ULTRA-LOW-POWER AI ARCHITECTURES FOR ERROR-RESILIENT MULTIMODAL GENOMIC AND BIOMEDICAL SIGNAL ANALYSIS

Abstract

Approximate computing techniques have been developed due to the growing need to achieve computational efficiency in error-aware applications like machine learning, image processing, and signal processing. Traditional exact computing models are accurate, but have a high power consumption, area requirement, and delays. To overcome these issues, the current paper suggests an ultra-efficient approximate multiplier having an inbuilt error compensation mechanism. The multiplier can also adaptively compensate for these errors due to approximation by adding an error compensation algorithm that delivers the accuracy without compromising efficiency. This approach defeats the shortcomings of current approximate multipliers which are normally affected by high preciseness loss. The suggested design consumes much less power, area and delay with an acceptable trade-off between precision and computational efficiency. Through experimental evidence, it is shown that the error resilient approximate multiplier is capable of performing better in energy efficiency and speed hence a practical solution to low power high-performance applications in the current computing systems. The proposed design is validated using MATLAB based simulation