Artificial neural networks having attention-based selective plasticity and methods of training the same

1. An autonomous system for a vehicle, the autonomous system comprising:

a controller configured to control the vehicle;

a plurality of sensors configured to detect objects in a path of the vehicle;

nonvolatile memory having an artificial neural network stored therein configured to classify the objects detected by the plurality of sensors, the artificial neural network comprising a plurality of neurons in each of an input layer, at least one hidden layer, and an output layer; and

a processor,

wherein the nonvolatile memory includes instructions which, when executed by the processor, cause the processor to:

train the artificial neural network on a first task;

identify, utilizing a contrastive excitation backpropagation algorithm, important neurons of the plurality of neurons for the first task;

identify, utilizing a learning algorithm, important synapses between the plurality of neurons for the first task based on the important neurons identified; and

rigidify the identified important synapses to achieve selective plasticity of the plurality of neurons in the artificial neural network when being trained on one or more new tasks.

2. The autonomous system of claim 1 , wherein the instructions, when executed by the processor, further cause the processor to train the artificial neural network on a second task different than the first task.

3. The autonomous system of claim 2 , wherein the instructions, when executed by the processor, further cause the processor to:

send at least one input of the second task to the input layer;

generate, at the output layer of the plurality of layers, at least one output based on the at least one input;

generate a reward based on a comparison between the at least one output and a desired output; and

modify weights of the synapses based on the reward.

4. The autonomous system of claim 3 , wherein, during training of the artificial neural network on the second task, the weights of the important synapses remain constant.

5. An autonomous system for a vehicle, the autonomous system comprising:

a controller configured to control the vehicle;

a plurality of sensors configured to detect objects in a path of the vehicle;

nonvolatile memory having an artificial neural network stored therein configured to classify the objects detected by the plurality of sensors, the artificial neural network comprising a plurality of neurons in each of an input layer, at least one hidden layer, and an output layer; and

a processor,

wherein the nonvolatile memory includes instructions which, when executed by the processor, cause the processor to:

train the artificial neural network on a first task;

identify, utilizing a contrastive excitation backpropagation algorithm, important neurons of the plurality of neurons for the first task;

identify, utilizing a learning algorithm, important synapses between the plurality of neurons for the first task based on the important neurons identified;

rigidify the identified important synapses to achieve selective plasticity of the plurality of neurons in the artificial neural network when being trained on one or more new tasks;

train the artificial neural network on a second task different than the first task;

send at least one input of the second task to the input layer of the plurality of layers;

generate, at the output layer of the plurality of layers, at least one output based on the at least one input;

generate a reward based on a comparison between the at least one output and a desired output; and

modify weights of the synapses based on the reward,

wherein, during training of the artificial neural network on the second task, the weights of the important synapses remain constant, and

wherein the learning algorithm to identify important synapses is a Hebbian learning algorithm as follows:

β ji l =β ji l +P ( a j l ( x n )) P ( a i l+1 ( x n )),

where β ji l is a synaptic importance parameter, x n is an input image, a j l is a j′th neuron in an l′th layer of the artificial neural network, a i l+1 is an i′th neuron in layer l+1 of the artificial neural network, and P is a probability.

6. An autonomous system for a vehicle, the autonomous system comprising:

a controller configured to control the vehicle;

a plurality of sensors configured to detect objects in a path of the vehicle;

nonvolatile memory having an artificial neural network stored therein configured to classify the objects detected by the plurality of sensors, the artificial neural network comprising a plurality of neurons in each of an input layer, at least one hidden layer, and an output layer; and

a processor,

wherein the nonvolatile memory includes instructions which, when executed by the processor, cause the processor to:

train the artificial neural network on a first task;

identify, utilizing a contrastive excitation backpropadation algorithm, important neurons of the plurality of neurons for the first task;

identify, utilizing a learning algorithm, important synapses between the plurality of neurons for the first task based on the important neurons identified;

rigidify the identified important synapses to achieve selective plasticity of the plurality of neurons in the artificial neural network when being trained on one or more new tasks;

train the artificial neural network on a second task different than the first task;

send at least one input of the second task to the input layer of the plurality of layers;

generate, at the output layer of the plurality of layers, at least one output based on the at least one input;

generate a reward based on a comparison between the at least one output and a desired output; and

modify weights of the synapses based on the reward,

wherein, during training of the artificial neural network on the second task, the weights of the important synapses remain constant, and

wherein the learning algorithm to identify important synapses is Oja's learning rule as follows:

γ ji l =γ ji l +∈( P c ( f j (l−1) ) P c ( f j l )− P c ( f i l ) 2 γ ji l ),

where i and j are neurons, l is a layer of the artificial neural network, P c is a probability, γ ji l is the importance of the synapse between the neurons f j (l−1) and f i l for the first task, ∈ is the rate of Oja's learning rule, and P c is a probability.

7. The autonomous system of claim 6 , wherein the instructions which, when executed by the processor, further cause the processor to update a loss function of the artificial neural network as follows:

(θ)= B (θ)+λΣ k γ k (θ k −θ* A,k ) 2 ,

where (θ) is the loss function, B (θ) is an original loss function for learning a second task different than the first task, λ is the regularization coefficient, γ k is the synaptic importance parameter of Oja's learning rule, θ k is the synaptic weights, and θ* A,k are the optimized synaptic weights for performing the first task.

