someone on something
predictions about the next 10 years are hard to make
the thing that's funny though is that nobody goes back and checks to determine accuracy
robot grasping has turned out to be extremely challenging, which was unexpected compared to vision etc
the answer to "what's not going to change" provides confidence that putting energy into them today will have staying power, stable in time not going to change
strategy should be based on customer needs that are stable over time, rather than competitors which dynamically change over time
a recently taken big bet at Amazon was Project Kuiper: Leo satellite, broadband everywhere
(a woman rushed on stage to interrupt the interview, it sounded like it had something to do with animal rights)
in being asked how he knows when to throw in the towel, Bezos said he doesn't like to give up
when the last high judgement champion has thrown in the towel, that's the time to do so
if an executive still thinks their's hope, Bezos will hang on and continue
Blue Origin is targeting the moon because it's the resource rich neighbor of Earth
water in the form of ice by poles
moon is close, only 3 days away
unlimited launch opportunities
logistically a good place to go
the reason to go to space is to save the earth
heavy industry needs to be moved off earth
operational, reusable launch vehicles is the number one important thing
infrastructure is always expensive
it was cheap to start Amazon because it relied on existing transportation networks like USPS and credit cards
the goal of Blue Origin is to build that infrastructure
you can't start a space firm from a dorm room like Facebook because the infrastructure is too expensive, the heavy lifting isn't done yet
advice to entrepreneurs for starting a business: be customer obsessed, how to absolutely delight them
you need to have passion for the arena in which you're going to work
you can't be a mercenary, you need to be a missionary
you have to be willing to take risk
if there's no risk with a particular business idea, it has probably already been done
you need to accept the fact that your business is an experiment, and it can fail

From the promotional materials:

M11: Predicting Weather to Save Energy Costs (Day 3 / 1:00pm – 2:00pm)

Learn how Kinect Energy Group uses advanced machine learning capabilities to predict electric spot prices for regional power markets using the Amazon SageMaker DeepAR time-series forecasting model, incorporating historical pricing and weather data to drive the machine learning models. Improved price predictions assist with increased trading volumes for forward pricing contracts.

Richard Delisser – Vice President, Global Infrastructure, Kinect Energy Group
Elena Ehrlich – Data Scientist, Amazon Web Services

My personal notes:

TODO: add notes

From the promotional materials:

M30-L: Investing in Technology Breakthroughs (Day 3 / 2:15pm – 2:45pm)

This session will outline managing investments in science and technology breakthroughs that challenge the status quo to bring futuristic ideas to life.

Josh Wolfe – Co-Founder and Managing Director, Lux Capital

My personal notes:

"we're in a pretty good time right now for some of these moonshots"
with respect to risk, "more things can happen than will"
where to get edge = time arbitrage + science fiction
there are about 25 VCs (venture capitalists) in attendance at this conference
not as many people doing what we're doing, but there are a lot of folks competing looking at some of the same things
it takes a behavioral disposition that's different, looking down the road
where to focus: ghosts
F. Scott Fitzgerald: "The test of a first-rate intelligence is the ability to hold two opposed ideas in the mind at the same time, and still retain the ability to function."
Mark Twain: "It ain't what you don’t know that gets you into trouble. It's what you know for sure that just ain’t so."
Schopenhauer: "Talent is hitting a target no one else can hit."
"best way to predict the future is to invent it"
reading science fiction is often cited as the source of inspiration
the gap between sci-fi and sci-fact is shrinking
"the future is here, it's just unevenly distributed"
Wolfe talked about how investments in one area lead to subsequent investments in related areas
"that's the stupidest fucking idea I've ever heard": words Wolfe has said in reaction to something that is actually going to be a good investment
what's important to think about is not when there is one thing, but when there is many, and the networking possibilities with them
find the thing that nobody knows about, and try to solve the problems they're trying to solve
dangerous words of investment: "this time it's different"
what is the unpredictable impact when everyone has one?
"Half-life of technology intimacy": you just know the way things are going to some extent: every half-life, the technology gets closer to you
the directional arrow of progress is more and more pointing to you
Wolfe explicitly invited emails from attendees

From the promotional materials:

A03-L: Follow the Money: What We Can Learn from Venture Capital Investment in Automation (Day 3 / 2:45pm – 3:15pm)

Venture capital firms are making big investments in the future of automation. Where they are putting their money is lending insight into the big trends that are expected to drive automation: mobility, artificial intelligence, drones, smart gripping, and machine vision — along with a few surprises.

