Method and apparatus for detecting spatial patterns

• US 7,937,342 B2
• Filed: 11/27/2007
• Issued: 05/03/2011
• Est. Priority Date: 11/28/2006
• Status: Active Grant

• Current Assignee: Numenta, Inc.
• Original Assignee: Numenta, Inc.
• Inventors: Hawkins, Jeffrey C, Jaros, Robert Gilchrist, George, Dileep
• Primary Examiner(s): VINCENT, DAVID ROBERT
• Application Number:
  US11/945,919
• Publication Number:
  US 20080140593A1
• Time in Patent Office: 1,253 Days
• Field of Search: 706/20, 706 45- 48, 706/16, 706/25, 382155-161, 707/6, 707/706, 707/713, 707/723, 707/737
• US Class Current: 706/16
• CPC Class Codes: G06N 3/049 Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs

Abstract

An HTM node learns a plurality of groups of sensed input patterns over time based on the frequency of temporal adjacency of the input patterns. An HTM node receives a new sensed input, the HTM node assigns probabilities as to the likelihood that the new sensed input matches each of the plurality of learned groups. The HTM node then combines this probability distribution (may be normalized) with previous state information to assign probabilities as to the likelihood that the new sensed input is part of each of the learned groups of the HTM node. Then, as described above, the distribution over the set of groups learned by the HTM node is passed to a higher level node. This process is repeated at higher level nodes to infer a cause of the newly sensed input.

Claims

1. A computer-implemented method, comprising:

in a learning stage;

receiving a first input data at a first computing module in a hierarchy of computing modules, the first input data including a plurality of first spatial patterns corresponding to different time points;

storing temporal relationships between the plurality of first spatial patterns based on the corresponding time points of the first spatial patterns; and

assigning the first spatial patterns into a plurality of groups based on the temporal relationships of the first spatial patterns, a spatial pattern assigned to a group more likely to be temporally proximate within the first input data to another spatial pattern assigned to the same group compared to a spatial pattern assigned to another group;

in an inference stage subsequent to the learning stage;

detecting second spatial patterns in second input data at the first computing module responsive to receiving the second input data at the first computing module;

generating a first output at the first computing module based on a group to which each of the second spatial patterns was assigned in the learning stage; and

sending the first output to a second computing module higher in the hierarchy compared to the first computing module.

2. The method of claim 1, wherein the second computing module assigns third spatial patterns in the first output to groups based on time points corresponding to each of the third spatial patterns in a learning stage, the second computing module generating a second output in an inference stage responsive to receiving an output for inference from the first computing module.

3. The method of claim 1, further comprising determining a time adjacency value for each combination of two spatial patterns in the first spatial patterns, the time adjacency value increased responsive to detecting the two spatial patterns appearing within a threshold temporal distance from each other, wherein spatial patterns in the first spatial patterns are assigned to the groups based on time adjacency values.

4. The method of claim 3, wherein the temporal adjacency value is weighted by the temporal distance between the two spatial patterns in each combination.

5. The method of claim 3, further comprising:

(a) adding temporal adjacency values of each spatial pattern relative to other spatial patterns;

(b) assigning a spatial pattern with a highest added temporal adjacency value to a first group;

(c) assigning, to the first group, fewer than a predetermined number of spatial patterns having highest temporal adjacency values relative to the spatial pattern with the highest added temporally adjacency value;

(d) assigning, to a nth group, a spatial pattern with a nth highest added temporal adjacency value;

(e) assigning, to the nth group, fewer than the predetermined number of unassigned spatial patterns having highest temporal adjacency values relative to the spatial pattern with the nth highest added temporally adjacency value; and

(f) repeating (d) and (e) until a predetermined condition is met.

6. The method of claim 3, further comprising:

selecting spatial patterns for determining time adjacency values;

for each unselected spatial pattern, determining a selected spatial pattern that is closest to the unselected pattern; and

for each unselected spatial pattern, assigning a weight based on spatial similarity between the unselected spatial pattern and a selected spatial pattern that is most spatially similar to the unselected spatial pattern, wherein time adjacency values for the selected spatial patterns are determined based further on the assigned weight.

