Handwritten Digit Detection and Prototyping on Field Programmable Gate Array (FPGA)

Neural Networks is a subset of Machine Learning (ML) where real-world data and algorithms are used to make the machine learn to mimic the way a human learns. An Extreme Learning Machine (ELM) is a specialized form of those Neural Networks for the training of Single Hidden Layer Feedforward Network (SLFN). The ELM algorithm has been presented to be faster than many other conventional algorithms with acceptable accuracy and good generalized performance. The algorithm allows the learning to be completed in a single step, which is made possible by using a pseudo-inverse matrix instead of back propagation to obtain the trained weights. The system uses random values for its initial set of weights, which are generated using a Linear Feedback Shift Register (LFSR). For the hardware implementation, a 256-bit input register is used to store the image. After the generation of weights, an intermediate signal is set to reset, moving the Finite State Machine (FSM) to the next weight fetching state. When the next set of weights has been calculated, the maximum output is given out through one hot encoding. An ELM is trained to detect handwritten digits from an image. Then a Register transfer level (RTL) is designed and further prototyped on a field programmable gate array (FPGA). Basys 3 is the FPGA board that will be used to implement the design. The one hot encoded output can be observed on the FPGA. The experimental results show a greater software accuracy of 99% and hardware accuracy of 90% in detecting handwritten digits from the dataset.