Jeff Burnstein – Michigan, Association for Advancing Automation

My personal notes:

Burnstein has 35+ years experience in robotics / automation
robotics, machine vision etc initially largely failed in 1987 because of GM canceling a bunch of orders
because of this, Burnstein began to wonder whether these were actually viable
where the money is currently going:
(1) mobility
(2) gripping
(3) AI
(4) cloud
(5) drones
(6) vision
mobility is giving machines legs
a lot of money going into this
only people have been able to do this type of task in the past
a lot of gripper types out there, and not one gripper can grip everything
everything is AI these days
cloud robotics is applying the cloud to the robotics space
Burnstein thought that the drones space would have declined by now, but he thinks he was wrong
U.S. firms are now taking leadership with respect to drones
industrial drones were originally a U.S. product, but there was no interest at the time, with union opposition, so it left the U.S. and went to Asia…
now it's coming back
one firm, Veo Robotics, is working to make robots aware of humans around them, as work is typically fully manual or fully robotic because robots are not aware of the environments around them
Burnstein commented that robots have got to be about making our lives better, as it can't be limited to just competing against one another
we know that when the number of robots increases, the number of jobs increases
"robots are job creators"
there are not enough people to fill jobs right now
conference attendees have a lot of opportunities ahead

From the promotional materials:

M22: Edge AI Technology Is Redefining Smart Devices (Day 3 / 3:30pm – 4:30pm)

The ever-growing demands for privacy, short response time, and offline availability, create the trend of processing Deep Learning locally at the edge devices. Deep Learning demands high computation complexity and memory usage. Edge AI applications need the concurrent incorporation of Deep Learning, Image Signal Processing, 3D Graphics, and wireless connectivity, etc. Devices are resource-constrained, especially in memory bandwidth and thermal budget. These requirements and constrains make SoC design of AI-powered devices very challenging. In this session, we will cover the application and technology trends driving SOC solutions to make smart devices truly intelligent – now and in the future.

Ryan Chen – GM of Computing and AI Technology Group, MediaTek

My personal notes:

face unlock takes about 0.2 sec locally and voice assistant takes about 0.5 to 1.5 sec via the cloud
DNNs (deep neural networks) achieve good accuracy at the cost of high computation complexity and memory usage
cloud based solutions are preferred in the early stages of DNNs for inference
there is still network latency, even though time reduced from 2014 to 2016
accuracy improved 24% between 2012 and 2016, but complexity went up
complexity measured in terms of GMACs
cloud vs edge
more complicated AI will remain in the cloud
cloud and edge will work together in a complimentary way
there are design challenges for SoC: DNN workloads and device constraints
the most important thing is to understand the entire system, not just the AI

From the promotional materials:

M09: Accelerating Machine Learning Projects (Day 3 / 4:45pm – 5:45pm)

Cut down on development time for machine-learning projects with third-party algorithms and models in a secure environment using Amazon SageMaker. In this session, we'll talk through the sample case of a machine-learning developer who wants to use machine learning to automate the auto insurance accident claim process.

Kanchan Waikar – Senior Solutions Architect, AWS Marketplace for Machine Learning, Amazon Web Services

My personal notes:

TODO: add notes

From the promotional materials:

A08: A Faster Path to Smart Device Development with No Cloud Required (Day 4 / 9:00am – 10:00am)

The Alexa Connect Kit (ACK) provides a quick and easy path for device makers to connect devices to Alexa at a hardware level without building cloud services or complex networking and security firmware. Learn about ACK and the technical details of how one of the first devices to use ACK was prototyped in just a few weeks.