7. A computer system including a processor and memory, the computer system implementing a hierarchy of computing modules configured to:

in a learning stage;

receive a first input data at a first computing module in the hierarchy of computing modules, the first input data including a plurality of first spatial patterns corresponding to different time points;

store temporal relationships between the plurality of first spatial patterns based on the corresponding time points of the first spatial patterns; and

assign the first spatial patterns into a plurality of groups based on the temporal relationships of the first spatial patterns, a spatial pattern assigned to a group more likely to be temporally proximate within the first input data to another spatial pattern assigned to the same group compared to a spatial pattern assigned to another group;

in an inference stage subsequent to the learning stage;

detect second spatial patterns in second input data at the first computing module responsive to receiving the second input data at the first computing module;

generate a first output at the first computing module based on a group to which each of the second spatial patterns was assigned in the learning stage; and

send the first output to a second computing module higher in the hierarchy compared to the first computing module.

8. The computer system of claim 7, wherein the hierarchy of computing modules implements a hierarchical temporal network.

9. The computer system of claim 7, wherein the second computing module assigns third spatial patterns in the first output to groups based on time points corresponding to each of the third spatial patterns in a learning stage, the second computing module generating a second output in an inference stage responsive to receiving an output for inference from the first computing module.

10. A computer-readable storage medium structured to store instructions executable by a processor to implement a hierarchy of computing modules, the instructions, when executed cause the processor to:

in a learning stage;

receive a first input data at a first computing module in the hierarchy of computing modules, the first input data including a plurality of first spatial patterns corresponding to different time points;

store temporal relationships between the plurality of first spatial patterns based on the corresponding time points of the first spatial patterns; and

assign the first spatial patterns into a plurality of groups based on the temporal relationships of the first spatial patterns, a spatial pattern assigned to a group more likely to be temporally proximate within the first input data to another spatial pattern assigned to the same group compared to a spatial pattern assigned to another group;

in an inference stage subsequent to the learning stage;

detect second spatial patterns in second input data at the first computing module responsive to receiving the second input data at the first computing module;

generate a first output at the first computing module based on a group to which each of the second spatial patterns was assigned in the learning stage; and

send the first output to a second computing module higher in the hierarchy compared to the first computing module.

11. The medium of claim 10, further comprising instructions to:

determine a time adjacency value for each combination of two spatial patterns in the first spatial patterns, the time adjacency value increased responsive to detecting the two spatial patterns appearing within a threshold temporal distance from each other, wherein spatial patterns in the first spatial patterns are assigned to the groups based on time adjacency values.

12. The medium of claim 10, further comprising instructions to determine a time adjacency value for each combination of two spatial patterns in the first spatial patterns, the time adjacency value increased responsive to detecting the two spatial patterns appearing within a threshold temporal distance from each other, wherein spatial patterns in the first spatial patterns are assigned to the groups based on time adjacency values.

13. The medium of claim 12, wherein each temporal adjacency value is weighted by the temporal distance between the two spatial patterns in each combination.

14. The medium of claim 12, further comprising instructions to:

(a) add temporal adjacency values of each spatial pattern relative to all other spatial patterns;

(b) assign a spatial pattern with a highest added temporal adjacency value to a first group;

(c) assign, to the first group, fewer than a predetermined number of spatial patterns having highest temporal adjacency values relative to the spatial pattern with the highest added temporally adjacency value;

(d) assign a spatial pattern with a nth highest added temporal adjacency value to a nth group;

(e) assign, to the nth group, fewer than the predetermined number of unassigned spatial patterns having highest temporal adjacency values relative to the spatial pattern with the nth highest added temporally adjacency value; and

(f) repeat (d) and (e) until a predetermined condition is met.

15. The method of claim 12, further comprising instructions to:

select spatial patterns for determining the time adjacency value;

for each unselected spatial pattern, determine a selected spatial pattern that is closest to the unselected pattern; and

for each unselected spatial pattern, assign a weight based on spatial similarity between the unselected spatial pattern and a selected spatial pattern that is most spatially similar to the unselected spatial pattern, wherein time adjacency values for the selected spatial patterns are determined based further on the assigned weight.

16. The medium of claim 10, further comprising instructions to assign, by the second computing module, third spatial patterns in the first output to groups based on time points corresponding to each of the third spatial patterns in a learning stage, the second computing module generating a second output in an inference stage responsive to receiving an output for inference from the first computing module.