Transcript Neural Network Algorithms-Quantum-Glia

Neural Network Algorithms
Review, Quantum and Glial Directions
Martin Dimkovski
CSE 6111 Presentation
York University
March 31st, 2011
Presentation Goals

Hope to increase awareness of NNs’ potential


(<10 min)




Very general toolkit, applicable to many problems
Overview:
Algorithm models
Computational power
Complexity, limitations
Interesting research directions

(<5 min)

(<5 min)
March 32st, 2011
Quantum
Glia
Neural Network Algorithms
2
Presentation Motivation

Can’t explain


Fight common over-simplification


0.5
ubiquitous (brains, computers,
society, space-time fabric), accessible & good
inspiration
On a serious note…:
Backpropagation-only/mostly view
Practical applications and solid theoretical
foundations exist for presented alternatives
March 32st, 2011
Neural Network Algorithms
3
Network Graphs

Unidirectional: simple, feed-forward


+ Bidirectional: recurrent, interactive


0.5
Backpropagation is here
Network dynamics comes up
++ lateral: resonance, competition,
pattern completion
March 32st, 2011
Neural Network Algorithms
4
Learning Approaches

Supervised vs Unsupervised

Error-driven








1.5
Straightforward
Mature
Adaptive basis functions
For well defined tasks
input>output, functions
Hebbian-Style


(ex: backprop)
(more on next slide)
More sophisticated, more bio-inspired, self-organizing
But not as mature (still weak like error-driven in 1960s)
Order and Chaos
Combinations (superior)
March 32st, 2011
Neural Network Algorithms
5
On Hebbian-Style

Network dynamics, more complex graphs




Build internal models of environ. - Identify principal features

Constraint satisfaction


Find energy function minima
Attractors = memorized patterns

1.5
Fire together – wire together
Compete
Resonate
Deal with corrupt and partial memory
March 32st, 2011
Neural Network Algorithms
6
More Modeling Aspects

Complex (superior) vs real valued

Temporal dynamics

Signal coding:



Static vs dynamic weights (in between training)



Stay the same for any input/output condition
Superior: Adjust to input/output condition
Excitation - inhibition modeling

1.5
Discreet, analog, pulse averaging, or
Superior: detailed pulse/spike pattern modeling
Superior: not on the same weight
March 32st, 2011
Neural Network Algorithms
7
Use of Probability

Bayesian: Solves a BIG problem

Over-fitting




Solves it because it samples from whole posterior and does not
depend on a single set of weights
To get a feeling, compare:



MAP = argmax P( | data)
E[] =   P( | data) d
The benefit of noise

1.5
When noise and peculiarities become more attended-to than the
general features of interest
Problem especially with error-driven, like backprop.
To avoid local minima/maxima (ketchup)
March 32st, 2011
Neural Network Algorithms
8
Research Directions



1.0
NNs have come a long way
Yet, still far below known upper bounds
 For precision, performance, and usability
What better place to turn for help, than back to our
original inspiration?
 In green are my personal speculations
March 32st, 2011
Neural Network Algorithms
9
Cues to Quantum

Classically unexplained brain features

Simultaneous synchronized stimulations in distant regions
for same stimuli





1.5
Perception unification – global attractor states
Speculation brain has ingredients for macroscopic
quantum state


highly structured in phase and amplitude
High metabolic energy; extreme dielectric prop.
Microtubules, superconducting waves, gap junctions
(anaesthesia)
Interest in just plain quantum computing power
March 32st, 2011
Neural Network Algorithms
10
A (Qu)bit of Quantum Basics

If left alone – a linear ‘combination’ of basis states (in
coherence)


Each |i is a single reality for us classical beings


Into one of the basis states, as per probabilities
Entanglement

1.5
|ci|2 giving the probability
If ‘touched’ by anything – decoheres


(ex: |0 or |1)
But in quantum world, they all exist at once


| =  ci|i
Instantaneous sync link between remote qubits
March 32st, 2011
Neural Network Algorithms
11
How can Quantum Translate
for Artificial NNs?
1.
Run existing NN algorithms on quantum computers


