Precision Scalable Activation Functions for Optimizing Inference Energy Consumption

Abstract

This work will investigate new frameworks that can mitigate inference energy consumption in deep neural networks while keeping the predictive accuracy high in resource-constrained artificial intelligence environments, specifically developing a Precision Scalable Activation Function (PSAF) framework. Existing activation functions generally rely on fixed high-precision computations that increase computational complexity, memory overhead, inference latency, and power consumption, limiting their suitability for edge computing and Internet of Things (IoT) applications. The proposed approach combines adaptive precision-aware activation computation in a lightweight neural network architecture. The framework dynamically adapts the activation precision based on the importance of neurons, the sensitivity of the feature map, and the conditions of the inference workload. The datasets have been preprocessed using normalization, resizing, and augmentation methods, and the models have been trained through the Adam optimizer with categorical cross-entropy loss. Performance evaluation was done using hardware-aware profiling data, such as floating-point operations, inference latency, memory usage, and power usage. The experimental results showed that the proposed PSAF framework was able to obtain the highest accuracy of 97.5%, precision of 97.3%, Recall of 97.1%, and F1-score of 97.2%, respectively. The framework cut down the inference energy consumption to 6.1 J and the total power consumption by around 31.4% as compared to conventional fixed-precision activation techniques. Moreover, PSAF lowered the computational complexity to 219 MFLOPs, inference latency to 17.3ms, and memory consumption to 96MB. The results of the study indicate that adaptive precision scaling in activation functions provides a great balance between energy efficiency and prediction accuracy, making the proposed framework suitable for sustainable edge AI and low-power intelligent computing systems.