Two hardware implementation of a feed-forward Convolutional Neural Network
- XNOR-Net : faster execution due to the replacement of vector-matrix multiplication to “XNOR + Popcount” operation
- Binary Connect : faster execution due to the presence of accumulator
- For the current scope of the project, we consider MNIST dataset which consists of 10 classes of images i.e. 0 to 9.
- Once the training of the neural network is done using any machine, our hardware design focuses on achieving speedup in the classification phase
- The weights learned by the neural network will be inserted in the module using a BRAM to the feed forward network
- These weights along with the images to be classified would be used as an input to the feed forward network on an FPGA
- The scope of our project is only defined for the implementation of the fully connected layer after the convolutions have been done
For more information click here