Artificial Brain - Overview 2013

3,841 views
3,395 views

Published on

Artificial Intelligence is being supplanted by "Artificial Brain," i.e. neuromorphic technologies. Yet there still a whopping gap that neuromorphic systems need to close before they will become a match for successful AI applications.

Published in: Technology, Education
1 Comment
3 Likes
Statistics
Notes
No Downloads
Views
Total views
3,841
On SlideShare
0
From Embeds
0
Number of Embeds
51
Actions
Shares
0
Downloads
296
Comments
1
Likes
3
Embeds 0
No embeds

No notes for slide

Artificial Brain - Overview 2013

  1. 1. Future with AI The Artificial Brain Odyssey Dr. Giedrius Tomas Burachas, Intentiva Inc.
  2. 2. I and my brain 100 billion neurons, 300 billion connections in 1cm3 = # stars in Milky way
  3. 3. AI: predictions, hopes and reality
  4. 4. ...but what is AI? John McCarthy (1955, roots in Lithuania) suggested: AI is a science and engineering aimed at creating intelligent machines But what is an “intelligent machine”? Turing test (1950): a machine can reason if: It can imitate humans so that an average person would assume he/she is dealing with another person New definitions: Jeff Hawkins: an intelligent system can correctly predict what’s coming. Created hierarchical temporal memory (Numenta, Grok; Vicarious Systems) Dr. Kestas Kveraga, Harvard University: visual brain systems use predictive feedback loops to speed up object recognition Penti Haikonen (Nokia): intelligent systems can predict outcomes of their actions Imagination is required in order to simulate consequences of choosing alternative actions. Igor Aleksander: computers with imagination
  5. 5. Achievements of AI Playing chess: IBM’s Deep Blue computer defeated the world champion Garry Kasparov in 1997. Used simple search algorithm Triumph of pure AI: no inspiration from ther brain Speech synthesis: Apple popularized it 25 years ago with Macintosh. computer algorithms have trouble pronouncing irregular words. Terrence Sejnowski, Salk professor, in 1987 created a neural net Nettalk that learned to pronounce them in a way like human baby. Showed advantages of brain-like algorithms. Face detection: Face detection algorithm is ubiquitous in photo cameras. Based on the algorithm invented by MIT professor Paul Viola. Used inspiration from visaul system of primate brain. Had many conversations with him Face recogniiton: A harder problem, but several companies solve it successfully, including Dr. Algimantas Malickas’ company in Vilnius: Neurotechnology. Relies on AI methods (machine learning, such as SVMs)
  6. 6. Achievements of AI, cnt’d Voice recognition: Nuance Dragon used in speech transcription, e.g. medicine. Google Voice – suitable for transcribing 70-80% voice messages. Weakness: break down in presence of accent, background noise. Language understanding: IBM Watson won “Jeopardy” game (2011) Largest project – Stanford Research Institute’s “personal assistant” (DARPA). SRI system was commercialized and Apple bought it in 2011: SIRI app for iPhone. Maintaining dialogue: No machine has passed the Turing test Recognition of human actions and intentions: Simple algorithms for human action recognition in security tapes (San Diego) Numenta (Vicarious ?): uses brain- inspired principles for human action detection, and other applications. Intentiva: recognizes human intentions Recognizing emotional and social states: San Diego company Emotient recognizes emotions from facial expressions.
  7. 7. AI vs. Artificial Brain - Principle of operation: logical rules created by engineers - Works with standard digital computers - Software is separated from hardware - Memory is separate from CPU - Advantage: the logics of operation is straightforward to understand - Disadvantages: - Limited tolerance for errors - Needs a lot of energy for each computational cycle - Principle of operation: adaptive processes, statistical learning rules - Requires neuromorphic electronics - Software is merged with hardware - Each processor has its own memory - Advantages: - Tolerance to errors and malfunctions, ability to compensate them - Asynchronous parallel operation needs little energy - Disadvantage: unexpected emergent phenomena can arise - e.g.: artificial “epilepsy”
  8. 8. When computers and AI surpass human intellect? Ray Kurzweil
