IP Library › Granted Patent US 12,406,172
Granted Patent B2
US 12,406,172 · App. 15/691,103 · Granted Sep 2, 2025

Computational method for feedback in a hierarchical neural network

Inventors: Ahmet S. Ozcan (San Jose, CA); J. Campbell Scott (San Jose, CA)
Assignee: INTERNATIONAL BUSINESS MACHINES CORPORATION
G06N3/049G06N3/045
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12,406,172
App. No.
15/691,103
Granted
Sep 2, 2025
Kind
B2
Abstract

Feedback within hierarchical neural networks such as hierarchical temporal memories is provided. In various embodiments, a plurality of input streams is received at a first region of a hierarchical artificial neural network. The plurality of input streams is pooled in the first region. One or more recurrent features are identified in the pooled input streams in the first region. A temporal pattern is recognized in the one or more recurrent features in the first region. An indication of the temporal pattern is provided to at least a second region of the hierarchical neural network. A response to a further input is modulated based on feedback from the second region and at least one additional region.

Claims (67)

1. A method, comprising:

identifying one or more recurrent features in an input stream in a first region of a hierarchical artificial neural network having a hierarchy;

recognizing a temporal pattern in the one or more recurrent features in the first region;

providing to at least a second region of the hierarchical neural network an indication of the temporal pattern;

based on the indication of the temporal pattern, forming a stable representation within the second region corresponding to a sequence within the input stream;

generating feedback in the second region based on the stable representation within the second region; and

modulating a response to a further input to the first region based on the feedback from the second region and at least one additional region, wherein the first region is directly below the second region within the hierarchy, the modulation comprising causing a neuron in the first region to fire irrespective of the further input, thereby sustaining the sequence corresponding to the stable representation despite noise in the input stream.

2. The method of claim 1 , wherein the input stream is a pooled input stream, the method further comprising:

receiving a plurality of input streams at the first region; and

pooling the plurality of input streams to form the input stream in the first region.

3. The method of claim 1 , wherein the stable representation is based on prior input to the hierarchical artificial neural network.

4. The method of claim 1 , further comprising:

providing to at least a third region of the hierarchical neural network an indication of the one or more recurrent features.

5. The method of claim 1 , wherein said providing the indication of the temporal pattern comprises activating one or more neurons in the second region.

6. The method of claim 1 , wherein said recognizing the temporal pattern comprises applying a trained sequence memory.

7. The method of claim 6 , further comprising training the sequence memory.

8. The method of claim 1 , wherein the first region is located below the second region and the least one additional region within the hierarchy of the hierarchical artificial neural network.

9. The method of claim 1 , wherein said modulating the response comprises sustaining the activity of one or more neurons in the first region.

10. A system, comprising:

a computing node comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a processor of the computing node to cause the processor to perform a method comprising:

receiving a plurality of input streams at a first region of a hierarchical artificial neural network having a hierarchy;

pooling the plurality of input streams in the first region;

identifying one or more recurrent features in the pooled input streams in the first region;

recognizing a temporal pattern in the one or more recurrent features in the first region;

providing to at least a second region of the hierarchical neural network an indication of the temporal pattern;

based on the indication of the temporal pattern, forming a stable representation within the second region corresponding to a sequence within the input stream;

generating feedback in the second region based on the stable representation within the second region; and

modulating a response to a further input to the first region based on the feedback from the second region and at least one additional region, wherein the first region is directly below the second region of the hierarchy, the modulation comprising causing a neuron in the first region to fire irrespective of the further input, thereby sustaining the sequence corresponding to the stable representation despite noise in the input stream.

11. A computer program product for providing feedback within an artificial neural network, the computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a processor to cause the processor to perform a method comprising:

receiving a plurality of input streams at a first region of a hierarchical artificial neural network having a hierarchy;

pooling the plurality of input streams in the first region;

identifying one or more recurrent features in the pooled input streams in the first region;

recognizing a temporal pattern in the one or more recurrent features in the first region;

providing to at least a second region of the hierarchical neural network an indication of the temporal pattern;

based on the indication of the temporal pattern, forming a stable representation within the second region corresponding to a sequence within the input stream;

generating feedback in the second region based on the stable representation within the second region; and

modulating a response to a further input to the first region based on the feedback from the second region and at least one additional region, wherein the first region is directly below the second region within the hierarchy, the modulation comprising causing a neuron in the first region to fire irrespective of the further input, thereby sustaining the sequence corresponding to the stable representation despite noise in the input stream.

