Halvtidsseminarium - Shiming Li: "Mitigating Ineffectual Computations for Efficient Neural Network Inference on Ultra-Low-Power MCUs"

Granskare: Christian Rohner

Abstract: Deep neural networks enable low-latency complex data processing at the network edge, but their execution efficiency is crucial for energy preservation for ultra-low-power devices. In this seminar, we discuss a problem in neural network applications on microcontroller-level devices: they often include ineffectual computations due to the activation function filtering out negative values.

We then discuss two techniques to enable sub-neuron level early-exits to mitigate this problem, whereby we only change how the NN inference code is generated, without modifying the NN model or the hardware.