A presentation given to the University of Central Florida Dept Industrial Organizational Psychology, on Jan 12th, 2018. Covers the what why and how of AI, machine learning, Deep Learning and its coming impact to Transportation, healthcare, and use all.
3. TURN SMART
Impacts
In transportation
In health care
In cities
Fin tech
In rural
scurve
Inflexion
Market insiders
rarely predict
the expansion
duration
correctly
8. PIVIT – TURN SMART E-Mail: karl@piviting.com l +1.321.750.5165
o AI will lead to $200 billion
in new revenue
o AI will become a positive
net job motivator, creating
768k new AI jobs while
eliminating 943k jobs
Gartner
predicts
by 2020
16. TURN SMART
NN: Network types and application
Image in,
classified
out
Image
captioning one
image in and
outputs a
sentence
Sentiment
analysis - given
a sentence
classified as + /
- sentiment
Machine Translation:
reads a sentence in
English and then outputs
a sentence in French
Video classification
where we wish to
label each frame of
the video
22. TURN SMART
IoT
Big
Data
AI/
ML
Senses Remembers
Decides
Billions of devices
Covers the planet
Internet connected
25 billion billion b/d
It’s a collective memory
Forensics for training
As-is for decisions
Unsupervised learning
Predictors
Optimizers
Advisors /Agents
Saves to…
TrainsCommands
23. TURN SMART
connected
devices by
2020. 10b
today
$15b
Spent on smart
home
$2t
Smart Industrial
by 2020
$500 b/y
Driverless market
$100 b/y
smart office
+50b
Spent on smart
things in next
5 years
$6t
biz
Lower costs,
increase
productivity,
open markets
gov
Improve
citizens
quality of life
$1t
Smart city market
$2t/y
smart factory
market
24. TURN SMART
Biomimicry training
EVTOLS
Cars as swarms:
• Separation
• Alignment
• Cohesion
…and the anticipated human reaction:
1. Awe
2. Wide Berth
3. Acceptance
4. Treat Like Second-Class Citizen
5. Begin to Ignore or Disdain
TRANSPORTATION
28. TURN SMART
YOU
Chat-bot services everywhere
Digitial-Twin - Digital-Self Advocacy
BOTs on Boards
AR/VR - Just-in-time advisors in your ears
and eyes
5K device interactions per day per person
Robotic wealth management
Insurance & citizen ranking
29. TURN SMART
3 AI Perspectives
AI Kills Us AI Saves Us AI Is Us
For me the AI revolution start 30 years ago when my cousin Henry stood in my driveway waving a research paper he had found on AI. We both had worked on the first Space Shuttle math model. We had both coded control systems for nuke power plants. But we wanted to do our own thing. We coded up the AI algorithms for inference engine-driven expert systems. Got it running on small early personal computers as opposed to mainframes. We placed a small ad in a few computer magazines for an everyman’s $95 AI tool and sold 30,000 copies. Bam, we were launched a company called LEVEL5. Sold it to a big software firm 3 years later.
I have since been involved in the commercial launch of 7 smart product lines. They all only did just OK in the market.
Now and finally we are seeing AI really breakout. Let’s talk about how, why and what to expect.
For a broad historical context, you just so happen to live in a time unlike any other time in the history of innovation. Bigger than the renaissance, than the industrial revolution, than electrification. It includes the computer revolution, the internet revolution, the internet of things, big data, cloud computing, CRSPR, autonomous vehicles, commercialization of space and the beginning AI breakout.
You live in that vertical line to the right where change happens virtually overnight..
Here is the life-cycle of change. Lots of fast attempts, quick failures, and then bam, an inflection. Expansion, proliferation to saturation.
The inflection occurs when tech gets easy enough, powerful enough, and cost effective.
My AI story speaks to the S-Curve of tech adoption. We tried, made some progress, but never broke out into widespread adoption. Expert systems with logical rules for symbolic expression of logic, were to hard to write, needed expert’s time, and the solutions were brittle, unique and transient. We never hit an inflection point.
Some example breakouts:
Horse to cars – a 15 year transitions
Black & White TV to color – 10 years
Land lines to cell phone dominance was 5 years
Now we are in the inflections of:
Gas car to electric vehicle dominance –will happen fast now
Oil energy versus solar - PV went through grid parity with natural gas in 2016 and is still headed steeply down in cost
The common venture capitalist view is AI has passed thru inflection and it is time to pour on the rocket fuel.
Andrew Ng at Stanford characterizes AI as “the new electricity”.
Some samples S-curves you have seen in your life.
Mobile users in light blue.
Smartphones in darker blue
PCs saturated and topping out
Where is AI at and headed.
