Portfolio

This paper performs an in-depth study of error-resiliency of Neural Networks(NNs). Our investigation has resulted into two important findings. First, we found that Binary Neural Networks (BNNs) are more error-tolerant than 32-bits NNs. Second, in BNNs the network accuracy is more sensitive to errors in Batch Normalization Parameters(BNPs) than that in binary weights. A detailed discussion on the same is presented in the paper. Based on these findings, we propose a split memory architecture for low power BNNs, suitable for IoTs. In the proposed split memory architecture, weights are stored in area-efficient 6T SRAM, and BNPs are stored in robust 12T SRAM. The proposed split memory architecture for BNNs synthesized in UMC 28nm is highly energy efficient as the V _min (minimum operating voltage) can be reduced to 0.36 V, 0.52 V, and 0.52 V for the MNIST, CIFAR10, and ImageNet datasets respectively, with accuracy drop of less than 1%. View full paper.

Neural networks are both compute and memory intensive, and consume significant power while inferencing. Bit reduction of weights is one of the key techniques used to make them power and area efficient without degrading performance. In this paper, we show that inferencing accuracy changes insignificantly even when floating-point weights are represented using 10-bits (lower for certain other neural networks), instead of 32-bits. We have considered a set of 8 neural networks. Further, we propose a mathematical formula for finding the optimum number of bits required to represent the exponent of floating point weights, below which the accuracy drops drastically. We also show that mantissa is highly dependent on the number of layers of a neural network and propose a mathematical proof for the same. Our simulation results show that bit reduction gives better throughput, power efficiency, and area efficiency as compared to those of the models with full precision weights. View full paper.

This brief compares quantized float-point representation in posit and fixed-posit formats for a wide variety of pre-trained deep neural networks (DNNs). We observe that fixed-posit representation is far more suitable for DNNs as it results in a faster and low-power computation circuit. We show that accuracy remains within the range of 0.3% and 0.57% of top-1 accuracy for posit and fixed-posit quantization. We further show that the posit-based multiplier requires higher power-delay-product (PDP) and area, whereas fixed-posit reduces PDP and area consumption by 71% and 36%, respectively, compared to (Devnath et al., 2020) for the same bit-width. View full paper.

Traditionally, the crop analysis and agricultural production predictions were done based on statistical models. However, with the climate of the world changing to drastic degrees, these statistical models have become very ambiguous. Hence, it becomes prudent that we resort to other less vague methods. Through a traditional model, user interacts primarily with a mathematical computations and its results and helps to solve well-defined and structured problems. Whereas, in a data driven model, user interacts primarily with the data and helps to solve mainly unstructured problems. At this point, enters the concept of Machine Learning. In this work we tried to find a new approach to reduce the input feature to reduce the processing power needed. We have attempted at predicting the agricultural outputs of rice production in an area by implementing a pixel count based classification machine learning model. Through this model, we tried to predict the approximate crop yield based on NDVI values analyzed for a particular season and area. View full paper.