8. A non-transitory computer-readable storage medium having software instructions stored therein, which, when executed by a processor, cause the processor to:

train an artificial neural network on a first task;

identify, utilizing a contrastive excitation backpropagation algorithm, important neurons of the artificial neural network for the first task;

identify, utilizing a learning algorithm, important synapses between the important neurons; and

rigidify the identified important synapses to achieve selective plasticity of the artificial neural network when being trained on one or more new tasks.

9. The non-transitory computer-readable storage medium of claim 8 , wherein the instructions, when executed by the processor, further cause the processor to train the artificial neural network on a second task different than the first task.

10. The non-transitory computer-readable storage medium of claim 9 , wherein the instructions, when executed by the processor, further cause the processor to:

send at least one input of the second task to an input layer of the artificial neural network;

receive at least one output from an output layer of the artificial neural network based on the at least one input;

generate a reward based on a comparison between at least one output and a desired output; and

modify weights of the synapses based on the reward.

11. The non-transitory computer-readable storage medium of claim 10 , wherein, during training of the artificial neural network on the second task, the weights of the important synapses remain constant.

12. The non-transitory computer-readable storage medium of claim 11 , wherein the learning algorithm is a Hebbian learning algorithm as follows:

β ji l =β ji l +P ( a j l ( x n )) P ( a i l+1 ( x n )),

where β ji l is a synaptic importance parameter, x n is an input image, a j l is a j′th neuron in an l′th layer of the artificial neural network, a i l+1 is an i′th neuron in layer l+1 of the artificial neural network, and P is a probability.

13. The non-transitory computer-readable storage medium of claim 11 , wherein the learning algorithm is Oja's learning rule as follows:

γ ji l =γ ji l +∈( P c ( f j (l−1) ) P c ( f j l )− P c ( f i l ) 2 γ ji l ),

where i and j are neurons, l is a layer of the artificial neural network, P c is a probability, γ ji l is the importance of the synapse between the neurons f j (l−1) and f i l for the first task, ∈ is the rate of Oja's learning rule, and P c is a probability.

14. The non-transitory computer-readable storage medium of claim 13 , wherein the instructions which, when executed by the processor, further cause the processor to update a loss function of the artificial neural network as follows:

ℒ

⁡

(

θ

)

=

ℒ

B

⁡

(

θ

)

+

λ

⁢

∑

k

⁢

γ

k

⁡

(

θ

k

-

θ

A

,

k

*

)

2

where (θ) is the loss function, B (θ) is an original loss function for learning a second task different than the first task, λ is the regularization coefficient, γ k is the synaptic importance parameter of Oja's learning rule, θ k is the synaptic weights, and θ* A,k are the optimized synaptic weights for performing the first task.

15. A method of training an artificial neural network having a plurality of layers, each layer of the plurality of layers comprising a plurality of neurons, and at least one weight matrix encoding connection weights between neurons in successive layers of the plurality of layers, the method comprising:

training the artificial neural network on a first task;

identifying, utilizing contrastive excitation backpropagation, important neurons for the first task;

identifying, utilizing a learning algorithm, important synapses for the first task based on the important neurons identified; and

rigidifying the identified important synapses to achieve selective plasticity of the plurality of neurons in the artificial neural network when being trained on one or more new tasks.

16. The method of claim 15 , further comprising training the artificial neural network on a second task different than the first task, the training of the artificial neural network on the second task comprising:

sending at least one input of the second task to an input layer of the plurality of layers;

generating, at an output layer of the plurality of layers, at least one output based on the at least one input;

generating a reward based on a comparison between the at least one output and a desired output; and

modifying the connection weights based on the reward.

17. The method of claim 16 , wherein, during the training of the artificial neural network on the second task, the weights of the important synapses remain constant.

18. The method of claim 17 , wherein the learning algorithm is a Hebbian learning algorithm as follows:

β ji l =β ji l +P ( a j l ( x n )) P ( a i l+1 ( x n )),

where β ji l is a synaptic importance parameter, x n is an input image, a j l is a j′th neuron in an l′th layer of the artificial neural network, a i i+1 is an i′th neuron in layer l+1 of the artificial neural network, and P is a probability.

19. The method of claim 17 , wherein the learning algorithm is Oja's learning rule as follows:

γ ji l =γ ji l +∈( P c ( f j (l−1) ) P c ( f j l )− P c ( f i l ) 2 γ ji l ),

where i and j are neurons, l is a layer of the artificial neural network, P c is a probability, γ ji l is the importance of the synapse between the neurons f j (l−1) and f i l for the first task, ∈ is the rate of Oja's learning rule, and P c is a probability.

20. The method of claim 19 , further comprising updating a loss function of the artificial neural network as follows:

(θ)= B (θ)+λΣ k γ k (θ k −θ* A,k ) 2

where (θ) is the loss function, B (θ) is an original loss function for learning a second task different than the first task, λ is the regularization coefficient, γ k is the synaptic importance parameter of Oja's learning rule, θ k is the synaptic weights, and θ* A,k are the optimized synaptic weights for performing the first task.