Performing multistep prediction using spatial and temporal memory system

• US 9,159,021 B2
• Filed: 10/23/2012
• Issued: 10/13/2015
• Est. Priority Date: 10/23/2012
• Status: Active Grant

• Current Assignee: Numenta, Inc.
• Original Assignee: Numenta, Inc.
• Inventors: Hawkins, Jeffrey C., Marianetti, Ronald
• Primary Examiner(s): Chen, Alan
• Assistant Examiner(s): BROWN JR, NATHAN H
• Application Number:
  US13/658,200
• Publication Number:
  US 20140114896A1
• Time in Patent Office: 1,085 Days
• Field of Search: 706/47, 706/50, 706/59
• US Class Current: 1/1
• 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 3/063 using electronic means

Abstract

Embodiments relate to making predictions for values or states to follow multiple time steps after receiving a certain input data in a spatial and temporal memory system. During a training stage, relationships between states of the spatial and temporal memory system at certain times and spatial patterns of the input data detected a plurality of time steps later after the certain time steps are established. Using the established relationships, the spatial and temporal memory system can make predictions multiple time steps into the future based on the input data received at a current time.

Claims

1. A computer-implemented method comprising:

receiving input data including a temporal sequence of spatial patterns at a predictive system, the input data in a distributed representation form;

spatially pooling the input data to generate sparse vectors in a sparse distributed representation form by a spatial pooler of the predictive system;

temporally processing transitions of the sparse vectors to establish relationships between temporal sequences of spatial patterns in the input data by a sequence processor of the predictive system;

detecting a state of the predictive system at a first time responsive to receiving the input data; and

generating a prediction for a second time following the first time after a plurality of time steps based on the detected state of the predictive system and stored relationships, the stored relationships mapping states of the predictive system at third times to spatial patterns derived from the input data at fourth times, each of the third times preceding a corresponding one of the fourth times by the plurality of time steps, the fourth times preceding the first time, wherein the states of the predictive system comprise states of the sequence processor.

2. The method of claim 1, wherein the prediction is distribution of likelihood of values or range of values in the input data at the second time.

3. The method of claim 2, further comprising recommending an action based on the values or the range of values and the distribution of likelihood.

4. The method of claim 1, wherein the input data is fed to the spatial pooler of the predictive system for pooling spatial patterns in the input data.

5. The method of claim 1, wherein temporally processing the transitions comprises storing relationships of activation states of columns of cells connected with other cells in other columns to receive activation states of the other cells.

6. The method of claim 5, wherein the relationships are stored as values in entries of a data structure, each of the entries associated with a cell in the sequence processor, wherein each of the stored values in the entries represents frequency or a number of occurrences in which an associated element in the spatial patterns was active the plurality of time steps after the associated cell was activated.

7. The method of claim 6, wherein generating the prediction comprises tallying stored values in the entries corresponding to cells activated at the first time.

8. The method of claim 1, further comprising decoding the prediction to determine the input data at the second time.

9. The method of claim 1, updating the stored relationships responsive to receiving the input data.

10. A predictive system comprising:

a spatial pooler configured to receive spatial patterns derived from input data in a distributed representation form and generate sparse vectors based on the received spatial patterns;

a sequence processor configured to temporally process transitions of the sparse vectors to establish relationships between temporal sequences of spatial patterns in the input data;

a multistep predictor configured to;

detect a state of the predictive system responsive to receiving a spatial pattern at a first time; and

generate a prediction for a second time following the first time after a plurality of time steps based on the detected state of the predictive system and stored relationships, the stored relationships mapping states of the sequence processor at third times to spatial patterns derived from the input data at fourth times, each of the third times preceding a corresponding one of the fourth times by the plurality of time steps, the fourth times preceding the first time.

11. The predictive system of claim 10, further comprising an encoder configured to generate the input data in the distributed representation form using an encoding scheme.

12. The predictive system of claim 10, wherein the prediction is distribution of likelihood of values or range of values in the input data at the second time.

13. The predictive system of claim 12, further comprising an action or value matrix configured to recommend an action based on the values or the range of values and the distribution of likelihood.

14. The predictive system of claim 10, wherein the sequence processor comprises a plurality of cells arranged in a plurality of columns, cells in a column having connections with other cells in other columns to predictively activate the cells in the column based on activation of the other cells in the other columns.

15. The predictive system of claim 14, wherein the multistep predictor is configured to store the relationships as values in entries of a data structure, each of the entries associated with a cell in the sequence processor, wherein each of the stored values in the entries represents frequency or a number of occurrences in which an associated element in the spatial patterns was active the plurality of time steps after the associated cell was activated.

16. The predictive system of claim 15, wherein the multistep predictor is configured to generate the prediction by tallying stored values in the entries corresponding to cells activated at the first time.

17. A non-transitory computer readable storage medium configured to store instructions, when executed by a processor cause the processor to:

receive input data including a temporal sequence of spatial patterns at a predictive system, the input data in a distributed representation form;

spatially pool the input data to generate sparse vectors in a sparse distributed representation form by a spatial pooler of the predictive system;

temporally process transitions of the sparse vectors to establish relationships between temporal sequences of spatial patterns in the input data by a sequence processor of the predictive system;

detect a state of the predictive system at a first time responsive to receiving the input data; and

generate a prediction for a second time following the first time after a plurality of time steps based on the detected state of the predictive system and stored relationships, the stored relationships mapping states of the predictive system at third times to spatial patterns derived from the input data at fourth times, each of the third times preceding a corresponding one of the fourth times by the plurality of time steps, the fourth times preceding the first time, wherein the states of the predictive system comprise states of the sequence processor.