Hierarchical temporal memory system with enhanced inference capability

• US 8,219,507 B2
• Filed: 06/26/2008
• Issued: 07/10/2012
• Est. Priority Date: 06/29/2007
• Status: Active Grant

• Current Assignee: Numenta, Inc.
• Original Assignee: Numenta, Inc.
• Inventors: Jaros, Robert G., George, Dileep
• Primary Examiner(s): Gaffin, Jeffrey A
• Assistant Examiner(s): BUSS, BENJAMIN J
• Application Number:
  US12/147,348
• Publication Number:
  US 20090006289A1
• Time in Patent Office: 1,475 Days
• Field of Search: 706/12, 706/52
• US Class Current: 706/12
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 3/049 Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
• CPC Class Codes: G06N 3/10 Interfaces, programming languages or software development kits, e.g. for simulating neural networks

Abstract

A node, a computer program storage medium, and a method for a hierarchical temporal memory (HTM) network where at least one of its nodes generates a top-down message and sends the top-down message to one or more children nodes in the HTM network. The first top-down message represents information about the state of a node and functions as feedback information from a current node to its child node. The node may also maintain history of the input patterns or co-occurrences so that temporal relationships between input patterns or co-occurrences may be taken into account in an inference stage. By providing the top-town message and maintaining history of previous input patterns, the HTM network may, among others, (i) perform more accurate inference based on temporal history, (ii) make predictions, (iii) discriminate between spatial co-occurrences with different temporal histories, (iv) detect “surprising” temporal patterns, (v) generate examples from a category, and (vi) fill in missing or occluded data.

Claims

1. A node in a computer-implemented hierarchical temporal memory network, the node associated with a child node, the node comprising:

a spatial pooler adapted to generate and store information about spatial co-occurrences of first input patterns received at the node based on spatial similarity of the first input patterns in a learning stage, the spatial pooler generating an output representing information about second input patterns received in an inference stage subsequent to the learning stage corresponding to the spatial co-occurrences based on the spatial co-occurrences information, at least part of the first input patterns and at least part of the second input patterns received from the child node;

a temporal pooler adapted to generate and store temporal statistics data indicating whether the second input patterns include a spatial co-occurrence based on spatial co-occurrences preceding the spatial co-occurrence in the second input patterns and temporal relationships of the spatial co-occurrences learned in the learning stage; and

a top-down message generator adapted to generate a first top-down message representing information about a state of the node in the inference stage based on the spatial co-occurrences information and the temporal statistics data, the top-down message generator adapted to send the first top-down message to the child node.

2. The node of claim 1, wherein the top-down message generator is adapted to generate the first top-down message further based on history of a previous state of the node in the inference stage.

3. The node of claim 1, wherein the top-down message generator is adapted to generate the first top-down message further based on a second top-down message received from a parent node associated with the node, the second top-down message representing a state of the parent node in the inference stage.

4. The node of claim 1, wherein the top-down message generator is adapted to generate the first top-down message by multiplying a downward dynamic variable, the output of the spatial pooler in the inference stage, and the temporal statistics data, the downward dynamic variable representing history of a state of the node.

5. The node of claim 4, wherein the temporal statistics data is computed based on whether a preceding spatial co-occurrence in the second input patterns is indicative of a transition between temporal groupings of the spatial co-occurrences.

6. The node of claim 1, wherein the spatial pooler is adapted to generate a first output for a third input pattern responsive to previously having received a same third input pattern, and generate a second output for the third input pattern responsive to previously having received a fourth input pattern different than the third input pattern, the first output different than the second output.

7. The node of claim 1, further comprising a bottom-up message generator adapted to generate a bottom-up message representing information that the second input patterns received at the node in the inference stage correspond to the learned spatial co-occurrences based on the second input patterns, the spatial co-occurrences information, and history of a state of the node.

8. The node of claim 1, wherein the temporal pooler is further adapted to:

replicate a spatial co-occurrence shared by two or more temporal groupings of the spatial co-occurrences to represent a higher-order temporal model of first input patterns; and

modify the temporal statistics data according to the replicated spatial co-occurrence.

