New-Thousand years Machine Learning.


43 views
Uploaded on:
Category: Food / Beverages
Description
Machine Learning Today: Costly Trial and Error. Customary machine learning:Learn when numerous reiterations of trial and errorStuck on capacity based modelE.g., Language: WSJ Corpus, 1987-1989, with 39 million wordsHurts with applications:Trial and blunder not great in situations where mistakes killMedical roboticsThousands of learning trials can be expensiveAcquainting a robot with another healing facility would
Transcripts
Slide 1

New-Millennium Machine Learning Selmer Bringsjord, Nick Cassimatis, Kostas Arkoudas, and Bettina Schimanski Department of Cognitive Science Department of Computer Science Rensselaer Polytechnic Institute (RPI) September 21, 2004

Slide 2

Machine Learning Today: Costly Trial and Error Traditional machine learning: Learn simply after numerous redundancies of experimentation Stuck on capacity based model E.g., Language: WSJ Corpus, 1987-1989, with 39 million words Hurts with applications: Trial and blunder not great in situations where mistakes kill Medical mechanical technology Thousands of learning trials can be costly Acquainting a robot with another healing facility would take days Teaching individuals new programming makes them less profitable in the short-term. Machines prepare us now rather than us preparing them . Learning trials regularly not accessible Homeland security: Not a large number of individuals in flight schools Robots and programming in this manner constrained to restricted assignments and firm We are compelled to collect machine information physically CYC has over a million certainties and is not even remotely entire

Slide 3

Motivating Examples Millions of understudies are right now adapting principally by perusing - and same e.g. for grown-up scientists like us!

Slide 4

Human One-shot Learning Example DAK CUP

Slide 5

Insert motion picture here (Nick has a duplicate)

Slide 6

Behavior of Micro PERI

Slide 7

Implications of One-Shot Learning and Reading One-shot learning case required: Rich arrangement of representation and thinking capacities right off the bat. Where was speaker looking when he said "USB Converter". Social thinking to track where speaker was looking. Spatial and transient thinking to construe what he was taking a gander at. Existing machine learning calculations have no idea of space, time or human consideration. Factual speculation only one of a few learning procedures: Inference from single cases. Similarity. Impersonation. Direction. Adapting a great deal more socially and physically intelligent. Make inquiries: Why? How? What if? Physically test their own speculations about the world. Learning by perusing...

Slide 8

Our Proposed Solution: A New Research Program Study the human case - at people (counting kids) who learn Developmental brain research has demonstrated that even newborn children and babies have rich thoughts of: Time, place, causality, conviction, crave, consideration, number, and so on., and of induction over these ideas Develop formal hypotheses that demonstrate to utilize these components to make adapting speedier and more compelling Develop machine learning calculations utilizing this substrate that learns by: Explicit perusing and guideline Analogy Inference Imitation The hardware of MARMML and the Bringsjord/Arkoudas machine learning framework Demonstrate affect on applications Elder care Homeland security Trace out the ramifications of these calculations for better educating/learning in the human circle

Slide 9

Objection How is this a change over GOFAI? i.e., Why isn\'t this the 1970s once more? Less information of human adapting then Formal strategies in their outset Nothing like Athena (used to demonstrate a decent piece of Unix sound)! Like two-layer neural systems contrasted with greater ones Formal foundation was divided. Not known how to consolidate consistent and probabilistic information? So scientists were either utilizing no representation and thinking substrate or they were utilizing the wrong one. Incorporated psychological models for joining strategies not created, Polyscheme, ACT-R, ... These strategies were not intuitive. No question getting some information about human aim No experimentation

Slide 10

Field is prepared for another approach Recognition of requirement for incorporated psychological frameworks developing: Example: AAAI Fall Symposium on Integrated Cognition Hundreds of studies in newborn child cognizance give us a smart thought of what the right substrate is. Coordinated intellectual models exist and are propelling each day. Computational foundation there: Abundant computational power for various strategies in one framework. Robot and machine vision framework set up: Object acknowledgment Face acknowledgment, eye-following Mobility and route Robot control

Slide 11

Robot Manipulation (i.e. PERI)

Slide 12

PERI Psychometric Experimental Robotic Intelligence Scorbot-ER IX Sony B&W XC55 Video Camera Cognex MVS-8100M Frame Grabber Dragon Naturally Speaking Software NL (Carmel & RealPro?) BH8-260 BarrettHand Dexterous 3-Finger Grasper System PERI was not intended to reenact how a human thinks - AI, not subjective demonstrating

Slide 13

Impact on machine learning and counterfeit consciousness More adaptable and ingenious learning and thinking calculations. Mentally adaptable robots (once more, e.g., PERI) Faster learning. Learning in circumstances that were outlandish some time recently. Combination of thinking group once more into learning group.

Slide 14

Impact on investigation of human learning Existing exact work hampered by unclear speculations that make consequences of straightforward tests questionable. Formal hypothesis ought to help this. Grow better comprehension of which guideline or learning strategies are best in which conditions.

Slide 15

Some applications High-stakes applications where experimentation excessively hazardous. Country security. Perilous waste evacuation. Robots and programming for less modern or learning-tested people utilize them. Impaired. Senior care. Senior care robots simpler to use by the more seasoned set. Rising Robotics Technologies & Applications Conference Proceedings, March 9-10, 2004 , Cambridge, MA Rodney Brooks said Elderly Care as one of the present future patterns in apply autonomy: Currently: None Future: Robotic Assistants in Millions of Households Less weak, more broad, less demanding to-learn and utilize robots and programming. Better learning situations: Direct/train robots (PERI). More precise pinpoint reasons for issue learning.

Slide 16

Our advantages Background in crossing point of thinking and formal techniques, and learning Selmer and Nick and Kostas and Bettina Prior R&D in machine learning. Selmer Background in kid advancement. Scratch Integrated subjective models All four Background in mechanical technology Selmer and Bettina and Nick

Slide 17

Wargaming RAIR Lab Sponsors Cracking Project; "Superteaching" A while back, RPI Strategic Investment speculation era; AI in support of IA Slate (Intelligence Analysis) Item era engineered characters/mental time

Slide 18

What we will do with the impetus cash? Investigate focus: khlkjh lkjh ljkh Workshop or some other vehicle to get this set up together? Understudy bolster. Models, confirmation of-idea. Senior care robot: Robot that more established individuals can use with their own particular phrasing, e.g., when alluding to spots, drugs, and so on. Something with perusing? Site, papers, presentations ??

Recommended
View more...