Parsimonious inference on convolutional neural networks
AssigneeIrida Labs S.A.
InventorsIlias Theodorakopoulos, Vassileios Pothos, Dimitris Kastaniotis, Nikos Fragoulis
Life of this patent
Estimating the term…
Abstract
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.
Claims5
A method, in a convolutional neural network (CNN), comprising:
- expressing the CNN final structure by a number of total convolutional layers, and
- automatically determining a number of total kernels in each specific layer utilizing a regular back-propagation optimization procedure, wherein a Learning Kernel Activation Module (LKAM) generates, for each input datum during inference, a gating vector based on feature maps of a preceding convolutional layer, and the gating vector selectively activates or deactivates convolutional kernels of a subsequent convolutional layer for that input datum, and wherein the CNN is implemented in VLSI hardware, and each kernel is connected to a CMOS transistor switch such that kernels deactivated by the gating vector are electrically disconnected from one or more of power supply and a clock signal.
A method to use a Learning Kernel Activation Module (LKAM) comprising:
- automatically defining an optimal number of layers and convolutional kernels for a given training task through a regular back-propagation optimization procedure, wherein the LKAM generates, for each input datum during inference, a gating vector based on feature maps of a preceding convolutional layer, and the gating vector selectively activates or deactivates convolutional kernels of a subsequent convolutional layer for that input datum, and wherein the LKAM is implemented in VLSI hardware, and each kernel is connected to a CMOS transistor switch such that kernels deactivated by the gating vector are electrically disconnected from one or more of power supply and a clock signal.
Dependencies are inferred from the claim text.