Multi-layer developmental network having in-place learning

• US 20080005048A1
• Filed: 03/27/2007
• Published: 01/03/2008
• Est. Priority Date: 03/27/2006
• Status: Active Grant

• Current Assignee: The Board of Trustees of Michigan State University (Michigan State University)
• Original Assignee: The Board of Trustees of Michigan State University (Michigan State University)
• Inventors: Weng, Juyang
• Granted Patent: US 7,711,663 B2
• Time in Patent Office: Days
• US Class Current: 706/14
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G06V 10/451 with interaction between the filter responses, e.g. cortical complex cells

Abstract

An in-place learning algorithm is provided for a multi-layer developmental network. The algorithm includes: defining a sample space as a plurality of cells fully connected to a common input; dividing the sample space into mutually non-overlapping regions, where each region is a represented by a neuron having a single feature vector; and estimating a feature vector of a given neuron by an amnesic average of an input vector weighted by a response of the given neuron, where amnesic is a recursive computation of the input vector weighted by the response such that the direction of the feature vector and the variance of signal in the region projected onto the feature vector are both recursively estimated with plasticity scheduling.

Claims

1. A multi-layer in-place learning developmental network, comprising:

an input layer having a plurality of artificial neurons;

multiple intermediate layers having a plurality of artificial neurons, each intermediate layer having artificial neurons that receive an input vector signal from neurons in previous layers;

an input vector signal from neurons in the same layer and an input vector signal from neurons in subsequent layers; and

an output layer having a plurality of artificial neurons, where each neuron in the network is responsible for adapting itself.

2. The in-place learning network of claim 1 wherein neurons in the network implement an in-place learning algorithm that adapts using a lateral inhibition mechanism and Hebbian learning.

3. The in-place learning network of claim 2 wherein the in-place learning algorithm estimates a feature vector of a given neuron by an amnesic average of input vectors weighted by corresponding responses (response-weighted inputs) of the given neuron, where amnesic is a recursive, incremental computation of response-weighted inputs, such that a direction of the feature vector and a variance of signals in the region projected to the feature vector are both recursively estimated with a plasticity scheduling.

4. The in-place learning network of claim 1 wherein neurons in the network are initialized upon creation using input from a preceding layer of the network.

5. The in-place learning network of claim 2 wherein the in-place learning algorithm computes a response for all neurons in a given layer of the network using a current input sample;

selects one or more neurons in the given layer having the largest response; and

updates the selected neurons using temporally scheduled plasticity.

6. The in-place learning network of claim 5 wherein the in-place learning algorithm computes a response in accordance with

7. The in-place learning network of claim 6 wherein the response goes through a sigmoidal function g of the neuron.

8. The in-place learning network of claim 1 wherein each artificial neuron in the output layer outputs a control signal to a motor.

9. The in-place learning network of claim 1 wherein each artificial neuron in the output layer correspond to a different class of a classifier, such that the neuron having largest output corresponds to class that has been classified by the network.

10. The in-place learning network of claim 1 wherein neurons in the network receive input signals from other decision areas of a system to which the network is part of.

11. An in-place learning algorithm for a multi-layer network, comprising:

defining a sample space as a plurality of cells fully connected to a common input;

dividing the sample space into mutually non-overlapping regions, where each region is represented by a neuron having a single feature vector; and

estimating a feature vector of a given neuron by an amnesic average of an input vector weighted by the response of the given neuron, where amnesic is a recursive computation of the input vector weighted by the response such that a direction of the feature vector and a variance of signals in the region projected onto the feature vector are both recursively estimated with plasticity scheduling.

12. The learning algorithm of claim 11 further comprises dividing the sample space into regions having a high concentration of a probability density of the input.

13. The learning algorithm of claim 11 wherein estimating a feature vector employs a lateral inhibition mechanism.

14. The learning algorithm of claim 11 wherein estimating a feature vector is based on Hebbian learning.

15. The learning algorithm of claim 11 wherein the sample space forms a layer in the multi-layer network and the input vector includes output from previous layers.

16. The learning algorithm of claim 11 wherein the sample space forms a layer in the multi-layer network and the input vector includes output from subsequent layers.

17. The learning algorithm of claim 11 further comprises supervising a given region using attention selection signals from other regions.

18. An in-place learning algorithm for a multi-layer network having a plurality of neurons organized into multiple layers, comprising:

computing a response for all neurons in a given layer of the network using a current input sample;

selecting one or more neurons in the given layer having the largest response;

updating the selected neurons using temporally scheduled plasticity.

19. The learning method of claim 18 further comprises initializing neurons in a given layer using input from preceding layers of the network prior to the step of computing a response.

20. The learning method of claim 18 further comprises retaining age and weight for all unselected neurons in the given layer.

21. The learning method of claim 18 further comprises repeating each step for each subsequent input sample.

22. The learning method of claim 18 further comprises performing each step for each layer of the network.

23. The learning method of claim 18 wherein computing a response further comprises estimating a feature vector of a given neuron by an amnesic average of an input vector weighted by corresponding response of the given neuron, where amnesic is a recursive computation of the input vector weighted by the response such that a direction of the feature vector and a variance of signals in the region projected onto the feature vector are both recursively estimated with plasticity scheduling.

24. The learning method of claim 18 selecting one or more neurons employs a lateral inhibition mechanism.

25. The learning method of claim 18 selecting one or more neurons further comprises sorting responses for all neurons in a given layer, selecting neurons having largest n responses and setting responses remaining neurons to zero.

26. The learning method of claim 18 wherein updating the selected neurons is based on Hebbian learning.

27. The learning method of claim 18 further comprises updating the selected neurons in accordance with

v_j^(t)=w₁v_j^(t−

1)+w₂z_jy(t),

where the scheduled plasticity is determined by its two age-dependent weights;

28. The learning method of claim 18 wherein updating the selected neurons further comprises incrementing age of selected neuron by one.

29. The learning method of claim 18 wherein updating the selected neurons further comprises updating neurons adjacent to the selected neurons.

30. The learning method of claim 18 wherein neurons in the given layer receive input signals from preceding layers, from the same layer, and from subsequent layers.