12. A hierarchical artificial neural network implemented via at least one processor, comprising:

a plurality of regions, each region comprising a pooler, a correlator, a sequence memory, and at least one connection to another of the plurality of regions, the hierarchical artificial neural network having a hierarchy and being configured to:

receive a plurality of input streams at a first region of the plurality of regions;

pool the plurality of input streams at the pooler of the first region;

identify one or more recurrent features in the pooled input stream at the correlator of the first region;

recognize a temporal pattern in the one or more recurrent features at the sequence memory of the first region;

provide to at least a second region of the plurality of regions an indication of the temporal pattern;

based on the indication of the temporal pattern, form a stable representation within the second region corresponding to a sequence within the input stream;

generate feedback in the second region based on the stable representation within the second region; and

modulate a response to a further input to the first region based on the feedback from the second region and at least one additional region of the plurality of regions, wherein the first region is directly below the second region within the hierarchy, the modulation comprising causing a neuron in the first region to fire irrespective of the further input, thereby sustaining the sequence corresponding to the stable representation despite noise in the input stream.

13. The hierarchical artificial neural network of claim 12 , wherein the stable representation is based on prior input to the hierarchical artificial neural network.

14. The hierarchical artificial neural network of claim 12 , wherein said providing the indication of the temporal pattern comprises activating one or more neurons in the second region.

15. The hierarchical artificial neural network of claim 12 , wherein said recognizing the temporal pattern comprises applying a trained sequence memory.

16. The hierarchical artificial neural network of claim 15 , the hierarchical artificial neural network being further configured to train the sequence memory.

17. The hierarchical artificial neural network of claim 12 , wherein the first region is located below the second region and the least one additional region within the hierarchy of the hierarchical artificial neural network.

18. The hierarchical artificial neural network of claim 12 , wherein said modulating the response comprises sustaining the activity of one or more neurons in the first region.

19. A method, comprising:

forming a stable representation within a second region of a hierarchical artificial neural network based on input to the hierarchical artificial neural network, the hierarchical artificial neural network having a hierarchy, and the stable representation corresponding to a sequence within the input;

identifying one or more recurrent features in an input stream in a first region of the hierarchical artificial neural network;

recognizing a temporal pattern in the one or more recurrent features in the first region;

providing to at least a second region of the hierarchical neural network an indication of the temporal pattern;

generating feedback in the second region based on the stable representation within the second region; and

modulating a response to a further input to the first region based on the feedback from the second region and at least one additional region based on the stable representation, wherein the first region is directly below the second region within the hierarchy, the modulation comprising causing a neuron in the first region to fire irrespective of the further input, thereby sustaining the sequence corresponding to the stable representation despite noise in the input stream.

20. The method of claim 19 , wherein the input stream is a pooled input stream, the method further comprising:

receiving a plurality of input streams at the first region; and

pooling the plurality of input streams to form the input stream in the first region.

21. The method of claim 19 , wherein said recognizing the temporal pattern comprises applying a trained sequence memory.

22. The method of claim 19 , wherein the first region is located below the second region and the least one additional region within the hierarchy of the hierarchical artificial neural network.

23. The method of claim 19 , wherein said modulating the response comprises sustaining the activity of one or more neurons in the first region.