AI has always been a difficult term. Inherently confusing. My first company’s t-shirts said “The only thing artificial about me is my intelligence”.
Think of AI as algorithms that try to mimic the results of thought. Software that attempts smart.
Smart software really has been here for a while and is a big part of all of our futures.
Looking at history we non-human animals have always created tools and machines to make our life easier and better.
That make our bodies and minds stronger. We create tools that amplify our human capabilities.
This process got us out of the cave, from behind the plow, out of the factory, and soon out of the office.
We seek to automate drudgery and repetitive tasks. This tends to free time for more valuable or more fun things.
Why not assume in looking forward that this is what we will continue to do with the tools of intelligent software.
How big a deal is this for us? The best word for me is disruption. Spanning major industry sectors, how our world economy functions and how the gears of our society turn.
AI is not in some distant future. It is here now and has been for awhile. You use it every day.
AI recommends products, music, helps search, trades on the market, builds things, is on your wrist, is starting to drive, fights your battles, reviews your taxes and calculates your credit risk.
Today there are over 7,500 projects in Google using AI/Deep Learning right now. Google is starting to call itself an AI-driven company versus a data-driven company.
AI is being built-in from the silicon up, using GPUs (graphic processing units invented for the gamers), TPUs (Tensor flow processing units for scalable cloud AI Apps). We are seeing a chip-set race to create special processors that algorithmically mimic neurological functions.
Near term impact scale:
Gartner projects in 2 years AI leads to $200 billion in new revenue. But it eliminates of some 943,000 jobs, although it also claims that will be counter-balanced by the creation of some 768,000 new AI-related positions. The + and - implies a huge need to transition skills.
By 2030, As many as 800 million workers worldwide may lose their jobs to robots and automation. This is equivalent to more than a fifth of today’s global labor force.
Clamp down on immigration, not enough farm workers, farmers are turning to robots.
Rapid reskilling will become key to avoiding a collapse in earnings and ability for a consumer-driven economy to function.
Automation of work is creeping up the ladder of the conventional job hierarchy. Displacing manual labor, displacing thought workers.
The AI breakout is driven by 3 changes.
#1 cheaper compute power, on the Moore’s law curve
#2 Access to enough data and rich data, long vector data
And #3 better algorithms such as deep learning neural nets.
We are moving from programming computers to showing stuff to computers, who then self-generate the solutions / applications.
Machine Learning is the focus of most AI breakout success
Who has used an auto-translator?
Talked to their phone?
Seen face recognition on images?
Gotten a personalized video recommendation?
Used adaptive speed control or lane minders?
Has a robotic vacuum cleaner?
Talked to your lights?
Talked to a chatbot?
Been told a joke by their computer?
Gotten automated medical advice?
Our machines are starting to see, hear, speak, think, walk and decide to get out of the way
Machine learning in the catalyst of the breakout.
Been around since the 1950’s. Recently started to show impressive results.
ML can be decomposed into : Unsupervised, Surpervised, and Reinforcement.
Unsupervised where the data itself provides enough information to identify a cluster, a best path, a new meaningful representation, a detectable anomaly.
Supervised where we have a human / an expert / a curator in the loop, providing guidance and feedback to the emerging model and its consumed training sets.
Reinforcement is where repetition and interaction with the world provides the feedback that can guide the model to converge on a good-enough solution.
More specifically in Machine Learning the biggest breakthroughs have come from Deep Learning.
In Deep Learning, features and clusters are discovered by the network. It has been achieving better than human expert performance. It is consuming very large corpuses of information and auto-deriving classifiers and pattern recognizers.
Not to say that most predictive models come for free. There is still considerable effort and expertise required to generate models that converge on good enough solutions. But that effort is largely data science algorithm selection, tuning and data wrangling.
The key in modern ML via Deep Learning is that non-experts with access to big data sets, cheap cloud compute resources, and Deep Learning model builders can produce predictors that beat experts.
How does Deep Learning work?
This of the case of identification if a cat or a dog is in an image.
An image is a matrix of red, blue, green intensity values.
A neural network node at the lowest layer can be trained to have the discrimination of whether a cluster of pixels represent an edge.
Another layer up. Node can be trained to determine, from the output of the lower layer, if the edges represent a cohesive shape.
Another layer up may determine is that shape is an animal
And so on until the top layer classify and generate a probability scalar result that the image contains a dog, and therefore not a cat.
Each node gets many inputs in, runs an tuned algorithm and generates a single scalar output out. ML training is tuning of these algorithms and their weights to converge on answers that mirror a training set.
ML is about building models that predict outcomes given data.
The key is having enough of the right kind of data.
