By Erik Gfesser — Jun 4, 2019

Amazon re:MARS 2019 Day 1 (June 4, 2019)

Amazon re:MARS 2019
Aria Resort & Casino
Day 1 (June 4, 2019)

From the promotional materials:

KEY1: Opening remarks & Innovation Spotlights (5:30pm – 6:30pm)

Opening remarks by Dave Limp, SVP Amazon Devices and Services. Keynote to include innovation spotlights talks by (1) Marc Raibert, CEO, Boston Dynamics, (2) Morgan Pope, Associate Research Scientist, Walt Disney Imagineering, (3) Tony Dohi, Principal R&D Imagineer, Walt Disney Imagineering, and (4) special celebrity talk by Robert Downey Jr, Producer and Actor.

My personal notes:

Dave Limp, SVP Amazon Devices and Services opened up
there will be 3 innovation spotlight talks
the original MARS (machine learning, automation, robotics, space) conference started about 4 years ago
trying to get builders and dreamers together
the gap that has existed between builders and dreamers is decreasing over time
with respect to MARS, machine learning came first and is the foundation to everything
automation done correctly is invisible to the customer
the goal of this conference is inspiration
the party on Thursday will include rides around a track in pace cars, and battle bots
Marc Raibert, CEO, Boston Dynamics presented the first spotlight talk
gave interesting talk addressing the type of intelligence being addressed
what kind of intelligence? the athlete vs the scholar
Handle, Spot, and Atlas are 3 robots currently being worked on at Boston Dynamics
Handle is purpose built, Spot is general purpose, and Atlas is R&D
talked through some video playback of Handle moving boxes between pallets in laboratory warehouse setting
then talked through video playback of Spot walking and climbing throughout the warehouse
it appeared as though Spot was tethered whereas it seemed like Handle was freely moving about
the first commercial product coming out in "a couple months" will be a version of Spot
it sounded like the purpose of this product will be to monitor construction sites, apparently for compliance
Marc then gave a demo of Spot physically on stage, and it was apparently let loose on the floor afterward (see my notes about this during the welcome reception below)
the least time was spent discussing Atlas, he just made a few comments
however, during this discussion he talked about the typical public perception of robots
essentially, public perception is built around the concept of physical design combined with software
however, Marc explained that the physical environment should be considered a third aspect
video playback of Atlas moving boxes between tables was then shown, and Atlas also appeared to be tethered
Morgan Pope, Associate Research Scientist, and Tony Dohi, Principal R&D Imagineer, both from Walt Disney Imagineering, jointly presented the second spotlight talk
Morgan and Tony opened up by discussing the Abraham Lincoln automaton / animatronic from the 1964 New York World's Fair
the original idea was to bring the 2-d world that Walt Disney was creating into the physical world
I recall watching the Lincoln automaton several years ago at Disneyland, and as they were talking about it I thought perhaps it may have influenced the Philip K. Dick short story "We Can Build You", which I read when I was younger during my blitz of reading everything Philip K. Dick, but looking it up it appears it was originally written in 1962
later, about 20 years ago, the desire was to untether such automatons from stationary positions like the 1964 Lincoln
they then talked through video playback of various incarnations evolving and being tested to do so, with the goal throughout this process to fail fast
the original such mechanism resembled a raccoon, and they figured they needed to start making mechanisms in human form
the next step was a human sized acrobat after they determined they could scale from earlier successes
some level of decision making was then introduced to the acrobats in terms of when to drop etc
when Tony was getting his PhD from Stanford, they would always say "that's really cool, but what's the business case?", but at Disney, "super cool is the business case"
Robert Downey Jr then gave a monologue, and at this point I was no longer taking notes since I just wanted to sit back
while his monologue was rather entertaining, Robert rambled quite a bit about dozens of topics for the duration of his time on stage
the 2 most entertaining sequences were his periodic, presumably simulated interactions with Alexa, and his dialogue with Matt Damon
at one point, Alexa abruptly interjected while Robert was talking, and he commented that he thought Alex was supposed to respond to inquiries, not lead
it seemed as though Amazon may have been hinting that the role of Alexa may be changing down the road
as much rambling as Robert may have done, I was pretty impressed with the extent to which he was able to memorize so many lines, including the back and forth with Alexa, and the back and forth with Matt Damon, which was via video that may have been prerecorded but seemed live
the other thing I was impressed with was his ability to joke about his lack of education, and the years he wasn't doing so well as an actor due to personal challenges
he also announced an initiative called "The Footprint Coalition" that he will officially launch in April 2020, the goal of which is to clean up the planet from an environment perspective, which can probably be done over the time span of only a decade according to some of whom with whom he has been meeting

From the promotional materials:

GEN1: Welcome Reception (6:30pm – 8:30pm)

We look forward to welcoming you to the event at the opening reception, which will take place in the Tech Showcase. This is a great opportunity to meet other attendees and familiarize yourself with what exhibits you may want to revisit later in the week.