24. The method of claim 1 , wherein the noise in the input stream comprises missing or erroneous data.

Assignments (1)
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Aug 30, 2017
From: OZCAN, AHMET S.; SCOTT, J. CAMPBELL
To: INTERNATIONAL BUSINESS MACHINES CORPORATION
Reel/Frame 043451/0778 →
Continuity (1)
Related Publication 20190065935A1 · Feb 28, 2019
References Cited (42)
US 7983998B2 · George · 2011 [cited by examiner]
US 8285667B2 · Jaros · 2012 [cited by examiner]
US 8775341B1 · Commons · 2014 [cited by examiner]
US 9390372B2 · Modha · 2016 [cited by applicant]
US 10395118B2 · Yu · 2019 [cited by examiner]
US 20060184462A1 · Hawkins · 2006 [cited by examiner]
US 20090006289A1 · Jaros et al. · 2009 [cited by applicant]
US 20090240639A1 · George et al. · 2009 [cited by applicant]
US 20110225108A1 · Hawkins et al. · 2011 [cited by applicant]
US 20130297539A1 · Piekniewski et al. · 2013 [cited by applicant]
US 20140089232A1 · Buibas et al. · 2014 [cited by applicant]
US 20150324655A1 · Chalasani · 2015 [cited by examiner]
CN 1212772A · 1999 [cited by applicant]
CN 103455843A · 2013 [cited by applicant]
CN 105095961A · 2015 [cited by applicant]
CN 111052150A · 2020 [cited by applicant]
EP 0877984B1 · 2000 [cited by applicant]
JP 2020532013A · 2020 [cited by applicant]
WO 2019042676A1 · 2019 [cited by applicant]
Kneller, et al., “Distal dendrite feedback in hierarchical temporal memory.” 2015 International Joint Conference on Neural Networks (IJCNN). IEEE, 2015 (Year: 2015). [cited by examiner]
George et al., “Towards a mathematical theory of cortical micro-circuits.” PLoS computational biology 5.10 (Year: 2009). [cited by examiner]
Wielgosza et al., “Noisy video classification with Spatial Pooler of Hierarchical Temporal Memory.” brain 3 (Year: 2016). [cited by examiner]
Kneller, et al., “Distal dendrite feedback in hierarchical temporal memory.” 2015 International Joint Conference on Neural Networks (IJCNN). IEEE (Year: 2015). [cited by examiner]
Byrne, Fergal. “Encoding reality: prediction-assisted cortical learning algorithm in hierarchical temporal memory.” arXiv preprint arXiv:1509.08255 (Year: 2015). [cited by examiner]
Hashmi et al. “Discovering Cortical Algorithms.” IJCCI (ICFC-ICNC) (Year: 2010). [cited by examiner]
Alonso et al., “Thalamus controls recurrent cortical dynamics,” Nat Neurosci, 18(12):1703-1704 (2015). [cited by applicant]
Chung et al., “Gated feedback recurrent neural networks,” (2015). [cited by applicant]
Fast Simulation with Multiple Multi-Layer Neural Networks, IP.com No. IPCOM000247603D, Sep. 19, 2016. [cited by applicant]
Felleman et al., “Distributed hierarchical processing in the primate cerebral cortex,” Cereb Cortex, 1(1):1-47 (1991). [cited by applicant]
Hawkins et al., “Biological and Machine Intelligence (BAMI),” Numenta, Inc., pp. 1-69, (2016). [cited by applicant]
Hebb: The Organization of Behavior, Wiley: New York; 1949. [cited by applicant]
Kawato et al., “A hierarchical neural-network model for control and learning of voluntary movement,” Biol Cybern, 57(3):169-185 (1987). [cited by applicant]
Machine Learning Algorithms for Smart Meter Diagnostics, IP.com No. IPCOM000242462D, Jul. 16, 2015. [cited by applicant]
Stollenga et al., “Deep networks with internal selective attention through feedback connections,” Adv Neur In, pp. 3545-3553 (2014). [cited by applicant]
George et al., “Invariant Pattern Recognition using Bayesian Inference on Hierarchical Sequence,” Retrieved from Internet, URL:http://www.cnbc.cmu.edu/cns/papers/DilJeffTechReport.pdf [retrieved on Oct. 29, 2018]. [cited by applicant]
Hawkins et al., “Hierarchical Temporal Memory Concepts, Theory, and Terminology,” Numenta Inc., pp. 1-20 (2006). [cited by applicant]
International Search Report and Written Opinion for International Application No. PCT/EP2018/070509 dated Oct. 15, 2018. [cited by applicant]
Yeap et al., “A Hierarchical Neural Network for Temoral Pattern Recognition,” IEEE International Symposium on Circuits and Systems,4:2967-2970 (1990). [cited by applicant]
Notice of Reasons for Refusal for JP Application No. 2020-511783 dated May 27, 2022. [cited by applicant]
English Translation of Notice of Reasons for Refusal for JP Application No. 2020-511783 dated Nov. 24, 2021. [cited by applicant]
Office Action for Chinese Patent Application No. 201880054348.5 dated Feb. 2, 2023. [cited by applicant]
Chinese Office Action for 201880054348.5 dated Sep. 23, 2023. [cited by applicant]