Feature map sparsification with smoothness regularization

1. A method comprising:

receiving an image by a deep neural network (DNN) implemented by at least one processor; and

obtaining, by the DNN, a first feature map based on the image while the DNN is in a trained state, wherein

the DNN is configured to perform a task based on the image, and is trained with a training image by using a feature sparsification with smoothness regularization process and a back propagation and weight update process that updates the DNN based on an output of the feature sparsification with smoothness regularization process, and

the method further comprises training the DNN, the training comprising:

obtaining, by the at least one processor, a second feature map based on the training image by performing a network forward computation, by the DNN, that includes the feature sparsification with smoothness regularization process;

computing, by the at least one processor, a regularization loss of the second feature map that is based on the training image;

computing, by the at least one processor, a smoothness loss of the second feature map that is based on the training image;

computing, by the at least one processor, a total gradient based on the regularization loss and the smoothness loss that are computed; and

updating, by the at least one processor, network coefficients of the DNN by performing the back propagation and weight update process based on the total gradient that is computed.

2. The method of claim 1 , wherein

the training the DNN further comprises computing, by the at least one processor, an empirical data loss of the task performed by the DNN, based on the training image, and

the computing the total gradient comprises computing the total gradient based on the empirical data loss, the regularization loss, and the smoothness loss that are computed.

3. The method of claim 1 , wherein

the training the DNN further comprises accumulating, by the at least one processor, total gradients of a batch of training data, including the total gradient, and

the updating the network coefficients of the DNN is performed based on the total gradients that are accumulated.

4. The method of claim 1 , wherein

the training the DNN is performed over multiple iterations, and the training the DNN further comprises changing hyperparameters over the iterations such that the training emphasizes learning a sparse feature map, and then emphasizes smoothing feature responses in later ones of the iterations.

5. The method of claim 1 , wherein

the training the DNN is performed over multiple iterations, and the training the DNN further comprises changing a hyperparameter over the iterations such that the training emphasizes smoothness within spatial dimensions, and then emphasizes channel wide smoothness in later ones of the iterations.

6. A method comprising:

receiving an image by a deep neural network (DNN) implemented by at least one processor; and

obtaining, by the DNN, a first feature map based on the image while the DNN is in a trained state, wherein

the DNN is configured to perform a task based on the image, and is trained with a training image by using a feature sparsification with smoothness regularization process and a back propagation and weight update process that updates the DNN based on an output of the feature sparsification with smoothness regularization process, and

the back propagation and weight update process updates network coefficients of the DNN, during the training, based on:

a regularization loss and a smoothness loss that are computed based on the output of the feature sparsification with smoothness regularization process, and

an empirical data loss of the task performed by the DNN that is computed.

7. The method of claim 1 , further comprising compressing the first feature map.

8. The method of claim 1 , wherein the DNN is configured to perform at least one from among semantic segmentation, image or video classification, object detection, and image or video super-resolution, as the task.

9. A system comprising:

at least one memory configured to store computer program code; and

a deep neural network (DNN) implemented by at least one processor configured to access the computer program code and operate as instructed by the computer program code, the computer program code including:

obtaining code configured to cause the at least one processor to obtain a first feature map based on an image inputted into the DNN while the DNN is in a trained state, wherein

the DNN is configured to perform a task based on the image, and is trained with a training image by using a feature sparsification with smoothness regularization process and a back propagation and weight update process that updates the DNN based on an output of the feature sparsification with smoothness regularization process, and

the DNN is trained by:

obtaining a second feature map based on the training image by performing a network forward computation, by the DNN, that includes the feature sparsification with smoothness regularization process;

computing a regularization loss of the second feature map that is based on the training image;

computing a smoothness loss of the second feature map that is based on the training image;

computing a total gradient based on the regularization loss and the smoothness loss that are computed; and

updating network coefficients of the DNN by performing the back propagation and weight update process based on the total gradient that is computed.

10. The system of claim 9 , wherein the DNN is further trained by:

computing an empirical data loss of the task performed by the DNN, based on the training image, and

the total gradient is computed based on the empirical data loss, the regularization loss, and the smoothness loss that are computed.

11. The system of claim 9 , wherein

the DNN is further trained by accumulating total gradients of a batch of training data, including the total gradient, and

the network coefficients of the DNN is updated based on the total gradients that are accumulated.

12. The system of claim 9 , wherein the DNN is trained by updating the DNN over multiple iterations, and changing hyperparameters over the iterations such that the training emphasizes learning a sparse feature map, and then emphasizes smoothing feature responses in later ones of the iterations.

13. The system of claim 9 , wherein the DNN is trained by updating the DNN over multiple iterations, and changing a hyperparameter over the iterations such that the training emphasizes smoothness within spatial dimensions, and then emphasizes channel wide smoothness in later ones of the iterations.

14. A system comprising:

at least one memory configured to store computer program code; and

a deep neural network (DNN) implemented by at least one processor configured to access the computer program code and operate as instructed by the computer program code, the computer program code including:

obtaining code configured to cause the at least one processor to obtain a first feature map based on an image inputted into the DNN while the DNN is in a trained state, wherein

the DNN is configured to perform a task based on the image, and is trained with a training image by using a feature sparsification with smoothness regularization process and a back propagation and weight update process that updates the DNN based on an output of the feature sparsification with smoothness regularization process, and

the back propagation and weight update process updates network coefficients of the DNN, during the training, based on:

a regularization loss and a smoothness loss that are computed based on the output of the feature sparsification with smoothness regularization process, and

an empirical data loss of the task performed by the DNN that is computed.

15. The system of claim 9 , the computer program code further comprises compression code configured to cause the at least one processor to compress the first feature map.

16. A non-transitory computer-readable medium storing computer instructions that are configured to, when executed by at least one processor implementing a deep neural network (DNN), cause the at least one processor to:

obtain a first feature map based on an image inputted into the DNN while the DNN is in a trained state, wherein

the DNN is configured to perform a task based on the image, and is trained with a training image by using a feature sparsification with smoothness regularization process and a back propagation and weight update process that updates the DNN based on an output of the feature sparsification with smoothness regularization process,

the computer instructions are further configured to cause the at least one processor to train the DNN by:

obtaining, by the at least one processor, a second feature map based on the training image by performing a network forward computation, by the DNN, that includes the feature sparsification with smoothness regularization process;

computing, by the at least one processor, a regularization loss of the second feature map that is based on the training image;

computing, by the at least one processor, a smoothness loss of the second feature map that is based on the training image;

computing, by the at least one processor, a total gradient based on the regularization loss and the smoothness loss that are computed; and

updating, by the at least one processor, network coefficients of the DNN by performing the back propagation and weight update process based on the total gradient that is computed.