  9. 9. Robotics: realizing brain-inspired algorithms in computers UCSD robot child Recognizes and reacts to human emotions All modern robots run on traditional CPUs/GPUs EB project “Artificial curiosity”: robots with internal motivation for learning Robots “animats” Emotional/social robots Robot training imitates development of a baby Robobee
  10. 10. The Brain Levels Cells: neurons Neuronal circuits Nuclei, “maps” Functional systems (e.g., vision, audition, speech, motivation, decision-making, motor) Global behavioral systems Function Integration of information and feature representation Neuronal algorithms (amplification, WTA, sparse code) Support a function Processes info for supporting elements of behavior Choreography of behaviors, planning Churchland & Sejnowski
  11. 11. Neurons- most complex cells Neuron’s job: integrate spikes coming from other neurons via synapses and decide whether to generate a spike - Spikes are on or off - Spike influence is determined by synaptic weights and synchrony - Excitation and inhibition are balanced Buračas et al.
  12. 12. Neuronal circuits – info processing units Most famous neuronal circuit: macrocolumn of cortex, Decision-making unit Minicolumn: ~80 neurons Macrocolumn = 100 minicolumns (0.5mm) Blue Brain project: Dr. Henry Markram EU project >1 bln Euros, Lausanne & Heidelberg univ. (86 institutions, 10 years) Evolved Machines: Dr. Paul Rhodes
  13. 13. Cortical “maps” & functional systems Kveraga et al. Visual system 2 mln macrocolumns in human cortex Prediction mechanisms speed up object recognition
  14. 14. Brain and behavior Brain rhythms Excitation/suppressio n Sleep/waking cycle Motivation Foraging behavior Fight/Flee responses Reproductive behavior Universal neurochemistry of love: dopamine->testosterone->occytocin
  15. 15. Cartography of human brain (fMRI) Language Introspection Speech generation Speech understanding The same parts of the brain are activated when observing, dreaming and imagining the same scene/object Network for Introspection (yellow) Brain/mind reading:
  16. 16. Neuromorphic systems SyNAPSE – the most ambitious program for building neuromorphic systems (Systems of Neuromorphic Adaptive Plastic Scalable Electronics) DARPA, HP, HRL, IBM and 5 universities (2009-now: ~$100M) Neurosynaptic chips Neurosynaptic processors Goal: Brain-like chips and systems with - 10 bln neurons - 100 trln synapses - 1kW - 2 liters of volume achievements: - 2009 cat brain-sized system simulated on Blue Gene - 2011 a chip with 256 nrn, 250K snps - Not finished: 2013 multi-chip neuroprocessor system (1M nrn on each chip)
  17. 17. Why neuromorphic systems are coming? The brain is more effective energetically than computers by a factor of 1mln – 1bln
  18. 18. Neuromorphic chips SyNAPSE DARPA chip: 256 neurons (digital) 262K programmable synapses 65K trainable synapses NeuroGrid, Stenford Uv, 2009: 1 mln. Neurons 6 bln. Synapses Real time(10 spikes/sek) 2W! But not trainable... Neuroinformatics Institute, ETH, Zurich, 2013: neuromorphic trainable chip Silicone physics used to simulate neuronal biophysics Implements “liquid states” just like the brain Qualcomm and The Brain Corporation (>$40 mln): Secretive, however, Qualcomm is building a chip Dr.E. Izhikevich Dr. Rodney Douglas Dr. Quabena Boahen Dr. Dharmendra Modha Schematics of the digital artificial neuron
  19. 19. Applications of neuromorphic systems Applications for vision: Several neuromorphic chips connected serially by analogy to human visual system: can attend to interesting objects in the visual field.
  20. 20. Neuroscience is moving towards neuroindustry: spiking camera
  21. 21. Conclusions Machine learning/neural network research has been pushing the limits of computer vision and AI, but to date it has been mostly based on traditional von Neumann type computer architectures As the amount of data and complexity of the systems grows, the need for neuromorphic solutions becomes more apparent Since neuromorphic systems as of 2014 are still very far from being of practical value, interim solutions take the shape of ASICs dedicated for machine learning tasks (e.g. Nervana Systems, San Diego) N.B. the images and video clips used herein are protected by copyrights of respective copyright holders

×