My personal notes:

not too many notes from me for this one, as I spent too much time watching the creations from Boston Dynamics walk about the exhibit hall
I didn't realize until tonight that these are controlled by remote, at least for this conference, presumably for control
it so happened by coincidence that a media interview with Marc Raibert started literally next to me
as I was watching one of the other exhibits, one of the founders of acloud.guru happened to stop in front of me, and I immediately recognized him due to my working through one of his courses to pass my first AWS cert earlier this year

From the promotional materials:

W08: Practical Machine Learning, From Unlabeled Data To A Production Model (9:00am – 12:00pm)

Attendees will learn how to do end-to-end Machine Learning starting with an unlabeled dataset of images (I-Naturalist or similar), and use Amazon SageMaker to: explore and prepare the dataset, employ SageMaker Ground Truth to inexpensively label the dataset with object bounding boxes, train an object detection model (single shot multibox detector or SSD) based on the labeled dataset, deploy the model to an endpoint and evaluate the model performance, learn about hyperparameter optimization, and use HPO to find a more accurate model for deployment.

My personal notes:

this workshop was led by Denis V. Batalov and Vikram Madan
from what I understood, Denis leads ML as well as the specialist solution architects at AWS, and Vikram works on SageMaker Ground Truth
Sagemaker is about 1.5 years old and was launched at re:Invent 2017
we largely followed the guide located at the following location: s3-us-west-2.amazonaws.com/sagemaker-remars/guide.zip
the first workshop step was to set up labeling with Ground Truth
prebuilt labeling workflow "bounding boxes" was used rather than other options such as image classification, text classification, semantic augmentation. or custom workflow
Vikram commented at one point that "there always needs to be human input, since you're ground truthing it"
needed to work in us-west-2 Oregon since that's where the dataset is located
in following the instructions, I missed the step that suggested "providing some helper text to guide the workers for both good and bad examples"
bounded boxes essentially are boxes that are drawn around the object of interest in each image, and for this workshop, the object of interest was bees
good bounding box examples include all visible parts of the insect: legs, antennae etc, and bad bounding box examples are when the bounding box is too big and/or excludes some visible parts of the insects
I talked to one of the AWS assistants at one point to ask about my not noticing the note in the Jupyter notebook to provide this helper text for workers, as it seemed to be taking a while for them to complete the tasks I had set up, but I wanted to see what would happen
the sample size of 500 images we used here was 2%, resulting in 10 images for which workers needed to created bounded boxes
he told me that 30-45 minutes should be expected for the 10 image sample, and his response to my comment that this seemed to be a long time was that you can pay more to get the work done more quickly
I'm not sure he followed everything I was saying, as I asked him how accountable the workers are in terms of accuracy of their work: are random samples perhaps taken and assessed to determine label accuracy? I asked twice, and he just kept saying that they are under contract with Amazon
Vikram commented about the crowdsourced global workforce available via the AWS marketplace: since it's an art, the price should be set higher if the task is more difficult / takes more time to execute
Vikram commented at some point that "I'm not a Mac person, can someone give me a hand?", and our table of developers was joking about this, as all of this work was just being performed through a browser anyway
Vikram also said that "there's variability in terms of how long it will take the task to complete"
Denis later said that "hyperparameters are the interesting bit": the weights, essentially metaconfiguration
my labeling job never finished, and got stuck on 9 out of 10 images complete
it's possible that my job would eventually complete, but most workshop attendees were waiting for workers to complete this task, and I needed to start cleaning up AWS resources to prevent further charges (we were all issued $50 credit cards as reimbursement for charges we incurred during the workshop)
the demo by Denis didn't finish either
Denis asked if anyone was having any issues, and a couple attendees said that only a certain instance type was supported, to which Denis responded that a limit increase request would need to be submitted to AWS, which was interesting because if this were the case we should have received a limit increase prior to the workshop
I was fortunate to have a woman at my table, as I did some quick math that considered about 21 tables with 8 people each, totaling 168, rounding down to 150 since some chairs were empty, and seeing only about 5 women total in attendance
unfortunately, very little here that I could have just learned by following some AWS tutorials, and the pace of this workshop was extremely slow, with the presenters providing very little value add
in addition, I was disappointed that this workshop did not end up being the full day that was originally advertised when I registered for the conference, as conference organizers apparently ended up deciding to instead offer 2 sessions, each only a half a day in length, after they realized how popular it was to attendees
as with other conferences, however, the great aspect was meeting people at my table who traveled from Denver, Atlanta, Dallas, and Tokyo to attend the conference, and I talked at length with the gentleman from Denver who is a university professor there who teaches statistics and also does consulting

Subscribe to Erik on Software