Field programmable gate array architecture for image analysis

A method comprising:

• receiving, from a memory, an input matrix, wherein each element of the input matrix corresponds to a value of a portion of an image;
• providing at least a portion of the input matrix and a first weight vector to a first set of logical blocks of a field programmable gate array (FPGA) device, wherein the FPGA device comprises input/output interfaces communicatively coupled to the memory and a plurality of logical blocks;
• determining, using at least a portion of the first set of logical blocks, a first convolutional kernel by performing exclusive nor operations, implemented by truth tables stored on the first set of logical blocks, between the input matrix and a first weight vector;
• determining, using at least a portion of the first set of logical blocks, based on the first convolutional kernel, a first binary kernel, the first binary kernel comprising a matrix of the same size as the first convolutional kernel with values adjusted to conform to a normal distribution;
• determining, using at least a portion of the first set of logical blocks, a first layer feature map by convoluting the input matrix using the first binary kernel;
• providing the first layer feature map and a second weight vector to the second set of logical blocks;
• determining, using at least a portion of the second set of logical blocks, a second convolutional kernel by performing exclusive nor operations, implemented by truth tables stored on the second set of logical blocks, between the first feature map and the second weight vector;
• determining, using at least a portion of the second set of logical blocks, based on the second convolutional kernel, a pooled kernel, the pooled kernel comprising, for each element of the pooled kernel, a representative value associated with a corresponding pooling region of the second convolutional kernel;
• determining, using at least a portion of the second set of logical blocks, based on the pooled kernel, a second binary kernel, the second binary kernel comprising a matrix of the same size as the pooled kernel with values adjusted to conform to the normal distribution;
• determining, using at least a portion of the second set of logical blocks, a second layer feature map by convoluting the first layer feature map using the second binary kernel;
• determining, based at least in part on the second layer feature map, a probability that the input matrix is associated with a predetermined class of images;
• and
• in response to determining that the probability is greater than a threshold value, providing classification results indicating the image is associated with the class of images.

A field programmable gate array (FPGA) device, the FPGA device comprising:

• input/output interfaces communicatively coupled to a memory of a computing device;
• and
• a plurality of logical blocks, each logical block coupled to a corresponding input/output interface, wherein the plurality of logical blocks comprise:
• a first set of logical blocks, each logical block of the first set of logical blocks comprising a corresponding truth table configured to implement an exclusive nor operation between a first layer input and a first weight vector;
• a second set of logical blocks, each logical block of the second set of logical blocks comprising a corresponding truth table configured to implement an exclusive nor operation between a second layer input and a second weight vector;
• and
• a third set of logical blocks configured to store the first weight vector for the first layer of the neural network model and the second weight vector for the second layer of the neural network model;
• the FPGA device configured to:
• receive, from the memory, the input matrix, wherein each element of the input matrix corresponds to a value of a portion of an image;
• provide at least a portion of the input matrix and the first weight vector to the first set of logical blocks;
• determine, using at least a portion of the first set of logical blocks, a first convolutional kernel by performing the exclusive nor operations, implemented by the truth table of the first set of logical blocks, between the input matrix and the first weight vector;
• determine, using at least a portion of the first set of logical blocks, based on the first convolutional kernel, a first binary kernel, the first binary kernel comprising a matrix of the same size as the first convolutional kernel with values adjusted to conform to a normal distribution;
• determine, using at least a portion of the first set of logical blocks, a first layer feature map by convoluting the input matrix using the first binary kernel;
• provide the first layer feature map and the second weight vector to the second set of logical blocks;
• determine, using at least a portion of the second set of logical blocks, a second convolutional kernel by performing the exclusive nor operations, implemented by the truth tables of the second set of logical blocks, between the first feature map and the second weight vector;
• determine, using at least a portion of the second set of logical blocks, based on the second convolutional kernel, a pooled kernel, the pooled kernel comprising, for each element of the pooled kernel, a representative value associated with a corresponding pooling region of the second convolutional kernel;
• determine, using at least a portion of the second set of logical blocks, based on the pooled kernel, a second binary kernel, the second binary kernel comprising a matrix of the same size as the pooled kernel with values adjusted to conform to the normal distribution;
• determine, using at least a portion of the second set of logical blocks, a second layer feature map by convoluting the first layer feature map using the second binary kernel;
• determine, based at least in part on the second layer feature map, a probability that the input matrix is associated with a predetermined class of images;
• and
• in response to determining that the probability is greater than a threshold value, provide classification results indicating the image is associated with the class of images.