The Random Neural Network: the model and some of its applications Erol Gelenbe 1,2 and Varol Kaptan...
-
Upload
eleanor-cunningham -
Category
Documents
-
view
228 -
download
1
Transcript of The Random Neural Network: the model and some of its applications Erol Gelenbe 1,2 and Varol Kaptan...
The Random Neural Network: the model and some of its applications
Erol Gelenbe1,2 and Varol Kaptan2
1 www.ee.imperial.ac.uk/gelenbeDenis Gabor Professor
Head of Intelligent Systems and Networks
2 Dept of Electrical and Electronic EngineeringImperial College
London SW7 2BT
Prof. Gelenbe is grateful to his PhD students who have worked or are working with him on random
neural networks and/or their applications
- Univ. of Paris: Andreas Stafylopatis, Jean-Michel Fourneau, Volkan Atalay, Myriam Mokhtari, Vassilada Koubi, Ferhan Pekergin, Ali Labed, Christine Hubert
- Duke University: Hakan Bakircioglu, Anoop Ghanwani, Yutao Feng, Chris Cramer, Yonghuan Cao, Hossam Abdelbaki, Taskin Kocak
- UCF: Rong Wang, Pu Su, Peixiang Liu, Will Washington, Esin Seref, Zhiguang Xu, Khaled Hussain, Ricardo Lent
- Imperial: Arturo Nunez, Varol Kaptan, Mike Gellman, Georgios Loukas, Yu Wang
Thank you to the agencies and companies who have generously supported RNN work over the last 15 yrs
- France (1989-97): ONERA, CNRS C3, Esprit Projects QMIPS, EPOCH and LYDIA
- USA (1993-): ONR, ARO, IBM, Sandoz, US Army Stricom, NAWCTSD, NSF, Schwartz Electro-Optics, Lucent
- UK (2003-): EPSRC, MoD, General Dynamics UK Ltd, EU FP6 for grant awards for the next three years, hopefully more ..
Outline
• Biological Inspiration for the RNN
• The RNN Model
• Some Applications– modeling biological neuronal systems– texture recognition and segmentation– image and video compression– multicast routing
Random Spiking Behaviour of Neurons
Work started as an individual basic research project, motivated by a critical look at modeling biological neurons, rather than using popular connectionist models
Biological characteristics of the model needed to include:- Action potential “Signals” in the form of spikes of fixed
amplitude- Modeling recurrent networks- Random delays between spikes- Conveying information along axons via variable spike rates- Store and fire behaviour of the soma (head of the neuron)- Reduction of neuronal potential after firing- Possibility of representing axonal delays between neurons
Random Spiking Behaviour of Neurons
Mathematical properties that we hoped for, but did not always expect to obtain, but which were obtained
-Existence and uniqueness of solution to the models -Closed form analytical solutions for large systems-Convergent learning for recurrent networks-Polynomial speed for recurrent gradient descent-Hebbian and reinforcement learning algorithms-Analytical annealing
We exploited the analogy with queuing networks, and this also opened a new chapter in queuing network theory now called “G-networks” where richer models were developed
Queuing Networks: Exploiting the Analogy
- Discrete state space, typically continuous time, stochastic models arising in studying populations, dams, production systems, communication networks ..
- Important theoretical foundation for computer systems performance analysis
- Open (external Arrivals and Departures), as in Telephony, or Closed (Finite Population) as in Compartment Models
- Systems comprised of Customers and Servers- Theory is over 100 years old and still very active .. e.g. - Big activity at Bell Labs, AT&T Labs, IBM Research- More than 100,000 papers on the subject ..