9. A node in a computer-implemented hierarchical temporal memory network between a parent node and a child node, comprising:

a spatial pooler adapted to generate and store information about spatial co-occurrences of first input patterns received at the node based on spatial similarity of the first input patterns in a learning stage, the spatial pooler generating an output representing information about second input patterns received in an inference stage subsequent to the learning stage corresponding to the spatial co-occurrences based on the spatial co-occurrences information, at least part of the first input patterns and at least part of the second input patterns received from the child node;

a temporal pooler adapted to generate and store temporal statistics data indicating whether the second input patterns include a spatial co-occurrence based on spatial co-occurrences preceding the spatial co-occurrence in the second input patterns and temporal relationships of the spatial co-occurrences learned in the learning stage; and

a bottom-up message generator adapted to generate a first bottom-up message representing information that the second input patterns received at the node in the inference stage correspond to the learned spatial co-occurrences based on the temporal statistics data and the spatial co-occurrences information, the bottom-up message generator adapted to send the first bottom-up message to the parent node.

10. The node of claim 9, wherein the bottom-up message generator is further adapted to generate the first bottom-up message based further on history of the state of the node in the inference stage.

11. The node of claim 9, wherein the temporal statistics data is computed based on whether a preceding spatial co-occurrence in the second input patterns is indicative of a transition between groupings of the spatial co-occurrences generated by the temporal pooler.

12. The node of claim 9, wherein the spatial pooler is adapted to generate a first output for a third input pattern responsive to previously having received a same third input pattern, and generate a second output for the third input pattern responsive to previously having received a fourth input pattern different than the third input pattern, the first output different than the second output.

13. The node of claim 9, wherein the temporal pooler is further adapted to:

replicate a spatial co-occurrence shared by two or more temporal groupings of the spatial co-occurrences to represent a higher-order temporal model of first input patterns; and

modify the temporal statistics data according to the replicated spatial co-occurrence.

14. A non-transitory computer-readable storage medium storing computer instructions adapted to operate a node in a hierarchical temporal memory network on a computer, the node associated with a child node in the hierarchical temporal memory network, the computer instructions when executed cause a processor in the computer to:

generate and store information about spatial co-occurrences of first input patterns received at the node based on spatial similarity of the first input patterns in a learning stage, at least part of the first input patterns received from the child node;

generate information about second input patterns received in an inference stage subsequent to the learning stage corresponding to the spatial co-occurrences based on the spatial co-occurrences information, at least part of the second input patterns received from the child node;

generate and store temporal statistics data indicating whether the second input patterns include a spatial co-occurrence based on spatial co-occurrences preceding the spatial co-occurrence in the second input patterns and temporal relationships of the spatial co-occurrences learned in the learning stage;

generate a first top-down message representing information about a state of the node in the inference stage based on the spatial co-occurrences information and the temporal statistics data; and

send the first top-down message to the child node.

15. The computer-readable storage medium of claim 14, wherein the first top-down message is further based on history of a previous state of the node in the inference stage.

16. The computer-readable storage medium of claim 14, wherein the first top-down message is further based on a second top-down message received from a parent node associated with the node, the second top-down message representing a state of the parent node in the inference stage.

17. The computer-readable storage medium of claim 14, further comprising computer instructions to generate the first top-down message by multiplying a downward dynamic variable, the output and the temporal statistics data, the downward dynamic variable representing history of a state of the node.

18. The computer-readable storage medium of claim 17, further comprising computer instructions to compute the temporal statistics data based on based on whether a preceding spatial co-occurrence in the second input patterns is indicative of a transition between temporal groupings of the spatial co-occurrences.

19. The computer-readable storage medium of claim 15, further comprising computer instructions to:

generate a first output for a third input pattern responsive to previously having received a same third input pattern; and

generate a second output for the third input pattern responsive to previously having received a fourth input pattern different than the third input pattern, the first output different than the second output.

