Applications of an algorithm that mimics cortical processing

• US 7,430,546 B1
• Filed: 06/02/2004
• Issued: 09/30/2008
• Est. Priority Date: 06/07/2003
• Status: Expired due to Fees

• Current Assignee: Roland Erwin Suri
• Original Assignee: Roland Erwin Suri
• Inventors: Suri, Roland Erwin
• Primary Examiner(s): Hirl; Joseph P.
• Assistant Examiner(s): Fernandez Rivas; Omar F
• Application Number:
  US10/859,794
• Time in Patent Office: 1,581 Days
• Field of Search: None
• US Class Current: 706/16
• CPC Class Codes: G06N 3/08 Learning methods

Abstract

An information processing system having neuron-like signal processors that are interconnected by synapse-like processing junctions that simulates and extends capabilities of biological neural networks. The information processing systems uses integrate-and-fire neurons and Temporally Asymmetric Hebbian learning (spike timing-dependent learning) to adapt the synaptic strengths. The synaptic strengths of each neuron are guaranteed to become optimal during the course of learning either for estimating the parameters of a dynamic system (system identification) or for computing the first principal component. This neural network is well-suited for hardware implementations, since the learning rule for the synaptic strengths only requires computing either spike-time differences or correlations. Such hardware implementation may be used for predicting and recognizing audiovisual information or for improving cortical processing by a prosthetic device.

Claims

1. A system for information processing that progressively minimizes at least one error function by adapting a plurality of parameters, comprising:

a plurality of signal processing nodes that receive a plurality of spike-coded node input signals to produce at least one spike-coded node output signal;

a plurality of processing junctions disposed to interconnect said plurality of signal processing nodes to form a network, wherein each of said processing junctions converts a prejunction signal to a postjunction signal wherein the postjunction signal is calculated using one parameter out of said plurality of parameters;

a plurality of network input connections that receive a plurality of signals for processing in said network;

a plurality of network output connections to transmit at least one signal out of said network;

at least one delay element to delay at least one of the signals processed by the network such that at least one of the signals in the network is influenced by at least two signal values obtained at different points in time;

at least one learning rule operable to adapt the values of the parameters of the at least one signal processing junction wherein (i) said learning rule is operable to use at least one node output signal containing more than one spike, and (ii) the change of the parameter for the at least one signal processing junction is calculated using an approximation to the correlation between a prejunction signal representation and a node output signal representation of the node receiving the output signal of said junction, said correlation being adjusted by a learning rate defined as a scalar for which the optimal value depends on the firing rates of each prejunction signal processing node and each postjunction signal processing node.

2. The system of claim 1 wherein at least one of the nodes is operable to perform a computational task selected from the group consisting of learning a principal component, learning an independent component, adaptive Kalman filtering, learning the regression slope, or system identification.

3. The system of claim 1 wherein at least one input signal of said network represent at least one source of information selected from the group consisting of visual information, audio-visual information, written information, auditory information, or electric signals from nervous tissue.

4. The system of claim 1 wherein at least one output signal of said network is used to induce at least one electric signal in nervous tissue.

5. The system of claim 1 wherein for a plurality of parameters used by a plurality of junctions said parameter adaptation minimizes a linear combination of two error terms wherein the first error term defines the optimal parameter values for system identification and the second error term defines the optimal parameters values for learning a principal component.

6. The system of claim 1 wherein the parameter of at least one junction that receives a node output signal from a first node and that provides its junction output signal to a second node is not equal to the parameter of a junction receiving its input signal from said second node and providing its output signal to said first node.

7. The system of claim 1 wherein said network includes at least one element selected from an electronic device, an optical device, a biological element or a chemical element.

8. The system of claim 2 wherein said network includes at least one element that is not implemented in software but instead selected from an electronic device, an optical device, a biological element, or a chemical material.

9. The system of claim 3 wherein said network includes at least one element that is not implemented in software but instead selected from an electronic device, an optical device, a biological element, or a chemical material.

10. The system of claim 5 wherein said network includes at least one element that is not implemented in software but instead selected from an electronic device, an optical device, a biological element, or a chemical material.

11. The system of claim 6 wherein said network includes at least one element that is not implemented in software but instead selected from an electronic device, an optical device, a biological element, or a chemical material.