Queuing Network <-> Random Neural Network
- Both Open and Closed Systems- Systems comprised of Customers and Servers- Servers = Neurons- Customer .. Arriving to server will increase the queue
length by +1- Excitatory spike arriving to neuron will increase its soma’s
potential by +1- Service completion (neuron firing) at server (neuron) will
send out a customer (spike), and reduce queue length by 1- Inhibitory spike arriving to neuron will decrease its soma’s
potential by -1- Spikes (customers) leaving neuron i (server i) will move to
neuron j (server j) in a probabilistic manner
Mathematical Model: A “neural” network with n neurons
• Internal State of Neuron i at time t, is an Integer xi(t) > 0
• Network State at time t is a Vector
x(t) = (x1(t), … , xi(t), … , xk(t), … , xn(t))
• If xi(t)> 0, we say that Neuron i is excited and it may fire at t+ (in which case it will send out a spike)
• Also, if xi(t)> 0, the Neuron i will fire with probability r it in the interval [t,t+t]
• If xi(t)=0, the Neuron cannot fire at t+
When Neuron i fires:
- It sends a spike to some Neuron j, w.p. p ij
- Its internal state changes xi(t+) = xi(t) - 1
Mathematical Model: A “neural” network with n neurons
The arriving spike at Neuron j is an:
- Excitatory Spike w.p. pij+
- Inhibitory Spike w.p. pij -
- pij = pij+ + pij
- with nj=1 pij < 1 for all i=1,..,n
From Neuron i to Neuron j:
- Excitatory Weight or Rate is wij+ = ri pij
+
- Inhibitory Weight or Rate is wij- = ri pij
-
- Total Firing Rate is ri = nj=1 (wij
+ + wij–)
To Neuron i, from Outside the Network
- External Excitatory Spikes arrive at rate i
- External Inhibitory Spikes arrive at rate i
State Equations
where,)1()( 'form'product '' thehasit then
solution, stationary a have Equations K-C theIf
]0)(Pr[lim],)(Pr[lim)(
:
]]0)([1)[(),(
]]0)([1),())(,([]),(]0)([1),([),(
:Equations The
:,
, and
process, Markov space-state discrete a is 0 where])(Pr[),(
n
1i
,
iki
it
it
iii
ii
iiiiii
iiijiijjijiji
ij
iiii
jiijjiij
qqkp
txqandktxkp
Let
tkrtkp
tktkpdrtkpprtkptkprtkptkpdt
d
ekkekk
eekkeekk
}{x(t):tktxtkp
i
:Theorem
Kolmogorov -Chapman
10
j jijjii
j jijji
i prqr
prqq
ji
Firing Rate ofNeuron i
External ArrivalRate of ExcitatorySpikes
External ArrivalRate of InhibitorySpikes
ji
Probability thatNeuron i is excited
o
o
o
ox
oo
v
i
j jijjii
j jijji
i
q
q
q
qxyforxyxg
tsqqneuronsoutputtwowithRNN
anexiststhereanyForboundedand
continuousbeRgLet
qtheallifsolutionuniqueanhas
niprqr
prqq
equationslinearnonofsystemThe
v
11)(|)()(|sup
..,
,0.
]1,0[: 99) al.et (Gelenbe
.1
1,
95)er Schassberg-Gelenbe 93, (Gelenbe
]1,0[
Theorem
Theorem
Random Neural Network
• Neurons exchange Excitatory and Inhibitory Spikes (Signals)
• Inter-neuronal Weights are Replaced by Firing Rates• Neuron Excitation Probabilities obtained from Non-
Linear State Equations• Steady-State Probability is Product of Marginal
Probabilities• Separability of the Stationary Solution based on
Neuron Excitation Probabilities• Existence and Uniqueness of Solutions for Recurrent
Network• Learning Algorithms for Recurrent Network are O(n3)• Multiple Classes (1998) and Multiple Class Learning
(2002)
Some Applications
• Modeling Cortico-Thalamic Response …
• Supervised learning in RNN (Gradient Descent)
• Texture based Image Segmentation
• Image and Video Compression
• Multicast Routing
Cortico-Thalamic Response to Somato-Sensory Input(i.e. what does the rat think when you tweak her/his whisker?)
