HIERARCHICAL TEMPORAL MEMORY SYSTEM WITH HIGHER-ORDER TEMPORAL POOLING CAPABILITY

• US 20090313193A1
• Filed: 06/12/2009
• Published: 12/17/2009
• Est. Priority Date: 06/12/2008
• Status: Active Grant

• Current Assignee: Numenta, Inc.
• Original Assignee: Numenta, Inc.
• Inventors: Astier, Frank E., Hawkins, Jeffrey C., George, Dileep, Raj, Anosh, Jaros, Robert G., Curry, Charles
• Granted Patent: US 8,407,166 B2
• Time in Patent Office: Days
• US Class Current: 706/13
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/047 Probabilistic or stochastic networks
• CPC Class Codes: G06N 3/049 Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G06N 7/01 Probabilistic graphical models, e.g. probabilistic networks

Abstract

A temporal pooler for a Hierarchical Temporal Memory network is provided. The temporal pooler is capable of storing information about sequences of co-occurrences in a higher-order Markov chain by splitting a co-occurrence into a plurality of sub-occurrences. Each split sub-occurrence may be part of a distinct sequence of co-occurrences. The temporal pooler receives the probability of spatial co-occurrences in training patterns and tallies counts or frequency of transitions from one sub-occurrence to another sub-occurrence in a connectivity matrix. The connectivity matrix is then processed to generate temporal statistics data. The temporal statistics data is provided to an inference engine to perform inference or prediction on input patterns. By storing information related to a higher-order Markov model, the temporal statistics data more accurately reflects long temporal sequences of co-occurrences in the training patterns.

Claims

1. A node in a computer-implemented Hierarchical Temporal Memory network for generating an output representing an object or a state of an object corresponding to input patterns, the node comprising:

a temporal pooler configured to generate temporal statistics data representing a higher-order Markov chain of temporal sequences of the spatial co-occurrences based on temporal relationships of the spatial co-occurrences of the first input patterns in the training stage; and

an inference engine configured to generate an output based on the temporal statistics data responsive to receiving a sequence of spatial co-occurrence information about second input patterns in an inference stage subsequent to the training stage, the output representing probabilities of the sequence of the spatial co-occurrence information corresponding to temporal sequences of the temporal statistics data.

2. The node of claim 1, wherein the temporal pooler is further configured to split a spatial co-occurrence into a plurality of sub-occurrences, each sub-occurrence assigned to represent the spatial co-occurrence in a different temporal sequence.

3. The node of claim 2, wherein the temporal pooler is further configured to store temporal relationships of the sub-occurrences in a connectivity matrix, each entry in the connectivity matrix representing frequency or transition counts from a first sub-occurrence to a second sub-occurrence, the first sub-occurrence split from a first spatial co-occurrence, and the second sub-occurrence split from a second spatial co-occurrence.

4. The node of claim 2, wherein the temporal pooler is configured to generate the temporal statistics data by receiving a single-round of the sequence of the spatial co-occurrence information.

5. The node of claim 2, wherein the temporal pooler comprises:

a window buffer for storing the sequence of the spatial co-occurrence information for a predetermined time period;

a connectivity matrix updater configured to determine transition from a first sub-occurrence at a first time to a second sub-occurrence at a second time subsequent to the first time based on the spatial co-occurrence information at a third time subsequent to the second time received from the window buffer;

a connectivity matrix store configured to store counts or frequency of the transition from the first occurrence to the second sub-occurrence; and

a sequence grouper configured to group sequences of co-occurrences based on the connectivity matrix.

6. The node of claim 5, wherein the connectivity matrix updater is configured to determine the transition based on a modified connectivity matrix generated by adding values to entries of the connectivity matrix to represent added likelihood of transitions between sub-occurrences not appearing in the connectivity matrix.

7. The node of claim 5, wherein the connectivity matrix updater is configured to determine the transition based on an activity variable representing probability of receiving a co-occurrence at the first time and previous co-occurrences before the first time according to the temporal relationships represented by the connectivity matrix.

8. The node of claim 1, wherein the temporal pooler is configured to generate the temporal statistics data by receiving a plurality rounds of the sequence of the spatial co-occurrence information.

9. The node of claim 7, wherein the temporal pooler comprises:

a mapper configured to map spatial co-occurrences of the first input patterns at a second time to sub-occurrences based on spatial co-occurrences of the first input patterns at a first time preceding the second time;

a connectivity matrix counter for adding counts to a connectivity matrix, the connectivity matrix configured to store counts or frequency of the transition from a first co-occurrence or a first sub-occurrence to a second co-occurrence or a second sub-occurrence;

a splitter configured to expand the connectivity matrix counter by selectively splitting a co-occurrence in the connectivity matrix, the splitter configured to update the mapper based on the splitting of the co-occurrence; and

a sequence grouper configured to group sequences of co-occurrences based on the connectivity matrix.

