harnessign_AI_schedule_2019

Harnessing AI

What is Al?

1

Origins and history

Having made steady progress since its founding in the 1950s, the AI field has now accelerated to the forefront of computer science, propelled by super-fast microprocessors, availability of big data, and our supreme appetite for automating human cognitive tasks.

Peter Denning
IntroductionLecture slides | Recommended reading | Watch the video

 

2

Hierarchy of Al machines

The many existing AI machines can be ranked by learning power into six tiers.  The machines of a tier can learn to do some tasks that lower tier machines cannot do.  Although they learn, none of these machines is intelligent.

Peter Denning
IntroductionLecture slides | Watch the video

 

Classifying Machines by Kinds of Learning

3

Automation

Automata are machines that can do jobs faster and better than humans.  Traditional automata may have adaptive feedback, but they do not learn new functions.  AI is transforming automation by providing automata that do learn new functions.

Joshua Kroll
IntroductionLecture Slides | Recommended ReadingWatch the video

 

4

Rule Based Al

Expert systems were among the first machines aimed at learning the skills of experts.  They encode rules described by experts into logic formulas in a database so that a machine can follow the same logic and solve the same problems.  However, because human experts take actions not describable as rules, no software expert system has become an expert.

Vinnie Monaco
Introduction | Lecture slides | Watch the video

 

5

Supervised Learning Al

Recognizing faces in images is a human function that we do not know how to describe with rules.  With a database of 100 million labeled images, we can teach an artificial neural network to name the faces when shown the images.  Unfortunately, these networks cannot explain what they do and are very sensitive to small, pixel-level changes in images.

Marko Orescanin
IntroductionLecture slides | Watch the video

 

6

Unsupervised Learning Al

Large, quality data sets sufficient to train neural networks are difficult to find or expensive to gather.  We have figured out how to get machines to become grandmasters at Chess or Go by playing millions of games against each other and rewarding themselves for winning.  No external supervision of learning is needed.  Can this translate to other AI?

Chris Darken
IntroductionLecture slides | Watch the video

 

7

Human-Machine Teaming Al

When IBM Blue beat him in 1997, Chess grandmaster Kasparov invented a new kind of chess played by human-computer teams.  The teams beat the best machines.  Finding ways to use machines to augment rather than replace human intelligence is a central question in AI.

Rudy Darken
Introduction | Lecture slides | Watch the video

 

8

Aspirational Al

Machines that can carry on intelligent conversations, think, understand, create, care, be self aware, or be sentient are well beyond our current understanding.  Such aspirations have inspired perseverance in the search for intelligent machines.  Our best near-term bet is to concentrate on machines that team with humans, augmenting and amplifying us.

Neil Rowe
Introduction | Lecture slides | Watch the video

 

Critical Domains

9

Data Science and Al

Statistical inference tools that find patterns and trends in large data sets have come extremely useful.  Making sense of the vast troves of sensor data from our ever-growing network of Internet "things" requires enormous computing power.  Working synergistically, data science provides the methods and AI the tools to learn from the data.

Major Ross Schuchard (USA)
Introduction | Lecture slides | Watch the video

 

10

Management and Al

Many aspects of organizations and enterprises can be automated – for example, personnel record keeping, project tracking, manpower forecasts, applicant screening and selection, workflow, budgeting, planning, and supply chain management.  But many jobs remain for managers such as performance assessment, counseling, coaching, mentoring, and building high performance teams.

Uday Apte
Introduction | Lecture slides | Watch the video

 

11

Computer Vision and Al

Automatically identifying persons and objects in images has been a long quest back to the 1950s.  Much has been learned about feature extraction – patterns of edges and areas that group into identifiable objects.  That is being combined with the "convolutional neural network" to provide sophisticated image recognition.

Mathias Kolsch
Introduction | Lecture slides | Watch the video

 

12

Cyber Security and Al

By using biometrics and monitoring user actions for deviations from profiles of authorized users, AI has enhanced cyber security.  But new AI tools have vulnerabilities we do not yet understand, such as the neural network that mistakes a stop sign for a speed limit sign when just a few pixels of the stop sign image are changed.  Will adversarial AI get so good that no one can rely on AI tools for military operations?

Britta Hale
Introduction | Lecture slides | Watch the video

 

13

Natural Language and Al

Finding ways that computers can understand text and speech and translate between languages has been a long quest in computing.  Progress has been slow but steady.  We now have limited AI tools such as Alexa and Siri as our interface to the computer, language translators, dictation recorders, and real-time voice translation between languages.

Neil Rowe
Introduction | Lecture slides | Watch the video

 

14

Robotics and Al

A robot is a machine that has human-like behaviors and performs human-like tasks.  Robots for assembling products in factories have become quite advanced.  Robots need enormous amounts of quality data about the tasks they must master.  Machine learning is only part of their training.  Finding good reward functions to accelerate the learning is a difficult engineering problem.

Brian Bingham
Introduction | Lecture slides | Watch the video

 

15

Ethics and Al

AI has raised numerous ethical dilemmas.  Who is responsible when an AI machine fails and causes harm?  Should weapons systems be allowed to on full automatic, making kill decisions without consulting with a human?  Is the programmer of an AI tool used in a military operation a combatant?  Is it wise to search for intelligent machines without knowing that we can control them?

Bradley J. Strawser
Introduction | Lecture handout | Watch the video

 

Moving Forward

16

Al and strategy

Military strategy is concerned with setting guidelines within the context of technology and geopolitics that minimize the chance of war but maximize the chance of winning should war occur.  AI has assumed a central role in the context of great powers competition.  What kind of AI would be most useful?  How should the US, whose political system encourages cooperation between government and industry, compete with China and Russia, whose systems mandate alignment between government and industry?

Wade Huntley
Introduction | Lecture slides | Watch the video

 

17

Risks of Al systems

AI can be alluringly attractive because it can do certain jobs well beyond the capabilities of humans.  It can also be alarmingly unattractive because it can make serious mistakes so fast that no human can intervene.  Where are the risks in the kinds of AI we have been discussing in this course?

Neil Rowe
Introduction | Lecture slides | Watch the video

 

18

Al in the DOD

Motivated by a concern to deal realistically with newly emerging great powers competition, the US DOD has issued various strategy documents for AI research, implementation, and industry cooperation.  What offices and programs has DOD set up to accelerate adoption of AI technologies?

Bret Michael
Introduction | Lecture slides | Watch the video

 

19

The Next Face of Battle

In 1979 John Keegan published a book, The Face of Battle, in which he analyzed the practical mechanics of battle and how they affect outcomes as much or more than strategy.  AI is radically transforming the space of possibilities available to commanders and warfighters.  It emphasizes small, networked, distributed, swarming, autonomous agents over large platforms.  AI also adds a new dimension, where the agents have their own decision-making authority.

John Arquilla
Introduction | Lecture handout | Automation Will Change Sea Power (essay) | Watch the video

 

Conclusion

 

20

Conclusions

This course has sought to cut through AI hype by exposing the base principles of AI.  Although we currently have no intelligent machines, we do have six varieties of machines that can rapidly learn to do complex human tasks.  These machines have produced significant advances and vulnerabilities in important domains such as vision, robotics, natural language processing, and cyber security.  The quest to make these machines reliable and secure has unearthed a host of dilemmas that implementers must face.

Peter Denning
Lecture slides | Recommended Reading | Watch the video