Browsing by Author "Abrham Ayal"

Now showing 1 - 1 of 1
  • No Thumbnail Available
    Item
    Implementation of the Programmable Memristor based – Long Short Term Memory (LSTM) Circuit for Time Series Prediction
    (Addis Ababa University, 2024-08) Abrham Ayal; Fetene Mulugeta (PhD)
    Recent advancements in recurrent deep neural networks with Long Short-Term Memory (LSTM) units have resulted in significant technological advancements in artificial intelligence. Recurrent Neural Network (RNN) and LSTM analysis of the time-series data enables us to recognise long-term trends and make predictions that can improve our way of life. State-of-the-art LSTM models with massively higher ambiguity and a large number of parameters, on the other hand, have a significant advantage. Previously, they tried to solve the time series prediction problems using LSTM based architecture, but the method to solve the problems is more complex, has less accuracy, limited memory capacity, data communication bandwidth and takes high power consumption. This work mainly focused on the implementation of already trained LSTM neural network and fully functional programmable memristor-based LSTM in analog circuitry for the solving time series prediction with increased the performance and to enhance power consumption of the system. Using the voltage-based memristive circuit on the LT-spice circuit simulator and machine learning on python programming with in the combination of the activation function, and multiplier describe the time series prediction of the LSTM architecture. Each state will have some noise, and as the noise level increases, so does the prediction accuracy in analog hardware. As a result we have get results using machine learning based LSTM and programmable memristor based LSTM of the MSE and RMSE value reduced by an average of 0.0101 and 0.100 to 0.00915 and 0.095 respectively. We use real-world regression and prediction problems to demonstrate the capabilities of our system, and the programmable memristor LSTM is a promising easy-configuration, low-power, and low-latency hardware platform for edge inference.