Energy Efficient BNN Accelerator using CiM and a Time-Interleaved Hadamard Digital GRNG in 22nm CMOS

In recent years, Neural Networks (NNs) have achieved tremendous success in a variety of fields, such as computer vision, natural language processing, speech recognition, autonomous driving, and healthcare [1] –[4]. However, conventional NNs rely heavily on large, labeled training datasets, which can lead to overfitting and overconfident decision making (especially when faced with unfamiliar, out-of-distribution inputs) [1] –[4]. Unlike conventional NNs, the weights of Bayesian Neural Network (BNNs) are represented by probability distributions, providing a mathematical framework to quantify the uncertainties in a model’s final prediction [3]. These uncertainty estimates allow BNNs to mitigate overfitting issues, enable training with smaller datasets, and help increase overall model accuracy through the implicit use of stochastic rounding [5]. Figure 1 shows an example with a BNN version of LeNet-5 [6], where weights of the network are represented by Gaussian distributions. The ambiguous digit is classified as a ‘5’ and ‘3’ with probabilities of 80% and 20%, respectively. However, these uncertainty estimates come at great computational cost, as multiple forward inference passes are required to generate the necessary posterior distributions. As such, BNNs require not only efficient, high-performance multiply-accumulation (MAC), but an efficient Gaussian Random Number Generator (GRNG) with high-quality statistics as well.