# Weightless Neural Networks for Efficient Edge Inference

**Authors:** Zachary Susskind, Aman Arora, Igor Dantas Dos Santos Miranda, Luis Armando Quintanilla Villon, Rafael Fontella Katopodis, Leandro Santiago de Araujo, Diego Leonel Cadette Dutra, Priscila Machado Vieira Lima, Felipe Maia Galvao Franca, Mauricio Breternitz Jr., Lizy K. John  
**arXiv:** [2203.01479](https://arxiv.org/abs/2203.01479)  
**Date:** 2022-03-03  
**Summary page:** https://www.summarizepaper.com/en/paper/2203.01479v1/

## Key points

- Weightless Neural Networks (WNNs) use table lookups for inference, while Deep Neural Networks (DNNs) rely on multiply-accumulate operations.
- WNN architectures have lower implementation costs than DNNs but typically lag behind in accuracy and have high memory requirements.
- Recent research has led to significant improvements in WNN performance.
- The authors propose a novel WNN architecture called BTHOWeN that incorporates algorithmic and architectural improvements over prior work, including counting Bloom filters, hardware-friendly hashing, and Gaussian-based nonlinear thermometer encodings.
- BTHOWeN is designed specifically for edge computing applications, providing superior latency and energy efficiency compared to comparable quantized DNNs.
- BTHOWeN reduces error by more than 40% and model size by more than 50% on average when compared with state-of-the-art WNNs across nine classification datasets.
- An FPGA-based accelerator for BTHOWeN consumes almost 80% less energy than MLP models with nearly 85% reduction in latency.
- Efficient machine learning on the edge is important, and WNNs offer a promising alternative due to their lower memory requirements and energy consumption.
- The authors' proposed BTHOWeN architecture represents a significant step forward in the development of efficient WNNs for edge computing applications.

## Summary

Weightless Neural Networks (WNNs) are a type of machine learning model that use table lookups to perform inference, as opposed to Deep Neural Networks (DNNs), which rely on multiply-accumulate operations. While WNN architectures have lower implementation costs than DNNs, they typically lag behind in accuracy for common image recognition tasks and suffer from high memory requirements. However, recent research efforts have led to significant improvements in WNN performance. In this paper, the authors propose a novel WNN architecture called BTHOWeN, which incorporates algorithmic and architectural improvements over prior work. These include counting Bloom filters, hardware-friendly hashing, and Gaussian-based nonlinear thermometer encodings to improve model accuracy while reducing area and energy consumption. BTHOWeN is designed specifically for the growing edge computing sector, providing superior latency and energy efficiency compared to comparable quantized DNNs. The authors demonstrate the effectiveness of BTHOWeN by comparing it against state-of-the-art WNNs across nine classification datasets. On average, BTHOWeN reduces error by more than 40% and model size by more than 50%. The authors also present an FPGA-based accelerator for BTHOWeN and compare its latency and resource usage against similarly accurate quantized DNN accelerators such as Multi-Layer Perceptron (MLP) and convolutional models. The proposed BTHOWeN models consume almost 80% less energy than MLP models with nearly 85% reduction in latency. The paper highlights the importance of efficient machine learning on the edge and suggests that WNNs deserve additional attention in this regard. Algorithmic and hardware improvements have driven rapid increases in DNN accuracies during the past decade; however, their large network sizes require significant computational resources that may not be feasible for edge devices with limited power budgets. In contrast, WNNs offer a promising alternative due to their lower memory requirements and energy consumption. The authors' proposed BTHOWeN architecture represents a significant step forward in the development of efficient WNNs for edge computing applications.