Co-Design of Approximate Multilayer Perceptron for Ultra-Resource Constrained Printed Circuits

Abstract

Printed Electronics (PE) exhibits on-demand, extremely lowcost hardware due to its additive manufacturing process, enabling machine learning (ML) applications for domains that feature ultra-low cost, conformity, and non-toxicity requirements that silicon-based systems cannot deliver. Nevertheless, large feature sizes in PE prohibit the realization of complex printed ML circuits. In this work, we present, for the first time, an automated printed-aware software/hardware co-design framework that exploits approximate computing principles to enable ultra-resource constrained printed multilayer perceptrons (MLPs). Our evaluation demonstrates that, compared to the state-of-the-art baseline, our circuits feature on average 6x (5.7x) lower area (power) and less than 1% accuracy loss.

Cross-Layer Approximation For Printed Machine Learning Circuits

“…Label Α I 1 [3] I 2 [6] I 2 [6] Label B implementation of parallel unary printed Decision Trees. As aforementioned, the unmatched customization in printed circuits enables hardcoding the model's parameters.…”

“…2 presents an example architecture of a Decision Tree when its inputs are available in a parallel unary representation. Instead of traditional comparisons, bespoke unary DTs can now be viewed as a simple logic over a set of unary digits corresponding to the trained parameters (e.g., I 1 [3], I 4 [2], I 2 [6]). The truth table for each label of the unary Decision Tree example is also depicted.…”

“…Our secondary objective is to select more power-efficient values/parameters for each comparison (i.e., optimize the order of the output digits in the ADC). Since feasibility is the foremost requirement for printed ML circuits, prioritizing it over strict accuracy constraints is a typical procedure [3]. Consequently, we also explore the trade-off between some accuracy degradation and the potential for additional hardware gains.…”

“…While printed circuits exhibit the potential to meet the demands of ultra-low cost (even sub-cent levels) due to their inexpensive additive manufacturing processes, the relatively large feature sizes, inherent in printed electronics, pose a challenge in realizing large-scale integration circuits for onsensor processing. This becomes particularly relevant for Machine Learning (ML) classification circuits, as classification is the primary task for numerous applications in aforementioned domains [2,3]. Such applications have to classify (analog) data collected from printed sensors to extract useful information.…”

“…While bespoke design indeed yields significant area and power gains, further improvements are mandatory to enable printed-battery-powered ML classifiers. The works in [3]- [5] and [6] combined bespoke design with approximate [8] and stochastic [9] computing, respectively, targeting Multilayer Perceptrons and/or Support Vector Machines. However, in most cases, the obtained gains are either inadequate or come at the cost of a considerable accuracy degradation.…”