Artificial neural network and method of training an artificial neural network with epigenetic neurogenesis

Patent metadata

Field	Value
Publication number	US-11113597-B2
Application number	US-201916561735-A
Country	US
Kind code	B2
Filing date	Sep 5, 2019
Priority date	Oct 23, 2018
Publication date	Sep 7, 2021
Grant date	Sep 7, 2021

Abstract

A method for retraining an artificial neural network trained on data from an old task includes training the artificial neural network on data from a new task different than the old task, calculating a drift, utilizing Sliced Wasserstein Distance, in activation distributions of a series of hidden layer nodes during the training of the artificial neural network with the new task, calculating a number of additional nodes to add to at least one hidden layer based on the drift in the activation distributions, resetting connection weights between input layer nodes, hidden layer nodes, and output layer nodes to values before the training of the artificial neural network on the data from the new task, adding the additional nodes to the at least one hidden layer, and training the artificial neural network on data from the new task.

First claim

What is claimed is:

1. A method for retraining an artificial neural network trained on data from an old task, the artificial neural network comprising an input layer having a plurality of input layer nodes, a plurality of hidden layers comprising at least a first hidden layer and a second hidden layer each having a plurality of hidden layer nodes, an output layer having a plurality of output layer nodes, and a plurality of old connections connecting the plurality of input layer nodes, the plurality of hidden layer nodes, and the output layer nodes, the method comprising:
training the artificial neural network on data from the old task and data from a new task different than the old task;
calculating a drift, utilizing Sliced Wasserstein Distance, in activation distributions of the plurality of hidden layer nodes during training of the artificial neural network with the data from the old task and data from the new task;
calculating a number of additional nodes to add to at least one of the plurality of hidden layers based on the drift in the activation distributions;
resetting connection weights between the plurality of input layer nodes, the plurality of hidden layer nodes, and the plurality of output layer nodes to values before the training of the artificial neural network on the data from the new task;
adding a first set of additional nodes of the number of additional nodes to the first hidden layer of the plurality of hidden layers, connecting the first set of additional nodes with new connections, and not connecting the first set of additional nodes added to the first hidden layer to the plurality of hidden layer nodes in the second hidden layer; and
training both the old connections and the new connections of the artificial neural network on data from the new task,
wherein the calculating the number of additional nodes is calculated according to Equation 1:
N nodes =c *log( D )+ b   (Equation 1)

wherein N nodes is the number of additional nodes, c and b are user-specified constants and D is the drift in the activation distributions.

2. The method of claim 1 , wherein data from the old task comprises training data retained from input and target output distributions of the old task.

3. The method of claim 1 , wherein the data from the old task comprises synthetic data generated from a model of input and target output distributions of the old task.

4. The method of claim 1 , wherein:
the number of additional nodes further comprises a second set of additional nodes, and
the method further comprises adding the second set of additional nodes to the second hidden layer of the plurality of hidden layers.

5. The method of claim 4 , further comprising connecting each additional node of the first set of additional nodes to each additional node of the second set of additional nodes.

6. The method of claim 5 , further comprising connecting the plurality of hidden layer nodes in the first hidden layer to the second set of additional nodes added to the second hidden layer.

7. The method of claim 1 , further comprising adding a plurality of new output layer nodes to the output layer.

8. The method of claim 7 , further comprising adding additional nodes of the number of additional nodes to a last hidden layer of the plurality of hidden layers adjacent to the output layer, and connecting the additional nodes only to the plurality of new output layer nodes.

9. The method of claim 8 , further comprising connecting each of the plurality of nodes of the last hidden layer adjacent to the output layer to each of the plurality of new output layer nodes.

10. The method of claim 1 , further comprising connecting each of the plurality of input layer nodes to each of the additional nodes in the first hidden layer.

11. The method of claim 1 , wherein the training the artificial neural network on the data from the new task includes minimizing a loss function with stochastic gradient descent.

Assignees

• Hrl Lab Llc

Inventors

• Martin Charles E
• Ketz Nicholas A
• Pilly Praveen K
• Kolouri Soheil
• Howard Michael D
• Stepp Nigel D

Classifications

• G06N3/006 Primary

  based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO] · CPC title

• G06V20/56

  exterior to a vehicle by using sensors mounted on the vehicle · CPC title

• G06V10/82

  using neural networks · CPC title

• G06V10/764

  using classification, e.g. of video objects · CPC title

• G06N3/047

  Probabilistic or stochastic networks · CPC title