Input from the brain stem (PrV) and response at thalamus (VPM) and cortex (SI), reprinted from M.A.L. Nicollelis et al. “Reconstructing the engram: simultaneous, multiple site, many single neuron recordings”, Neuron vol. 18, 529-537, 1997
Simultaneous MultipleCell Recordings(Nicollelis et al., 1997)
Predictions of CalibratedRNN Mathematical Model(Gelenbe & Cramer ’98, ’99)
Comparing Measurements and Theory:Calibrated RNN Model and Cortico-Thalamic Oscillations
Extracting Abnormal Objects from MRI Images of the Brain
Extracting Tumors from MRI
T1 and T2 Images
Separating HealthyTissue from Tumor
Simulating and PlanningGamma Therapy &
Surgery
2) RNN based Adaptive Video Compression:Combining Motion Detection and RNN Still Image
Compression
RNN
Neural Still Image CompressionFind RNN R that Minimizes
|| R(I) - I ||Over a Training Set of Images {I }
3) Analytical Annealing with the RNN: Multicast Routing(Similar Results with the Traveling Salesman Problem)
• Finding an optimal many-to-many communications path in a network is equivalent to finding a Minimal Steiner Tree which is an NP-Complete problem
• The best heuristics are the Average Distance Heuristic (ADH) and the Minimal Spanning Tree (MSTH) for the network graph
• RNN Analytical Annealing improves the number of optimal solutions found by ADH and MST by approximately 10%
Selected ReferencesE. Gelenbe, ``Random neural networks with negative and positive signals and product
form solution,'' Neural Computation, vol. 1, no. 4, pp. 502-511, 1989.E. Gelenbe, ``Stability of the random neural network model,'‘Neural Computation, vol. 2,
no. 2, pp. 239-247, 1990.E. Gelenbe, A. Stafylopatis, and A. Likas, ``Associative memory operation of the random
network model,'' in {\it Proc. Int. Conf. Artificial Neural Networks}, Helsinki, pp. 307-312, 1991.
E. Gelenbe, F. Batty, ``Minimum cost graph covering with the random neural network,'' Computer Science and Operations Research, O. Balci (ed.), New York, Pergamon, pp. 139-147, 1992.
E. Gelenbe, ``Learning in the recurrent random neural network,'‘ Neural Computation, vol. 5, no. 1, pp. 154-164, 1993.
E. Gelenbe, V. Koubi, F. Pekergin, ``Dynamical random neural network approach to the traveling salesman problem,'' Proc. IEEE Symp. Syst., Man, Cybern., pp. 630-635, 1993.
A. Ghanwani, ``A qualitative comparison of neural network models applied to the vertex covering problem,'' Elektrik, vol. 2, no. 1, pp. 11-18, 1994.
E. Gelenbe, C. Cramer, M. Sungur, P. Gelenbe ``Traffic and video quality in adaptive neural compression'', Multimedia Systems, Vol. 4, pp. 357--369, 1996.
…
Selected References… C. Cramer, E. Gelenbe, H. Bakircioglu ``Low bit rate video compression with neural
networks and temporal subsampling,'' Proceedings of the IEEE, Vol. 84, No. 10, pp. 1529--1543, October 1996.
E. Gelenbe, T. Feng, K.R.R. Krishnan ``Neural network methods for volumetric magnetic resonance imaging of the human brain,'' Proceedings of the IEEE, Vol. 84, No. 10, pp. 1488--1496, October 1996.
E. Gelenbe, A. Ghanwani, V. Srinivasan, ``Improved neural heuristics for multicast routing,'' IEEE J. Selected Areas in Communications, vol. 15, no. 2, pp. 147-155, 1997.
E. Gelenbe, Z. H. Mao, and Y. D. Li, ``Function approximation with the random neural network,'' IEEE Trans. Neural Networks, vol. 10, no. 1, January 1999.
E. Gelenbe, J.M. Fourneau ``Random neural networks with multiple classes of signals,'' Neural Computation, vol. 11, pp. 721--731, 1999.
E. Gelenbe, E. Seref, and Z. Xu, ``Simulation with learning agents’’ Proceedings of the IEEE, 89 (2) pp.148-157 (2001)
E. Gelenbe, K. Hussain ``Learning in the multiple class random neural network’’, IEEE Trans. Neural Networks, vol. 13, no. 6, pp. 1257—1267, 2002.
E. Gelenbe, R. Lent and A. Nunez ``Self-aware networks and QoS ‘’, Proceedings of the IEEE, 92 ( 9) pp. 1479-1490 (2004)