# Enabling Incremental Training with Forward Pass for Edge Devices

arXiv: https://arxiv.org/abs/2103.14007 · HTML: https://mlsanity.com/paper/2103.14007/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 7 of its 16 references that are in the map.
- First posted on arXiv 25 Mar 2021.

## Builds on (most cited first, 7 of 7)

- [Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding](https://mlsanity.com/paper/1510.00149/) (2015)
- [Evolution Strategies as a Scalable Alternative to Reinforcement Learning](https://mlsanity.com/paper/1703.03864/) (2017)
- [Quantized Neural Networks: Training Neural Networks with Low Precision Weights and Activations](https://mlsanity.com/paper/1609.07061/) (2016)
- [Fast inference of deep neural networks in FPGAs for particle physics](https://mlsanity.com/paper/1804.06913/) (2018)
- [Low-Memory Neural Network Training: A Technical Report](https://mlsanity.com/paper/1904.10631/) (2019)
- [FINN-R: An End-to-End Deep-Learning Framework for Fast Exploration of Quantized Neural Networks](https://mlsanity.com/paper/1809.04570/) (2018)
- [On the Relationship Between the OpenAI Evolution Strategy and Stochastic Gradient Descent](https://mlsanity.com/paper/1712.06564/) (2017)

## Abstract (arXiv)

Deep Neural Networks (DNNs) are commonly deployed on end devices that exist in constantly changing environments. In order for the system to maintain it's accuracy, it is critical that it is able to adapt to changes and recover by retraining parts of the network. However, end devices have limited resources making it challenging to train on the same device. Moreover, training deep neural networks is both memory and compute intensive due to the backpropagation algorithm. In this paper we introduce a method using evolutionary strategy (ES) that can partially retrain the network enabling it to adapt to changes and recover after an error has occurred. This technique enables training on an inference-only hardware without the need to use backpropagation and with minimal resource overhead. We demonstrate the ability of our technique to retrain a quantized MNIST neural network after injecting noise to the input. Furthermore, we present the micro-architecture required to enable training on HLS4ML (an inference hardware architecture) and implement it in Verilog. We synthesize our implementation for a Xilinx Kintex Ultrascale Field Programmable Gate Array (FPGA) resulting in less than 1% resource utilization required to implement the incremental training.