Incorporating top-down information in deep neural networks via the bias term

1. A computer-implemented method for classifying an element in an input by biasing a deep neural network, comprising:

determining, prior to classifying the element in the input to the deep neural network, whether the element has an increased probability of being present in the input to the deep neural network based on at least one of a location of the input, a time of the input, or another element classified in the input;

adjusting a bias of activation functions of neurons in the deep neural network to increase sensitivity to the element based on whether the element has an increased probability of being present in the input, the bias being adjusted without adjusting weights of the deep neural network;

adjusting an output of the deep neural network based at least in part on the adjusted bias; and

classifying the element based on the output of the deep neural network.

2. The computer-implemented method of claim 1 , further comprising determining an amount of adjustment for the bias as a constant, function of a synaptic weight, or function of activations in response to presentation of a target class.

3. The computer-implemented method of claim 1 , in which adjusting the bias comprises scaling the bias.

4. The computer-implemented method of claim 3 , in which a scale of the adjusting is based at least in part on a priori knowledge of what is likely to be present in the input and/or what is not likely to be present in the input.

5. The computer-implemented method of claim 1 , in which the adjusting is performed at an internal level of the deep neural network.

6. An apparatus for classifying an element in an input by biasing a deep neural network, comprising:

means for determining, prior to classifying the element in the input to the deep neural network, whether the element has an increased probability of being present in the input to the deep neural network based on at least one of a location of the input, a time of the input, or another element classified in the input;

means for adjusting a bias of activation functions of neurons in the deep neural network to increase sensitivity to the element based on whether the element has an increased probability of being present in the input, the bias being adjusted without adjusting weights of the deep neural network;

means for adjusting an output of the deep neural network classifying the element based at least in part on the adjusted bias; and

means for classifying the element based on the output of the deep neural network.

7. The apparatus of claim 6 , further comprising means for determining an amount of adjustment for the bias as a constant, function of a synaptic weight, or function of activations in response to presentation of a target class.

8. The apparatus of claim 6 , in which the means for adjusting the bias comprises means for scaling the bias.

9. The apparatus of claim 8 , in which a scale of the adjusting is based at least in part on a priori knowledge of what is likely to be present in the input and/or what is not likely to be present in the input.

10. The apparatus of claim 6 , in which the means for adjusting is performed at an internal level of the deep neural network.

11. An apparatus for classifying an element in an input by biasing a deep neural network, comprising:

a memory unit; and

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

to determine, prior to classifying the element in the input to the deep neural network, whether the element has an increased probability of being present in the input to the deep neural network based on at least one of a location of the input, a time of the input, or another element classified in the input;

to adjust a bias of activation functions of neurons in the deep neural network to increase sensitivity to the element based on whether the element has an increased probability of being present in the input, the bias being adjusted without adjusting weights of the deep neural network;

to adjust an output of the deep neural network based at least in part on the adjusted bias; and

to classify the element based on the output of the deep neural network.

12. The apparatus of claim 11 , in which the at least one processor is further configured to determine an amount of adjustment for the bias as a constant, function of a synaptic weight, or function of activations in response to presentation of a target class.

13. The apparatus of claim 11 , in which the at least one processor is further configured to adjust the bias by scaling the bias.

14. The apparatus of claim 13 , in which a scale of the adjusting is based at least in part on a priori knowledge of what is likely to be present in the input and/or what is not likely to be present in the input.

15. The apparatus of claim 11 , in which the at least one processor is further configured to adjust the bias at an internal level of the deep neural network.

16. A non-transitory computer-readable medium having program code recorded thereon for classifying an element in an input by a deep neural network, the program code being executed by a processor and comprising:

program code to determine, prior to classifying the element in the input to the deep neural network, whether the element has an increased probability of being present in the input to the deep neural network based on at least one of a location of the input, a time of the input, or another element classified in the input;

program code to adjust a bias of activation functions of neurons in the deep neural network to increase sensitivity to the element based on whether the element has an increased probability of being present in the input, the bias being adjusted without adjusting weights of the deep neural network;

program code to adjust an output of the deep neural network based at least in part on the adjusted bias; and

program code to classify the element based on the output of the deep neural network.

17. The computer-readable medium of claim 16 , in which the program code further comprises program code to determine an amount of adjustment for the bias as a constant, function of a synaptic weight, or function of activations in response to presentation of a target class.

18. The computer-readable medium of claim 16 , in which the program code to adjust the bias comprises program code to adjust the bias by scaling the bias.

19. The computer-readable medium of claim 18 , in which a scale of the adjusting is based at least in part on a priori knowledge of what is likely to be present in the input and/or what is not likely to be present in the input.

20. The computer-readable medium of claim 16 , in which the program code to adjust the bias comprises program code to adjust the bias at an internal level of the deep neural network.