Configuring sparse neuronal networks

1. A computer-implemented method for selecting model neurons in an artificial neural network, comprising:

generating a first sparse set of decoding weights, each decoding weight corresponding to a synapse between a first neuron layer and a second neuron layer;

selecting, from the first neuron layer, a first set of model neurons corresponding to a first set of non-zero decoding weights selected from the first sparse set of decoding weights;

updating the first neuron layer to consist of the first set of model neurons; and

operating the artificial neural network with the updated first neuron layer.

2. The computer-implemented method of claim 1 , further comprising performing a least squares optimization to generate the first sparse set of decoding weights.

3. The computer-implemented method of claim 2 , in which the least squares optimization comprises least absolute shrinkage and selection operator (LASSO) regularization.

4. The computer-implemented method of claim 2 , further comprising selecting a regularization term to reduce a number of non-zero weights in the first sparse set of decoding weights.

5. The computer-implemented method of claim 1 , further comprising:

generating a second sparse set of decoding weights, each decoding weight from the second sparse set corresponding to a synapse between the updated first neuron layer and the second neuron layer; and

operating the artificial neural network with a second set of model neurons in the updated first neuron layer corresponding to a second set of non-zero decoding weights selected from the second sparse set of decoding weights.

6. The computer-implemented method of claim 1 , further comprising:

repeating the generating and the selecting a plurality of times to create a pooled set of model neurons corresponding to non-zero decoding weights; and

operating the artificial neural network with a third set of model neurons selected from the pooled set of model neurons.

7. An apparatus for selecting model neurons in an artificial neural network, comprising:

a memory; and

at least one processor coupled to the memory, the at least one processor being configured:

to generate a first sparse set of decoding weights, each decoding weight corresponding to a synapse between a first neuron layer and a second neuron layer;

to select, from the first neuron layer, a first set of model neurons corresponding to a first set of non-zero decoding weights selected from the first sparse set of decoding weights;

to update the first neuron layer to consist of the first set of model neurons; and

to operate the artificial neural network with the updated first neuron layer.

8. The apparatus of claim 7 , in which the at least one processor is further configured to perform a least squares optimization to generate the first sparse set of decoding weights.

9. The apparatus of claim 8 , in which the least squares optimization comprises least absolute shrinkage and selection operator (LASSO) regularization.

10. The apparatus of claim 8 , in which the at least one processor is further configured to select a regularization term to reduce a number of non-zero weights in the first sparse set of decoding weights.

11. The apparatus of claim 7 , in which the at least one processor is further configured:

to generate a second sparse set of decoding weights, each decoding weight from the second sparse set corresponding to a synapse between the updated first neuron layer and the second neuron layer; and

to operate the artificial neural network with a second set of model neurons in the updated first neuron layer corresponding to a second set of non-zero decoding weights selected from the second sparse set of decoding weights.

12. The apparatus of claim 7 , in which the at least one processor is further configured:

to repeat the generating and the selecting a plurality of times to create a pooled set of model neurons corresponding to non-zero decoding weights; and

to operate the artificial neural network with a third set of model neurons selected from the pooled set of model neurons.

13. An apparatus for selecting model neurons in an artificial neural network, comprising:

means for generating a first sparse set of decoding weights, each decoding weight corresponding to a synapse between a first neuron layer and a second neuron layer;

means for selecting, from the first neuron layer, a first set of model neurons corresponding to a first set of non-zero decoding weights selected from the first sparse set of decoding weights;

means for updating the first neuron layer to consist of the first set of model neurons; and

means for operating the artificial neural network with the updated first neuron layer.

14. The apparatus of claim 13 , further comprising means for performing a least squares optimization to generate the first sparse set of decoding weights.

15. The apparatus of claim 14 , in which the least squares optimization comprises least absolute shrinkage and selection operator (LASSO) regularization.

16. The apparatus of claim 14 , further comprising means for selecting a regularization term to reduce a number of non-zero weights in the first sparse set of decoding weights.

17. The apparatus of claim 13 , further comprising:

means for generating a second sparse set of decoding weights, each decoding weight from the second sparse set corresponding to a synapse between the updated first neuron layer and the second neuron layer; and

means for operating the artificial neural network with a second set of model neurons in the updated first neuron layer corresponding to a second set of non-zero decoding weights selected from the second sparse set of decoding weights.

18. The apparatus of claim 13 , further comprising:

means for repeating the generating and the selecting a plurality of times to create a pooled set of model neurons corresponding to non-zero decoding weights; and

means for operating the artificial neural network with a third set of model neurons selected from the pooled set of model neurons.

19. A non-transitory computer readable medium having encoded thereon program code for selecting model neurons in an artificial neural network, the program code executed by a processor and comprising:

program code to generate a first sparse set of decoding weights, each decoding weight corresponding to a synapse between a first neuron layer and a second neuron layer;

program code to select, from the first neuron layer, a first set of model neurons corresponding to a first set of non-zero decoding weights selected from the first sparse set of decoding weights;

program code to update the first neuron layer to consist of the first set of model neurons; and

program code to operate the artificial neural network with the updated first neuron layer.

20. The non-transitory computer readable medium of claim 19 , further comprising program code to perform a least squares optimization to generate the first sparse set of decoding weights.

21. The non-transitory computer readable medium of claim 20 , in which the least squares optimization comprises least absolute shrinkage and selection operator (LASSO) regularization.

22. The non-transitory computer readable medium of claim 20 , further comprising program code to select a regularization term to reduce a number of non-zero weights in the first sparse set of decoding weights.

23. The non-transitory computer readable medium of claim 19 , further comprising:

program code to generate a second sparse set of decoding weights, each decoding weight from the second sparse set corresponding to a synapse between the updated first neuron layer and the second neuron layer; and

program code to operate the artificial neural network with a second set of model neurons in the updated first neuron layer corresponding to a second set of non-zero decoding weights selected from the second sparse set of decoding weights.

24. The non-transitory computer readable medium of claim 19 , further comprising:

program code to repeat the generating and the selecting a plurality of times to create a pooled set of model neurons corresponding to non-zero decoding weights; and

program code to operate the artificial neural network with a third set of model neurons selected from the pooled set of model neurons.