12. The system of claim 1 further comprising:

at least one sensor element, producing at least one sensor output signal, and operable to detect signals in a first location of nervous tissue of an animal;

at least one connection element operable to transmit information represented by said sensor output signal to said network input signal;

at least one effector element operable to receive at least one effector input signal and to induce at least one electrical signal in either said first location or in a second location of nervous tissue of said animal; and

at least one connection element operable to transmit information represented by at least one of the network output signals to said effector element,

whereby the network enhances the memory and information processing capabilities of said animal.

13. A method for information processing, comprising:

providing a plurality of nodes wherein each of said nodes computes at least one node output signal that is influenced by a plurality of node input signals wherein at least one of said node output signals comprises at least two characteristic temporal patterns as a means to label by each of said temporal patterns a time point in said node output signal;

providing a plurality of junction elements disposed to interconnect said plurality of nodes to form a network, (i) wherein each of said junction elements uses one parameter out of a plurality of parameters to compute a postjunction signal from a prejunction signal, and (ii) wherein the said prejunction signals contain a multitude of temporal patterns as a means to label by each of said temporal pattern a time point;

providing at least one network output connection to that transmit at least one signal out of said network;

providing at least one network input connection to influence at least one signal in said network;

delaying at least one of the signals processed by one of the elements of the network such that at least one of the network signals is influenced by at least two signal values of network signals that said network signals take at two different points in time;

providing processing elements for adapting the values of a non-empty subset of said parameters according to at least one learning rule wherein said learning rule is operable to use at least one node output signal containing more than one of said temporal patterns; and

adapting at least one of said parameters according to said learning rule wherein the change of the parameter is calculated using an approximation to the correlation between a representation of its prejunction signal and a representation of the node output signal of the node receiving the output signal of said junction, said correlation being adjusted by a learning rate defined as a scalar for which the optimal value depends on the firing rates of each prejunction signal processing node and each postjunction signal processing node.

14. The method of claim 13 wherein, for at least one node, a subset of the junction parameters is adapted such that an error function is progressively minimized.

15. The method of claim 14, wherein, for a plurality of junctions, the adaptation of the parameters value due to said learning rule is calculated using the gradient of the error function with respect to said parameter.

16. The method of claim 15, wherein the parameter of at least one junction that receives a node output signal from a first node and that provides its junction output signal to a second node is not equal to the parameter of a junction receiving its input signal from said second node and providing its output signal to said first node.

17. The method of claim 16 wherein said network includes at least one element that is not implemented in software but instead selected from an electronic device, an optical device, a biological substrate, or a chemical material.

18. The method of claim 17, wherein for a plurality of parameters used by a plurality of junctions said parameter adaptation minimizes a linear combination of two error terms wherein the first error term defines the optimal parameter values for system identification and the second error term defines the optimal parameters values for learning a principal component.

19. The method of claim 18, wherein, for a plurality of said nodes, the average value of the covariance between two different prejunction signals, expressed in the representation used by said error function, is at least six times smaller than the average value of the covariance of the prejunction signals with themselves.

20. A system for information processing that progressively minimizes at least one error function by adapting a plurality of parameters, comprising:

a plurality of signal processing nodes that respond to a plurality of spike-coded node input signals with at least one spike-coded node output signal;

a plurality of processing junctions disposed to interconnect said plurality of signal processing nodes to form a network, wherein each of said processing junctions converts a prejunction signal to a postjunction signal wherein the postjunction signal is calculated using one parameter out of said plurality of parameters;

a plurality of network input connections to influence at least one signal in said network;

a plurality of network output connections to transmit at least one signal out of said network;

at least one delay element that delays at least one of the signals processed by the network such that at least one of the signals in the network is influenced by at least two signal values that network signals take at different points in time;

at least one learning rule to change the values of a non-empty subset of the parameters wherein (i) said learning rule is operable to use at least one output signal containing more than one spike, (ii) said parameter adaptation minimizes a linear combination of two error terms wherein the first error term defines the optimal parameters values for system identification and the second error term defines the optimal parameters values for learning a principal component, and (iii) said change of a parameter is calculated using an approximation to the correlation between a representation of its prejunction signal and a representation of the node output signal of the node receiving the output signal of said junction, said correlation being adjusted by a learning rate defined as a scalar for which the optimal value depends on the firing rates of each prejunction signal processing node and each postjunction signal processing node.