Some historical reference set is split 80 / 20 for training data and test data is set aside. A model is selected and tuned using data science and or training algorithms. Creating a tuned model. Real incident data is run through that model, results are predicted. Results are measured against the test data for quality. New model weights are adjusted. The model is updated. Testing and model tuning cycles repeat until “good enough” solution is reached.
Overtraining or over fitting a Deep Neural Network causes it to become brittle. Overfitting is using to many nodes. It becomes trained to too closely match the results of only the training data.
You want to generate a predictive model with the least nodes possible to achieve a generalized solution that can mitigate ambiguity in the data.
The example if the child getting their hand burnt on a stove. They extrapolate really well to things that are hot and a deep respond for touching stuff on the stove, after a single training example.
Neural networks come in different topologies
One 2 one – an image in a classification out
One to many – an image in a caption out
Many to one – a tweet in a sentiment out
Many to many – an English sentence in a French sentence out
Many to many – a video clip in where each frame is a part of a sequence and meta data out
We want AI to give us general capabilities: we want Robby the Robot who can do the dishes, drive us around, walk the dog, stitch up your cut, be your lawyer and babysit the kids.
However, what we can do today is deep and narrow. It does one thing really well and better than you. Plays Jeopardy, Chess and Go better than humans. Reads MRIs, and X-rays better than the docs. We get a smart washing machine. Smart Machines that see patterns we miss, optimizes their better and faster than us, that behave consistently, do not sleep.
For example, dermatologists make a lot of money and take many years to train to be good. A couple of non-medical, rather computer science grad student at a class in Stanford, used Deep Learning NN, and a big bunch of curated skin disease images. Their model matched and in some cases surpassed the top 3 dermatologist in the world. Non-experts, using a new flint knife and bear skin better world class expertise.
So why is this not already on your phone. Taking pictures of every funny bump on your skin. It s not yet ready, approved, situated in a treatment protocol. But it is coming.
Remember do not expect this dermo app to drive your car.
What patterns can we detect today?
Image classifications
Customer intent and sentiment
Autonomous navigation on the ground, in the air, on the battlefield
…and the ability to discover new relationships from data
A breakdown of our near-term capacities:
You can achieve better results to the questions you have now.
You now have the ability to ask new kinds of questions
You can now consume and use data that computers could not utilize before. Speech, images, video, unstructured text.
Pull back and we see commercialization of image, text and language processing. “Alexa”, “Hey Google”, “Siri”.
CNNs are Convolutional Neural Networks. Used to break a component (such as an image) into subcomponents (such as an object in an image, such as the outline of the object in the image, etc.)
RNNs are Recurrent Neural Networks. Used to detect features and classify from sequences of images (aka video), sentences of words, sequences of sounds, aka voice and music.
Approaching commercialization: Q&A engines, trainable Chabot's, best-path services, biomimicry in robotics, that is robots learning by watching us.
Not yet ready but proving to be powerful are adversarial and reinforcement learning systems.
Adversarial training pits one learning model against another. One trying to find the right answer in the data. Its adversary trying to spoof or fool the learner. This iterates until converging on a more hardened, safer from hackable and more robustly trained generalized model.
In reinforcement the focus is on on-line performance, which involves finding a balance between exploration of the world and and exploitation of current knowledge. A robot experimentally learning how to pickup a strange new object.
AI and machine learning services are commoditized now as building block capabilities that are hosted by the cloud service providers.
The intent is that they get integrated into all next-gen applications.
This accelerates the cost-effective adoption of AI / machine learning into any and all automated solutions.
3 major technology surges, synergistically combine and cross support each other for a greater impact
The Internet of Things senses. Billions of sensing device blanket the planet, connect to the Internet.
IoT saves its data streams to in-cloud Big Data. Big Data is out collective memory. It is the resource we mine and learn from. It is the stream of now that we react to.
AI and ML eat Big Data and learn. Converging on good enough predictive models.
Models detect patterns.
Cognitive services think, predict, decide, and act.
AI talks back to IoT closing the loop. Sending commands to turn left, or buy that stock, or send you a hint.
Slow reactions cycle through the cloud. Fast reactions happen at the edge, by the swarm, sometimes called the fog.
How big is the bang?
$50 plus billion internet connected devices by 2020. We already at 10 billion today.
$2 trillion spent in smart industrial
$500 billion / year in the driverless market
$6 trillion spent on smart things in 5 years
$1 trillion spent on making our cities smart
$15 billion spent on the smart home
$2 trillion / year on the smart factory market
$100 billion / year on the smart office
Governments care – hoping to improve quality of life
Business cares – shooting to lower costs, increase output and open new markets
Let’s drill down a little into a few sectors.
