Parsimonious inference on convolutional neural networks
The disclosed system incorporates a new learning module, the Learning Kernel Activation Module (LKAM), at least serving the purpose of enforcing the utilization of less convolutional kernels by learning kernel activation rules and by actually controlling the engagement of various computing elements: The exemplary module activates/deactivates a sub-set of filtering kernels, groups of kernels, or groups of full connected neurons, during the inference phase, on-the-fly for every input image depending on the input image content and the learned activation rules.
1. A method in a convolutional neural network (CNN) with more than three layers, the method comprising:
operating the CNN during an inference phase by utilizing a learning kernel-activation module (LKAM) which is inserted between a first and a second convolutional layer in the CNN, wherein said LKAM is a second CNN with one or two layers and wherein the LKAM has as inputs feature maps that are output from the first convolutional layer and has as output a vector of numbers indicating whether the convolutional kernels in the second convolutional layer are on or off; and
switching off during the inference phase at least one convolutional kernel of the CNN based on the output vector of the LKAM;
wherein the LKAM performs operations on the feature maps realizing a set of learned rules and constituted by a combination of a number of N×N convolutional kernal banks, operations of average or max pooling operating globally or locally and a sigmoid function.
2. The method of claim 1 , wherein a level of engagement is determined through optimization of a cost function that utilizes a regularization term proportional to a number of kernels that are engaged in each forward propagation step.
3. The method of claim 2 , wherein the cost function is
L
aug
=
G
i
2
m
∑
i
sw
i
,
where sw i are the elements of SW vector which is the output of LKAM, G i is a gain factor and m is the length of the SW vector.
4. The method of claim 1 , where the convolutional kernels that are off are electrically switched-off when the neural networks are implemented in VLSI.
5. The method of claim 1 , wherein an external parameter is used to control an amount of computations of the CNN during the inference phase.
6. A non-transitory computer readable information storage media having stored therein instructions, that when executed by one or more processors, cause to be performed a method in a convolutional neural network (CNN) with more than three layers, the method comprising:
operating the CNN during an inference phase by utilizing a learning kernel-activation module (LKAM) which is inserted between a first and a second convolutional layer in the CNN, wherein said LKAM is a second CNN with one or two layers and wherein the LKAM has as inputs feature maps that are output from the first convolutional layer and has as output a vector of numbers indicating whether the convolutional kernels in the second convolutional layer are on or off; and
switching off during the inference phase at least one convolutional kernel of the CNN based on the output vector of the LKAM;
wherein the LKAM performs operations on the feature maps realizing a set of learned rules and constituted by a combination of a number of N×N convolutional kernal banks, operations of average or max pooling operating globally or locally and a sigmoid function.
7. The media of claim 6 , wherein a level of engagement is determined through optimization of a cost function that utilizes a regularization term proportional to a number of kernels that are engaged in each forward propagation step.
8. The media of claim 7 , wherein the cost function is
L
aug
=
G
i
2
m
∑
i
sw
i
,
where sw i are the elements of SW vector which is the output of LKAM, G i is a gain factor and m is the length of the SW vector.
9. The media of claim 6 , where the convolutional kernels that are off are electrically switched-off when the neural networks are implemented in VLSI.
10. The media of claim 6 , wherein an external parameter is used to control the amount of computations of the CNN during the inference phase.
11. A system comprising a convolutional neural network (CNN) with more than three layers, the system comprising:
means, including a processor and memory, for operating the CNN during an inference phase by utilizing a learning kernel-activation module (LKAM) which is inserted between a first and a second convolutional layer in the CNN, wherein said LKAM is a second CNN with one or two layers and wherein the LKAM has as inputs the feature maps that are output from the first convolutional layer and has as output a vector of numbers indicating whether the convolutional kernels in the second convolutional layer are on or off; and
means for switching off during the inference phase at least one convolutional kernel of the CNN based on the output vector of the LKAM,
wherein the LKAM performs operations on the feature maps realizing a set of learned rules and constituted by a combination of a number of N×N convolutional kernal banks, operations of average or max pooling operating globally or locally and a sigmoid function.