Published Paper Details:

A novel approach for design of low-power approximate multipliers by leveraging Long Short-Term Memory (LSTM) networks within a deep learning (DL)-based framework. Our proposed architecture, referred to as the DL -Based Approximate Multiplier (DLAM), exploits the sequence-to-sequence learning capabilities of LSTMs to enhance the efficiency of approximate multiplication in terms of both accuracy and power consumption. The DLAM model is trained on a diverse dataset, incorporating various input patterns and corresponding approximate multiplication outcomes. Through the integration of LSTM units, the model captures long-range dependencies within the input sequences, enabling more accurate predictions of approximate multiplication results. The trained DLAM exhibits superior performance in terms of both precision and energy efficiency when compared to traditional approximate multiplier designs. Furthermore, we explore optimization techniques to minimize power consumption without compromising the accuracy of multiplication results. Our test findings show that the DLAM accomplishes a significant reduction in power consumption while maintaining competitive levels of accuracy, making it a promising candidate for low-power applications in energy-constrained environments.