Transportation is a trillion $ sector in major upheaval. In part fueled by autonomous driving by AI, gig economies, EV battery tech
Where learning to drive differs per car, per region, per city, per time of day, per type of traffic pattern. Where minicry of humans has proven effective in training the self-driving models.
Where groups of cars traveling near each other can operate as a swarm. Vehicles senory networks have 3 bubbles. On car, collective perceptions extended by sharing info across a group or swarm of cars, and the 3rd is the fine grained digital map of the world map and transportation laws and constraints.
We are expecting to see a wave of reactions as self-driving vehicles surround us.
The big picture of this transformation upends entrenched constituents.
Road infrastructure itself transforms. Lanes narrow, roads recharge, roads are solar panels. Road disappear as we go VTOL.
Do we own or pay-as-you-go?
Who fault is a crash?
If we do not buy what do dealerships become, regional fleet managers?
Gas stations – dead
Why have stop lights? The vehicle phones in and reserves and air slot for passing through.
Parking is a huge $20 B industry in cities. If I do not park, or own, and vehicles are shared and in +60% utilization, why park?
Healthcare
I was invited to a future tech conference hosted by the governor of Wyoming just after I went through hurricane Irma. Wyoming, I come to find out has a billion dollar war chest. Their problems stem from a small remote population largely rural population. My problem after the hurricane is I was disconnected from information, felt small, powerless, remote and blind. I did not have access to what I needed or no other way than manual means to find things out. I wanted long-range low power reliable connectivity and an in-cloud avatar helping me out. Do not make me use resources (gas) to find resources (gas).
Advances in miniaturization, batteries and low power sensors are leading to a family of “insideables”. In your blood, intestinal track, eyes, teeth, bones, muscles.
Machine-learning models can monitor and predict remission, relapse in cancer patients
Arm band motion sensors, and face recognition can detect and predicting depression
Robotized Exoskeletons
Automated and personalized wellness coaches
Enablement of aging-in-place with smart floors, sinks, toilets, chairs, mirrors and nurses.
Industry – makers of stuff
You can bring factories back to the USA but the manufacturing line will be manned by robots. The jobs created are to aid and maintain the bots.
Efficient co-gen facilities
Fine-grained asset tracking, down to the smart box that knows where it has been and what it has been exposed to.
Autonomous EV fleets
Smart integrated end to end supply chains.
Predictive maintenance
More efficient, cleaner, cheaper ops, less defects, happier customers.
The city becomes the smart city.
Cities vie for smart businesses and higher paid work forces.
Cities attract talent and attention. Mayors gain power by piloting smart projects.
A smart city offers better services, is cleaner, is safer, is flexible, is convenient, is more entertaining, is connected. First wave is street lights. Next wave is parking. Then co-power gen.
Asia is surging in this area..
…and you
Automated bots for all initial service touch-points
Avatars that monitor and defend your digital persona
By 2024 most boards will have an advisor bot on the board of directors
Intelligent advisors, trainers, assistants at work, in home and in wearables
By 2025, an average connected person anywhere in the world will interact with connected devices nearly 4,800 times per day – basically one interaction every 18 seconds.
That which you own, save, buy, sell and exchange with be managed via your own AI advocates
Because we know so much about you, you will be scored, as to whether are you a risk, are you a civil threat, are you healthy, are you a “good” citizen.
Watch China on citizenship scores.
Lastly, we should talk about the “creepy-factor” of AI
In my experience, there are 3 main camps people fall into. AI kills us. AI saves us. AI is us.
AI kills us
AI is like nuclear weapons
Where we live on the edge of potential annihilation. This is the stuff of AI in sci-fi movies.
Many very smart people say AI must be monitored
It must not get out of control. It must not get self-conscious. It should not be weaponized.
Like nukes we care about AI getting in the wrong hands.
Certain AI should have legitimate terms of engagement
AI saves us.
It helps to bail us out.
It is fundamental to solving our intractable world problems.
Food production, water management, Alt fuels
It is fundamental to our security.
It is fundamental to global progress
It is fundamental to life, liberty and health.
It makes life much much much better
It steps in where we have skill, staff, dangerous occupation gaps.
Take Japan. A major labor shortage, getting worse as its population ages. Expected to be filled by automation.
Lastly – AI is us
AI apps get closer and closer to us. Know us. Get inside us. Become part of how we hear, think, see, feel.
And / or we get our mental memory, persona, and though processes replicated into the machine. We drop the body or bind it to the machine.
Given the exponential curves of power and performance in computing to we converge on a place where we go into the matrix
And is that a life as a separate consciousness or a collective mind hive?
So I see this debate as important, full of cultural change, and steep in ethical ramifications.
Show of hands. What camp are you in?
Thanks for your time and attention.
Here is all of my contact info. Feel free to reach out and contact my avatar.
Questions?