Tunable Floating-Point for Artificial Neural Networks

Abstract

Approximate computing has emerged as a promising approach to energy-efficient design of digital systems in many domains such as digital signal processing, robotics, and machine learning. Numerous studies report that employing different data formats in Deep Neural Networks (DNNs), the dominant Machine Learning approach, could allow substantial improvements in power efficiency considering an acceptable quality for results. In this work, the application of Tunable Floating-Point (TFP) precision to DNN is presented. In TFP different precisions for different operations can be set by selecting a specific number of bits for significand and exponent in the floating-point representation. Flexibility in tuning the precision of given layers of the neural network may result in a more power efficient computation.

Tunable Floating-Point Adder

“…The NN is used to interpolate a function y = f (x) approximating the distribution of some sample points in the XY-plane. More detail on the example is given in [10].…”

“…We apply TFP to the operations in the NN of Fig. 14 for both training and inference, and we evaluate the trade off precision vs. error and power dissipation. Differently from [10], both the TFP multiplier (TFP-mul) and the TFP-add implement round-to-the-nearest unbiased rounding (RTNE). Table 6 reports the trade off for training the NN with a "cosine-like" distribution for several TFP precisions.…”

“…In the next experiment, we estimate the performance of several combinations of the precision m={24, 16, 8} (e=8) for the classification of points within a given distance from the interpolated curve/function (more detail is given in [10]). Due to the approximation error (value of weights), out of 1,000 randomly distributed points in the XY-plane, 55 result miss-classified (about 5%).…”