Feedback in group based hierarchical temporal memory system

• US 7,983,998 B2
• Filed: 03/21/2008
• Issued: 07/19/2011
• Est. Priority Date: 03/21/2008
• Status: Active Grant

• Current Assignee: Numenta, Inc.
• Original Assignee: Numenta, Inc.
• Inventors: George, Dileep, Jaros, Robert G.
• Primary Examiner(s): HOLMES, MICHAEL B
• Application Number:
  US12/053,204
• Publication Number:
  US 20090240639A1
• Time in Patent Office: 1,215 Days
• Field of Search: 706/12
• US Class Current: 706/12
• 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

Abstract

A Hierarchical Temporal Memory (HTM) network has at least first nodes and a second node at a higher level than the first nodes. The second node provides an inter-node feedback signal to the first nodes for grouping patterns and sequences (or co-occurrences) in input data received at the first nodes at the first nodes. The second node collects forward signals from the first nodes; and thus, the second node has information about the grouping of the patterns and sequences (or co-occurrences) at the first nodes. The second node provides inter-node feedback signals to the first nodes based on which the first nodes may perform the grouping of the patterns and sequences (or co-occurrences) at the first nodes. Also, a node in a Hierarchical Temporal Memory (HTM) network comprising a co-occurrence detector and a group learner coupled to the co-occurrence detector. The group learner provides an intra-node feedback signal to the co-occurrence detector including information on the grouping of the co-occurrences. The co-occurrence detector may select co-occurrences to be split, merged, retained or discarded based on the intra-node feedback signals.

Claims

1. An Hierarchical Temporal Memory (HTM) network comprising:

first nodes for receiving a training input data representing an object or a status of an object in a learning phase, the first nodes in the learning phase grouping patterns and sequences in the training input data and in an inference phase subsequent to the learning phase receiving sample input data generating first vectors representing information about patterns and sequences in the sample input data corresponding to the patterns and sequences grouped in the learning phase; and

a second node associated with the first nodes for receiving output signals from the first nodes for generating and outputting a second vector based on the first vectors in the inference phase, the second vector representing information about causes of the sample input data, the second node providing inter-node feedback signals to the first nodes for grouping of the training input data at the first nodes.

2. The HTM network of claim 1, wherein each of the first nodes comprises:

a co-occurrence detector receiving the training input data to identify co-occurrences from the patterns and sequences in the training input data; and

a group learner receiving information on the co-occurrences from the co-occurrence detector, the group learner grouping the co-occurrences based on temporal relationship between the co-occurrences.

3. The HTM network of claim 2, wherein the group learner comprises a time-adjacency table for tracking temporal relationships between the co-occurrences by scoring temporal adjacency between the co-occurrences, increments for scoring the temporal adjacency adjusted by the inter-node feedback signal.

4. The HTM network of claim 2, wherein the group learner modifies the grouping of the occurrences at the group learner based on the inter-node feedback signal.

5. The HTM network of claim 2, wherein the group learner generates and provides an intra-node feedback signal to the co-occurrence detector, the intra-node feedback signal including information for merging multiple co-occurrences into one co-occurrence or splitting a single co-occurrence into multiple co-occurrences at the co-occurrence detector.

6. The HTM network of claim 1, wherein the inter-node feedback signal includes information identifying grouping of patterns and sequences in at least one first node.

7. A node in a Hierarchical Temporal Memory (HTM) network, the node comprising:

a co-occurrence detector for identifying co-occurrences in patterns and sequences of training input data representing an object or a status of an object in a learning phase, the co-occurrence detector in an inference phase subsequent to the learning phase outputting information representing probabilities that patterns and sequences in sample input data correspond to the co-occurrences identified in the learning phase; and

a group learner for receiving and grouping the co-occurrences identified by the co-occurrence detector into groups based on temporal relationships between the co-occurrences, the group learner providing to the co-occurrence detector a first intra-node feedback signal indicating grouping of the co-occurrences.

8. The node of claim 7, wherein the co-occurrence detector selects multiple co-occurrences to be merged into one co-occurrence based on the first intra-node feedback signal.

9. The node of claim 7, wherein the co-occurrence detector selects a co-occurrence to be split into multiple co-occurrences based on the first inter-node feedback signal.

10. The node of claim 9, wherein the group learner merges the co-occurrences including the multiple co-occurrences split from the selected co-occurrence, the group learner providing a second intra-node feedback signal to the co-occurrence detector for selecting co-occurrences of the multiple co-occurrence to be retained for the inference phase.

11. A computer-implemented method of determining an object or a state of an object causing an input data:

first nodes generating first groups of first co-occurrences from patterns and sequences in training input data representing the object or the state of the object in a learning phase;

a second node in the learning phase generating second groups of second co-occurrences responsive to receiving output signals from the first nodes;

the second node in the learning phase generating and providing to the first nodes inter-node feedback signals for generating the first groups of the first co-occurrences at the first nodes;

the first nodes in an inference phase subsequent to the learning phase generating first vectors representing information about patterns and sequences in sample input data corresponding to the patterns and sequences grouped to the first groups;

the second node in the inference phase generating a second vector based on the first vectors, the second vector representing information about causes of the sample input data; and

storing the second vector generated by the second node.

12. The method of claim 11, wherein generating first groups of first co-occurrences comprises managing time-adjacency tables at the first nodes for tracking temporal relationships between the co-occurrences by scoring temporal adjacency between the co-occurrences, increments for scoring the temporal adjacency adjusted by the inter-node feedback signals.

13. The method of claim 11, wherein generating the first groups of the first co-occurrences comprises managing time-adjacency tables at the first nodes for tracking a temporal relationship between the co-occurrences, the first groups of the first occurrences modified by the inter-node feedback signals.

14. The method of claim 11, wherein the inter-node feedback signal includes information for grouping of patterns and sequences in at least one first node.

15. A method of managing co-occurrences in a node of a Hierarchical Temporal Memory (HTM) network, comprising:

identifying the co-occurrences in patterns and sequences of a training input data representing an object or a status of an object in a learning phase;

grouping the co-occurrences identified by the co-occurrence detector into groups based on temporal relationships between the co-occurrences in the learning phase;

outputting information representing probabilities that patterns and sequences in a sample input data correspond to the identified co-occurrences in an inference phase subsequent to the learning phase;

generating a vector representing information about patterns and sequences in the sample input data corresponding to the co-occurrences of the groups in the inference phase; and

generating a first intra-node feedback signal to the co-occurrence detector for grouping the co-occurrences in the learning phase.

16. The method of claim 15, further comprising selecting multiple co-occurrences to be merged into one co-occurrence based on the first intra-node feedback signal.

17. The method of claim 15, further comprising selecting a co-occurrence to be split into multiple co-occurrences based on the first intra-node feedback signal.

18. The method of claim 17, further comprising:

grouping the co-occurrences including the multiple co-occurrences split from the selected co-occurrence; and

generating a second intra-node feedback signal for selecting co-occurrences of the multiple co-occurrence to be retained for the inference phase.