Multi-kernel configuration for convolutional neural networks

A system comprising:

• a processor;
• a structure configured to be in communication with the processor, the structure comprises:
• a plurality of multi-kernel modules, wherein a multi-kernel module comprises a plurality of unit cells, a unit cell corresponds to an element of a kernel being implemented in a convolutional neural network, and wherein the unit cell comprises:
• a control circuit comprising a storage component configured to store a weight of a corresponding element of the kernel;
• a plurality of operation circuits connected to the control circuit, an operation circuit of the plurality of operation circuits being operable to receive an input signal associated with a data unit among an array of data units;
• wherein the control circuit is operable to control an application of the stored weight to the input signals received by the plurality of operation circuits to implement the convolutional neural network;
• the processor being configured to:
• receive an input comprising the array of data units;
• receive a request to classify the input by implementing the convolutional neural network;
• send the input to the structure to perform the classification of the input.

A computer-implemented method comprising:

• receiving an input comprising an array of data units;
• receiving a request to classify the input by implementing a convolutional neural network;
• sending input signals associated with the array of data units of the input to a plurality of unit cells of a structure, wherein the structure comprises a plurality of multi-kernel modules, a multi-kernel module comprises a respective set of unit cells, a unit cell corresponds to an element of a kernel being implemented in the convolutional neural network, and wherein the unit cell comprises a storage component configured to store a weight of a corresponding element of the kernel;
• activating a first pass gate of each unit cell in the structure to pass a stored weight of each unit cell to a plurality of operation circuits in the corresponding unit cell, wherein passing the stored weight to the plurality of operation circuits causes the unit cells to apply the stored weight to input signals received at the plurality of operation circuits;
• generating a set of outputs based on the application of the stored weights to the input signals;
• and
• classifying the input based on the set of outputs.