20. The computer-readable storage medium of claim 15, further comprising computer instructions to generate a bottom-up message representing information that the second input patterns received at the node in the inference stage correspond to the learned spatial co-occurrences based on the second input patterns, the spatial co-occurrences information, and history of a state of the node.

21. The computer-readable storage medium of claim 15, further comprising computer instructions to:

replicate a spatial co-occurrence shared by two or more temporal groupings of the spatial co-occurrences to represent a higher-order temporal model of input patterns; and

modify the temporal statistics data according to the replicated spatial co-occurrence.

22. A non-transitory computer-readable storage medium storing a computer program product including computer instructions adapted to instantiate a node of a hierarchical temporal memory network on a computer, the node between a parent node and a child node in the hierarchical temporal memory network, the computer instructions when executed cause a processor in the computer to:

generate and store information about spatial co-occurrences of first input patterns received at the node based on spatial similarity of the first input patterns in a learning stage, at least part of the first input patterns received from the child node;

generating information about second input patterns received in an inference stage subsequent to the learning stage corresponding to the spatial co-occurrences based on the spatial co-occurrences information, at least part of the second input patterns received from the child node;

generate and store temporal statistics data indicating whether the second input patterns include a spatial co-occurrence based on spatial co-occurrences preceding the spatial co-occurrence in the second input patterns and temporal relationships of the spatial co-occurrences learned in the learning stage;

generate a first bottom-up message representing information that the second input patterns received at the node in the inference stage correspond to the learned spatial co-occurrences based on the temporal statistics data and the spatial co-occurrences information; and

send the bottom-up message to the parent node.

23. The computer-readable storage medium of claim 22, further comprising computer instructions to generate the first-bottom-up message based further on history of the state of the node in the inference stage.

24. The computer-readable storage medium of claim 22, wherein the temporal statistics data is computed based on whether a preceding spatial co-occurrence in the second input patterns is indicative of a transition between groupings of the spatial co-occurrences.

25. The computer-readable storage medium of claim 22, further comprising computer instructions to:

generate a first output for a third input pattern responsive to previously having received a same third input pattern; and

generate a second output for the third input pattern responsive to previously having received a fourth input pattern different than the third input pattern, the first output different than the second output.

26. The computer-readable storage medium of claim 22, further comprising computer instructions to:

replicate a spatial co-occurrence shared by two or more temporal groupings of the spatial co-occurrences to represent a higher-order temporal model of first input patterns; and

modify the temporal statistics data according to the replicated spatial co-occurrence.

27. A method for operating a hierarchical temporal memory network on a computer, comprising:

generating information about spatial co-occurrences of first input patterns received at a node based on spatial similarity of the first input patterns in a learning stage, at least part of the first input patterns received from a child node;

generating information about second input patterns received in an inference stage subsequent to the learning stage corresponding to the spatial co-occurrences based on the spatial co-occurrences information, at least part of the second input patterns received from the child node;

by a processor, generating temporal statistics data indicating whether the second input patterns include a spatial co-occurrence based on spatial co-occurrences preceding the spatial co-occurrence in the second input patterns and temporal relationships of the spatial co-occurrences learned in the learning stage;

generating a first top-down message representing information about a state of the node in the inference stage based on the spatial co-occurrences information and the temporal statistics data; and

sending the first top-down message to the child node.

28. A method for operating a hierarchical temporal memory, comprising:

generating information about spatial co-occurrences of first input patterns received at the node based on spatial similarity of the first input patterns in a learning stage, at least part of the first input patterns received from a child node;

generating information about second input patterns received in an inference stage subsequent to the learning stage corresponding to the spatial co-occurrences based on the spatial co-occurrences information, at least part of the second input patterns received from the child node;

by a processor, generating temporal statistics data indicating whether the second input patterns include a spatial co-occurrence based on spatial co-occurrences preceding the spatial co-occurrence in the second input patterns and temporal relationships of the spatial co-occurrences learned in the learning stage;

generate a first bottom-up message representing information that the second input patterns received at the node in the inference stage correspond to the learned spatial co-occurrences based on the temporal statistics data and the spatial co-occurrences information; and

send the bottom-up message to a parent node.