Rolando Cavazos – Division Vice President – NPD – Home and Personal Care, Spectrum Brands
Ben McInnis – Lead, Alexa Connect Kit, Amazon

My personal notes:

TODO: add notes

From the promotional materials:

M41: Extracting Complex Concepts to Unlock Value from Clinical Language Data (Day 4 / 10:15am – 11:15am)

As healthcare shifts to the value-based world, providers, payers, and other stakeholders are seeking to improve patient outcomes and financials. Language data is abundant in the world of healthcare and offers a rich source of insights about patient stories, but extracting anything more complex than the simplest concepts requires highly manual and laborious effort. In this session, Roam Analytics will share how they use AWS to enable healthcare companies to extract complex concepts from clinical language data to understand patient population trends, build predictive models, and leverage more data to inform treatment choices.

Alex Turkeltaub – Co-Founder and CEO, Roam Analytics
Kyle Johnson – Practice Manager, Data Science, Amazon Web Services

My personal notes:

patient data is largely in "language data" such as medical notes etc
most of the structured data in EHR systems is what is billed
macro trends exist in healthcare
in the past, most of this data was "scribbled"…now most of this data is digitized, but is in "horrible shape"
most of what you read in the papers about "outcome based" care is "baloney"
what is changing is how payers are assuming risk
there's a rise of paying for outcomes in a very different way
"(IBM) Watson has done very poorly in the market"
EHR systems have reduced the productivity of physicians because of their time spent typing
things are going the way of recording patient / physician interactions
most people talk about "concept extraction", i.e. does so and so suffer from such and such
the point of language data is not who has heart failure, but what conversations have taken place
"what Epic has done is criminal"
Epic is purposely built to keep hospitals away from cheaper options
and Epic is not interoperable
models in healthcare will get worse over time because the way they do things changes over time
models are created as follows:
(1) take your data and annotate it
(2) build predictive model on top of that data
(3) deploy model
there isn't a single hospital who can tell you who has heart failure!
all they can do is tell you whether they've had heart failure in the past
"physicians spend a lot of time doing really stupid things"
with the Roam system, "bureaucratic" things get sent to operations, and medical things get sent to it
"sounds prosaic, but is hugely complicated"
1/3 of the student body wanted to be physicians at Stanford when Turkeltaub was in school, and now this is down to 10%
"a lot of drugs don't work in the real world like they do in clinical trials"
paying attention to subtleties in the language data can point to why patients are not taking medication etc
you get more predictive power from the unstructured data than from the structured data because of all the information that it provides
it's important to understand that we have all this data
"what the papers like to talk about is the sexy new drug…especially ones that fail", but these are largely irrelevant
Turkeltaub presented a diagram listing the AWS services they are using, but commented that the analytics is done within the Roam platform
the number one issue mentioned at a healthcare symposium he spoke at a month ago is that there is no standard workflow…
software engineers have this, but data scientists and machine learning engineers don't
don't forget about MLU (Machine Learning University)
Turkeltaub mentioned during the Q&A that some have tried to do MDM (master data management) across different federal health agencies etc to come up with single definitions of things, e.g. because heart failure might be described in 14 different ways, but instead of making these 14 conform to 1 thing they just store all 14 and check against all of them
it's hard to build a business case around small providers for various reasons, such as a lot of small providers making use of mom and pop shop EHR systems etc…
so in order to penetrate this space there are a few options, such as providing access to Roam free of charge in exchange for some level of anonymized data in return so that Roam can feed this data into models on their end

From the promotional materials:

M42: Transforming Trash Collection: Turning Sensor Data into Machine Learning Models with Amazon SageMaker (Day 4 / 11:30am – 12:30pm)

Zolitron’s Z-Node autonomously captures data from the physical world like temperature, vibration frequencies, and GPS data. Zilitron uses Amazon SageMaker to create use-case specific machine learning models using these data streams. For example, Z-Node's vibration data is used to determine when a trash collection bin is or will be full, so that waste collection services can optimize their routes, saving time and energy. Zolitron will discuss how they clean data, and then build, train, and deploy machine learning models using AWS.

Arndt-Hendrik Zinn – Founder and CEO, Zolitron Technology
Shyam Srinivasan – Sr. Product Marketing Manager, AI/ML, AWS, Amazon Web Services

My personal notes:

TODO: add notes

Subscribe to Erik on Software