may 5, 2017 tiecon, @santa clara santa clara university...
TRANSCRIPT
How AI is Eating the WorldSanjiv R Das
Santa Clara University TiEcon, @Santa Clara
May 5, 2017
Thanks to Arpita Khandelwal for her extensive research in preparing these slides.
Outline
● What is AI?
● Recent applications of AI.
● Why has Deep Learning transformed AI?
● Dangers of AI.
Definition of AIIntelligence exhibited by machines● Artificial Intelligence (AI) /Narrow or Weak AI: “Expert systems that
match or exceed human intelligence in a narrowly defined area, but not in broader areas” e.g. Siri
● Artificial General Intelligence: An artificial neural network that wouldn’t need to be preprogrammed with fixed rules
● Machine Learning (Tom Mitchell): "A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P if its performance at tasks in T, as measured by P, improves with experience E."
Ability to rewire itself to reflect patterns in the data it absorbed, a mechanism adaptable to its environment, in which advanced skills would emerge organically: “Humans don’t learn to understand language by memorizing dictionaries and grammar books, so why should we possibly expect our computers to do so?”
http://srdas.github.io/MLBook/
Two Types of AI● Rule-Based AI
● Data-driven AI
Example: Checkers
Albert Samuel @IBM began writing code for a checkers game program in 1949. In 1956, the program was demonstrated to the public on live television. In 1962, the computer beat checkers master player Robert Nealey, and IBM’s stocks rose 15 percent overnight.
Rule-based AI can never be more intelligent than its creators, but data-driven AI can!
Usefulness of AIAI in Wearables:
Emotion detectors Could Tell You When The Conversation Is Getting Awkward!
https://www.fastcompany.com/3066867/wearables-that-detect-emotion-could-tell-you-when-the-conversation-is-getting-awkward
Conversational AI
Ford:● Invests in new artificial intelligence software company, Argo AI● Develop the virtual driver system — the brains of their self-driving vehicles● Designed to sense and interpret the world in real-time
Drive.ai: ● Test Vehicle depicts incident-free, unbroken footage shot on a rainy night
Google’s Self-Driving Cars:● The National Highway Transportation and Safety Administration - artificial
intelligence system that controls its self-driving car can be considered a driver under federal law!
AI in Automobiles
Deep Patient:
● An Unsupervised Representation
to Predict the Future of Patients from
the Electronic Health Records
● Discovered patterns hidden in the
hospital data to indicate people’s disposition
to a wide range of ailments, including
cancer of the liverhttps://www.nature.com/articles/srep26094
AI in HealthCare
Image RecognitionThe CIFAR-10 dataset● Consists of 60000 colour images in 10 completely mutually exclusive
classes
● 50000 training images and 10000 test images; 18% test error without data augmentation and 11% with
AI and Gaming
Brains Vs. AI, which began Jan. 11 at Rivers Casino in Pittsburgh, pits Chou and three other leading players — Dong Kim, Jason Les and Daniel McAulay — against Libratus in a 20-day contest in which they will play 120,000 hands of Heads-Up, No-Limit Texas Hold’em poker. All four pros specialize in this two-player, unlimited bid form of Texas Hold’em and are considered among the world’s top players of the game.
Libratus had amassed a lead of $459,154 in chips in the 49,240 hands played by the end of Day Nine.
Positive Reinforcement: ● When machines figure out a way by trial and error and then associate the
positive outcome with the actions that led up to it
● The machine learns without instruction or even explicit examples
● E.g. AlphaGo’s historic Victory against Lee Sedol
AI and Leisure - Gaming
https://www.technologyreview.com/s/603216/5bigpredictionsforartificialintelligencein2017/
Deep Learning
Old Mathematics + New Technology
Artificial Neural Networks + Modern Chip Technology
http://srdas.github.io/DLBook/
Implementing Deep Learning to Detect Breast Cancer
Wisconsin Breast Cancer Dataset
Training the deep learning net
Fitting statistics
60,000 x 784
Validation Metrics (hidden layers 10,10,10)
7510 parameters
Validation Metrics (50,50,50,50,50)
46610 parameters
Validation Metrics (100,100,100)93010 parameters
Predicting stock index direction
● Use all stocks in the S&P500 from 1963 till today
● Fit a deep learning model to 80% of data randomly
● Validate on remaining 20%● 5 hidden layers of 512
neurons each● How accurate is the
prediction of the market direction next day? = 57%
● The statistical significance is a t-stat = 7.54
Latency
Vs
Throughput (#cores)
The Dark Side of AI
The biggest change is going to be seen in the labor markets. Every use of AI is a negative effect on a traditional labor market.
The Atomic Level of Work
David Beyer
The “one second” rule.
What tasks get automated first?
http://www.datasciencecentral.com/profiles/blogs/how-to-put-ai-to-work?xg_source=activity
AI and Humans!
