Supervision based grouping of patterns in hierarchical temporal memory (HTM)
A HTM network that uses supervision signals such as indexes for correct categories of the input patterns to group the co-occurrences detected in the node. In the training mode, the supervised learning node receives the supervision signals in addition to the indexes or distributions from children nodes. The supervision signal is then used to assign the co-occurrences into groups. The groups include unique groups and nonunique groups. The co-occurrences in the unique group appear only when the input data represent certain category but not others. The nonunique groups include patterns that are shared by one or more categories. In an inference mode, the supervised learning node generates distributions over the groups created in the training mode. A top node of the HTM network generates an output based on the distributions generated by the supervised learning node.
1. An Hierarchical Temporal Memory (HTM) network system comprising a top node and a supervised learning node, the top node configured to generate an output representing an object or a state of an object responsive to receiving sample input data in an inference mode, comprising:
a supervised learning node configured to receive a training input data in a training mode and group patterns in the training input data responsive to receiving a supervision signal representing a correct category of the object or the state of the object for the training input data, and the supervised learning node in the inference mode subsequent to the training mode generating first information about patterns in sample input data corresponding to the patterns grouped in the training mode, the top node in the inference mode generating the output based on the first information generated at the supervised learning node.
2. The HTM network system of claim 1 , wherein the supervised learning node is further adapted to create a unique group including one or more patterns in the training input data indicated by the supervision signal as representing one category, and a nonunique group including one pattern in the training input data indicated by the supervision signal as representing two or more categories.
3. The HTM network system of claim 1 , wherein the supervised learning node is configured to receives the training input data for a first receptive field in the training mode and the sample input data for a second receptive field in the inference mode, the second receptive field smaller than the first receptive field.
4. The HTM network system of claim 3 , wherein the supervised learning node is configured to generating a correlation table in the training mode for a first pattern and a second pattern by comparing and scoring patterns surrounding the first pattern and the second pattern, the supervised learning node grouping the first pattern and the second pattern into a same group based on the number of matching patterns that surround the first pattern and the second pattern.
5. The HTM network system of claim 1 , wherein the supervised learning node is configured to generate the first information about the patterns in sample input data based on a number of times a pattern was detected in the training input data.
6. The HTM network system of claim 1 , wherein the supervised learning node is further adapted to create a unique group including one or more patterns representing one object or one state of the object but not others by assigning a pattern to the unique group responsive to the supervision signal indicating one category for the assigning pattern over a threshold percentage of times.
7. The HTM network system of claim 1 , wherein the supervised learning node further comprises a temporal group learner configured to group the patterns in the training input data based on temporal proximity of the patterns in the input data, the supervised learning node configured to generate second information about patterns in sample input data corresponding to the patterns grouped by the temporal group learner, the top node in the inference mode generating the output based on the second information.
8. A computer-implemented method of determining an object or a state of an object in a supervised learning node in a Hierarchical Temporal Memory (HTM) network system, the HTM network comprising a top node, comprising:
grouping patterns in training input data responsive to receiving training input data in a training mode and a supervision signal representing a correct category of the object or the state of the object for the training input data;
generating first information about patterns in sample input data corresponding to the patterns grouped in the training mode in an inference mode subsequent to the training mode; and
sending the first information to the top node in the inference mode for generating an output representing the object or the state of the object corresponding to the input data based on the first information.
9. The method of claim 8 , further comprising:
creating a unique group including one or more patterns in the training input data indicated by the supervision signal as representing one category; and
creating a nonunique group including one pattern in the training input data indicated by the supervision signal as representing two or more categories.
10. The method of claim 8 , wherein the supervised learning node is configured to receives the training input data for a first receptive field in the training mode and the sample input data for a second receptive field in the inference mode, the second receptive field smaller than the first receptive field.
11. The method of claim 10 , further comprising
generating a correlation table in the training mode for a first pattern and a second pattern by comparing and scoring patterns surrounding the first pattern and the second pattern; and
grouping the first pattern and the second pattern into a same group based on the number of matching patterns that surround the first pattern and the second pattern.
12. The method of claim 8 , further comprising generating the first information about the patterns in sample input data based on a number of times a pattern was detected in the training input data.
13. The method of claim 8 , further comprising creating a unique group including one or more patterns representing one object or one state of the object but not others by assigning a pattern to the unique group responsive to the supervision signal indicating one category for the assigning pattern over a threshold percentage of times.
14. The method of claim 8 , further comprising:
grouping the patterns in the training input data based on temporal proximity of the patterns in the input data;
generating second information about patterns in sample input data corresponding to the patterns grouped based on the temporal proximity; and
sending the second information to the top node for generating the output.
15. A computer-readable storage medium storing instructions for determining an object or a state of an object in a supervised learning node in a Hierarchical Temporal Memory (HTM) network system, the instructions when executed by a processor causes the processor to
group patterns in training input data responsive to receiving training input data in a training mode and a supervision signal representing a correct category of the object or the state of the object for the training input data;
generate first information about patterns in sample input data corresponding to the patterns grouped in the training mode in an inference mode subsequent to the training mode; and
send the first information to the top node in the inference mode for generating an output representing the object or the state of the object corresponding to the input data based on the first information.
16. The computer-readable storage medium of claim 15 , further comprising instructions that cause the processor to:
create a unique group including one or more patterns in the training input data indicated by the supervision signal as representing one category; and
create a nonunique group including one pattern in the training input data indicated by the supervision signal as representing two or more categories.
17. The computer-readable storage medium of claim 15 , wherein the supervised learning node is configured to receives the training input data for a first receptive field in the training mode and the sample input data for a second receptive field in the inference mode, the second receptive field smaller than the first receptive field.
18. The computer-readable storage medium of claim 17 , further comprising instructions that cause the processor to:
generate a correlation table in the training mode for a first pattern and a second pattern by comparing and scoring patterns surrounding the first pattern and the second pattern; and
group the first pattern and the second pattern into a same group based on the number of matching patterns that surround the first pattern and the second pattern.
19. The computer-readable storage medium of claim 15 , further comprising instructions that cause the processor to generate the first information about the patterns in sample input data based on a number of times a pattern was detected in the training input data.
20. The computer-readable storage medium of claim 15 , further comprising instructions causing the processor to create a unique group including one or more patterns representing one object or one state of the object but not others by assigning a pattern to the unique group responsive to the supervision signal indicating one category for the assigning pattern over a threshold percentage of times.
21. The computer-readable storage medium of claim 15 , further comprising instructions causing the processor to:
group the patterns in the training input data based on temporal proximity of the patterns in the input data;
generate second information about patterns in sample input data corresponding to the patterns grouped based on the temporal proximity; and
send the second information to the top node for generating the output.