US 11,928,602 B2
Systems and methods to enable continual, memory-bounded learning in artificial intelligence and deep learning continuously operating applications across networked compute edges
Matthew Luciw, Boston, MA (US); Santiago Olivera, Brookline, MA (US); Anatoly Gorshechnikov, Newton, MA (US); Jeremy Wurbs, Worcester, MA (US); Heather Marie Ames, Milton, MA (US); and Massimiliano Versace, Milton, MA (US)
Assigned to Neurala, Inc., Boston, MA (US)
Filed by Neurala, Inc., Boston, MA (US)
Filed on May 9, 2018, as Appl. No. 15/975,280.
Claims priority of provisional application 62/612,529, filed on Dec. 31, 2017.
Claims priority of provisional application 62/503,639, filed on May 9, 2017.
Prior Publication US 2018/0330238 A1, Nov. 15, 2018
Int. Cl. G06N 3/08 (2023.01); G06F 18/23211 (2023.01); G06F 18/40 (2023.01); G06N 3/04 (2023.01); G06N 3/044 (2023.01); G06N 3/045 (2023.01); G06N 3/084 (2023.01); G06V 10/20 (2022.01); G06V 10/44 (2022.01); G06V 10/764 (2022.01); G06V 10/82 (2022.01); G06V 10/94 (2022.01); G06V 20/17 (2022.01); G06V 20/13 (2022.01)
CPC G06N 3/084 (2013.01) [G06F 18/23211 (2023.01); G06F 18/40 (2023.01); G06N 3/0409 (2013.01); G06N 3/044 (2023.01); G06N 3/045 (2023.01); G06N 3/08 (2013.01); G06V 10/255 (2022.01); G06V 10/454 (2022.01); G06V 10/764 (2022.01); G06V 10/82 (2022.01); G06V 10/95 (2022.01); G06V 20/17 (2022.01); G06V 20/13 (2022.01)] 12 Claims
OG exemplary drawing
 
1. A method of implementing a Lifelong Learning Deep Neural Network (L-DNN) including a slow-learning subsystem and a fast-learning subsystem having previously determined weights in a real-time operating machine, the method comprising:
predicting, by the L-DNN, a first action for the real-time operating machine based on (i) an observation, by a sensor, of an environment of the real-time operating machine, (ii) weights of the fast-learning subsystem, and (iii) features extracted from the observation by the slow-learning subsystem;
determining, by the L-DNN, a mismatch between an expectation and a perception of the real-time operating machine based on the observation;
in response to the mismatch, triggering a fast learning mode by the L-DNN, the fast learning mode updating a weight vector of a corresponding category node of the fast-learning subsystem based on the observation or adding a category node with a weight vector based on the observation to the fast-learning subsystem without changing the previously determined weights of the slow-learning subsystem;
consolidating the weights of the fast-learning subsystem to restrict a memory footprint by the fast-learning subsystem and bound memory growth of the L-DNN to no faster than linear with a number of objects that the L-DNN is trained to recognize; and
applying the consolidated weights of the fast-learning subsystem to predict a second action for the real-time operating machine based on a subsequent observation, by the sensor, of the environment of the real-time operating machine.