● Algorithmic redlining by deep learning might arrive at the same result but without violating “protected characteristics”
● Broward county (FL) used COMPAS to evaluate re-offense probabilities, but predicted rates were much higher than realized
https://www.theguardian.com/technology/2017/mar/02/robot-tax-job-elimination-livable-wage
Historian Yuval Noah Harari makes a bracing prediction: just as mass industrialization created the working class, the AI revolution will create a new un-working class.
Technology
+
Specialization
The crucial problem is creating new jobs where humans are better than algorithms
AI and JobsThe White House Report (December 2016):
● Millions of lost jobs - the workers earning less than $20 per hour /without a high school diploma most affected
● 9% jobs - completely displaced in the next two decades ● 47% jobs at risk ● To reconcile - expand access to education in technical fields and increase the
scope of unemployment benefits● Calls for further investment in artificial intelligence research and
development - specifically, in cyber defense and fraud detection
“Anything we can do to have more AI will lead to more productivity growth.”
https://www.washingtonpost.com/news/innovations/wp/2016/12/20/ai-could-cost-millions-of-jobs-the-white-house-says-we-need-more-of-it/?utm_t
erm=.8e9c6e3e8073 http://www.oxfordmartin.ox.ac.uk/downloads/academic/The_Future_of_Employment.pdf
AI and JobsFINANCE:
Bridgewater Associates : ● World’s largest hedge fund - project to automate decision-making to save
time and eliminate human emotional volatility● PriOS, the over-arching management software
Goldman Sachs: ● Two out of the 600 equity traders left● Found that four traders can be replaced by one computer engineer● Reducing the number of investment bankers - great cost savings
https://www.theguardian.com/technology/2016/dec/22/bridgewater-associates-ai-artificial-intelligence-management
AI in Hedge Funds
Eurekahedge report : AI hedge fund firms with machine learning outperforming “traditional quants,” and doing so with low correlation
http://www.valuewalk.com/2017/01/ai-hedge-fund-returns/
Last Days of the Stock Picker● BlackRock Inc. announced that it is replacing human stock pickers with machine-run
algorithms for some of its equity funds, signals that the money management industry
is getting the message.
● The world’s largest hedge fund, Bridgewater Associates (US$160 billion in AUM), is
using software to automate its day-to-day decision making.
● The popularity of computerized quantitative trading strategies, and the growing use
of artificial intelligence (AI) techniques, stems in large part from their impressive
returns.
● AI and machine learning hedge funds outperformed both traditional quantitative and
the average global hedge fund, with annualized gains of 10.6 per cent over a two year
period, according to Eurekahege.http://business.financialpost.com/investing/last-days-of-the-stock-picker-as-money-managers-embrace-artificial-intelligence
https://www.linkedin.com/pulse/importance-ethics-governance-artificial-intelligence-marks?trk=v-feed&lipi=urn%3Ali%3Apage%3Ad_flagship3_search_srp_content%3Bz%2FBwgA7F4ss5MPHy8%2FYlkw%3D%3D
WHY THE NEED?
● On one hand potential to do a lot of good ; On the other hand, also being used in ways we don't necessarily notice or understand eg., profiling us for crimes we have yet to commit
● Critical to encourage an interdisciplinary approach to creating regulatory frameworks
Ethics and Governance in AI
Ethical Issues
Unemployment
Robot RightsEvil Genies
Inequality
Artificial Stupidity!
Behavior &
Interaction
Singularity &
Staying in Control
Racist Robots??Security
What’s Being Done?OpenAI :
● founded on the principle that AI should be advanced in a way to benefit humanity, unconstrained by the need to generate financial return.
The Ethics and Governance of Artificial Intelligence Fund:
● An attempt by the Knight Foundation, Reid Hoffman, Pierre Omidyar, the MIT Media Lab amongst others
● Encourage transparent, cross-disciplinary research into how to best manage AI
One Hundred Year Study on Artificial Intelligence (AI100):
● Stanford Study, a long-term investigation of the field of AI and its influences on people, their communities, and society
Can Super-powerful AI Eliminate Humans?
● How do we manage uncontrollable AI?
● Encrypted Deep Learning: Build a super-intelligent AI to detect and cure cancer (for example), but encrypt it so that it cannot replicate itself, nor can it be used without human control.
● Encryption prevents it from being stolen by a malevolent player.
● Homomorphic encryption: Only permit a limited range of encrypted predictions. (Allows you to operate/modify information without being able to see it.)
https://iamtrask.github.io/2017/03/17/safe-ai/
Summary● AI is a transformation, not a change.
● Super AI is now possible.
● AI is everywhere, from marketing to medicine and lots more.
● AI will disrupt the labor market, and only some jobs are safe.
● AI has a dark side, from law to lending.
● The second coming of AI is based on Deep Learning.
● AI raises ethical issues.
● AI will need controlling through encryption.
● Artificial General Intelligence (AGI) has arrived! Read more at Stuart Russell’s terrific collection of links:https://people.eecs.berkeley.edu/~russell/research/future/