10. A computer-implemented method of determining an object or a state of an object represented by input patterns in a node of a Hierarchical Temporal Memory network, comprising:

generating information about spatial co-occurrences of first input patterns received by the node based on spatial similarity of the first input patterns in a training stage,

generating temporal statistics data representing a higher-order Markov chain of temporal sequences of the spatial co-occurrences based on temporal relationships of the spatial co-occurrences in the training stage; and

generating an output based on the temporal statistics data responsive to receiving a sequence of spatial co-occurrence information about second input patterns, the output representing probabilities of the sequence of the spatial co-occurrence information corresponding to temporal sequences of the temporal statistics data in an inference stage subsequent to the training stage.

11. The method of claim 10, further comprising splitting a spatial co-occurrence into a plurality of sub-occurrences, each sub-occurrence assigned to represent the spatial co-occurrence in a different temporal sequence of the higher-order Markov chain.

12. The method of claim 11, further comprising storing temporal relationships of the sub-occurrences in a connectivity matrix, each entry in the connectivity matrix representing frequency or transition counts from a first sub-occurrence to a second sub-occurrence, the first sub-occurrence split from a first spatial co-occurrence, and the second sub-occurrence split from a second spatial co-occurrence.

13. The method of claim 11, further comprising generating the temporal statistics data by receiving a single round of the sequence of the spatial co-occurrence information.

14. The method of claim 11, further comprising:

storing the sequence of the spatial co-occurrence information for a predetermined time period;

determining transition from a first sub-occurrence at a first time to a second sub-occurrence at a second time subsequent to the first time based on the spatial co-occurrence information at a third time subsequent to the second time received from the window buffer;

storing configured to store counts or frequency of the transition from the first occurrence to the second sub-occurrence; and

grouping sequences of co-occurrences based on the connectivity matrix.

15. The method of claim 14, wherein determining the transition comprises:

generating a modified connectivity matrix by adding values to entries of the connectivity matrix to represent added likelihood of transitions between sub-occurrences not appearing in the connectivity matrix;

computing an activity variable representing probability of receiving a co-occurrence at the first time and previous co-occurrences before the first time according to the temporal relationships represented by the connectivity matrix; and

determining the transition based on the modified connectivity matrix and the activity variable.

16. The method of claim 10, wherein the temporal statistics data is generated by receiving a plurality rounds of probabilities of co-occurrences.

17. The method of claim 16, further comprising:

mapping spatial co-occurrences of the first input patterns at a second time to sub-occurrences based on spatial co-occurrences of the first input patterns at a first time preceding the second time;

adding counts to a connectivity matrix, the connectivity matrix configured to store counts or frequency of the transition from a first co-occurrence or a first sub-occurrence to a second co-occurrence or a second sub-occurrence;

expanding the connectivity matrix counter by selectively splitting a co-occurrence in the connectivity matrix, the splitter configured to update the mapper based on the splitting of the co-occurrence; and

grouping sequences of co-occurrences based on the connectivity matrix.

18. A computer program storage medium storing computer instructions to operate a Hierarchical Temporal Memory network on a computer, the computer instructions when executed cause a processor in the computer to:

generate information about spatial co-occurrences of first input patterns received by a node in the Hierarchical Temporal Memory system based on spatial similarity of the first input patterns in a training stage,

generate temporal statistics data representing a higher-order Markov chain of temporal sequences of the spatial co-occurrences based on temporal relationships of the spatial co-occurrences in the training stage; and

generate an output based on the temporal statistics data responsive to receiving a sequence of spatial co-occurrence information about second input patterns, the output representing probabilities of the sequence of the spatial co-occurrence information corresponding to temporal sequences of the temporal statistics data in an inference stage subsequent to the training stage.

19. The computer program storage medium of claim 18, further comprising instructions to split a spatial co-occurrence into a plurality of sub-occurrences, each sub-occurrence assigned to represent the spatial co-occurrence in a different temporal sequence of the higher-order Markov chain.

20. The computer program storage medium of claim 18, further comprising instructions to:

store the sequence of the spatial co-occurrence information for a predetermined time period;

determine transition from a first sub-occurrence at a first time to a second sub-occurrence at a second time subsequent to the first time based on the spatial co-occurrence information at a third time subsequent to the second time received from the window buffer;

store configured to store counts or frequency of the transition from the first occurrence to the second sub-occurrence; and

group sequences of co-occurrences based on the connectivity matrix.