LSTM Neural Networks for Brain Signals and Neuromorphic Chip

Long Short-Term Memory (LSTM) is a type of RNN architecture commonly employed in natural language processing, speech recognition, and various sequence modeling applications. A normal recurrent neural network (RNN), which features a simple architecture of input, hidden state, and output, an LSTM is distinguished by a more complex structure that incorporates supplementary memory cells and gates. This architecture allows the LSTM to selectively retain or discard information from previous time steps. The fundamental component of these devices is a memory cell, which has two input gates, one output gate, and a forget gate. These gates manage the inflow and outflow of data, allowing LSTMs to choose to retain or eliminate information. The articles concentrate on the hardware chip design of the LSTM-like mechanism in neuromorphic systems, which incorporates two LSTM layers and processes eight neurons together with their weights in Xilinx ISE 14.7 and simulates the design for logic verification test cases. LSTM networks could inspire neuromorphic technologies that process time-sensitive information and memory like the brain. The novelty of the work is in the scalable hardware chip design and its performance is synthesized on the FPGA that supports the frequency 570.00 MHz. The design has shown the optimal performance in terms of parameters such as delay, frequency, memory, flip flops, and IoBs onVirtex-5 FPGA for scalable LSTM that is tested up to 64 neurons. The hybrid architecture is based on two intermediate LSTM layers processing 8 neurons in a single slot.