2.
Could we simulate the quantum ‘spooky’ effects in
new NN algorithms?

0.5
…getting there, but will take a long time
Extra slide in appendix
March 32st, 2011
Using our classical computers
Neural Network Algorithms
12
Simulating Quantum Effects


Maybe the brain uses certain quantum features for
evolutionary reasons.
Could we program/simulate?:


Synchronization and unification of distant physically
unconnected neurons?
Coherence and decoherence of macroscopic quantum states


1.0
even though we would have to use many more bits
Interference and quantum functions in discretized
approximations?
March 32st, 2011
Neural Network Algorithms
13
Old View on Glia

Myelinate for insulation only

Clean-up and recycle neurotransmitters

Feed and heal neurons

…But, Einstein’s brain

0.5
Double the glia
March 32st, 2011
Neural Network Algorithms
14
Recent Findings
Glia-Neuron and Glia-Glia Information
Processing

Listen to all neurotransmitters, and uses them to communicate with
both glia and neurons

Control synapse formation and operations


Connect neurons which have no synapses between them, and correlate
them

Run separate network in parallel to NNs

Control speed of neuron’s output (axon)

Most regulated genes during REM are in glia (integration/consolidation)


2.0
As many as 100,000 synapses per glia
It’s a whole new brain out there…
And there’s more:
March 32st, 2011
Neural Network Algorithms
15
Glia Quantum Correlates

Brain-wide calcium broadcast network




Glia Quantum Correlates



1.0
Connect through
 gap junctions
Calcium messaging affects neural circuits
Drive global broadcast waves
Calcium stores related to microtubules
Gap junctions as hypothesized
Quantum aspects might play a big function in glia networks
March 32st, 2011
Neural Network Algorithms
16
Glia as Biological Bayesian?

Accumulated effect from previous inputs (old posterior) serves
as baseline for new input (new prior).
Posterior (t )  Norm _ Const (t ) * Posterior (t  1) * Data(t )
2.0

Glia excitations last second to minutes, compared to ms for
neurons, and it span much wider

This could produce something alike cumulative data likelihood
during the period t of glia excitation

Glia could then adjust/sample weights for neurons as per latest
posterior (weight factors coupled to posterior)

Would need a mechanism for normalization to 1
March 32st, 2011
Neural Network Algorithms
17
Conclusion

Existing NN algorithms offer a rich toolkit for
computing




1.0
Much more beyond plain backpropagation
Take advantage of combinations and complex graphs
Use as many of the superior modeling aspects as affordable
Use probability theory

Glia networks and interactions with neurons can be
modeled in new algorithms

Might be possible to simulate quantum effects for
more enhancements
March 32st, 2011
Neural Network Algorithms
18
The End

Questions?
March 32st, 2011
Neural Network Algorithms
19
References

O’Reilly, Y. Munakata, “Computational Explorations in Cognitive Neuroscience, The MIT Press, 2000

V. Ivancevic, T. Ivancevic, “Quantum Neural Computation”, Springer, 2010

U. Ramacher, C. v.d. Malsburg, “On the Construction of Artificial Brains”, Springer, 2010

(Ed.) A. Volterra, P. Magistretti, P. Haydon, “The Tripartite Synapse”, Oxford University Press
2002

R. M. Neal, “Bayesian Learning for Neural Networks”, Springer, 1996




R. D. Fields, “The Other Brain”, Simon & Schuster 2009
S. Gupta, R. Zia, “Quantum Neural Networks”, Journal of Computer and Systems Sciences 63, 355383, 2001
A. A. Ezhov, D. Ventura, “Quantum Neural Networks”, Future Directions for Intelligent Systems and
Information Science. Physica-Verlang, 2000
J. J. Hopfield, “Neural networks and physical systems with emergent collective computational
abilities”, Proc. Natl. Acad. Sci. USA Vol 79, pp. 2554-2558, April 1982
March 32st, 2011
Neural Network Algorithms
20
Additional Slides
Existing NNs on
Quantum Computers

Quantum Computing
1.
2.
3.
4.
N-qubit register can contain all 2N values
at once
You can have a quantum ‘circuit’
‘computing’ on all of them at once
But when you ‘touch it’, you will get one
value only.
Goal – how to touch it, to get the value
you want, with high probability
March 32st, 2011
Neural Network Algorithms
22