# Hebbian Deep Learning Without Feedback

**Authors:** Adrien Journé, Hector Garcia Rodriguez, Qinghai Guo, Timoleon Moraitis  
**arXiv:** [2209.11883](https://arxiv.org/abs/2209.11883)  
**Date:** 2022-09-23  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2209.11883v1/

## Key points

- Authors address limitations of backpropagation (BP) in neural network training
- Proposed approach called multilayer SoftHebb overcomes challenges without feedback, target, or error signals
- SoftHebb leverages Hebbian learning principles in soft winner-take-all networks
- Eliminates weight transport, non-local plasticity, time-locking of layer updates, iterative equilibria, and supervisory feedback signals for efficiency
- Maintains high accuracy levels while enhancing efficiency and biological plausibility
- Demonstrates superior performance on popular datasets like MNIST, CIFAR-10, STL-10, and ImageNet
- Achieves accuracies reaching 99.4%, 80.3%, 76.2%, and 27.3% respectively with up to five hidden layers and a linear classifier
- Challenges conventional wisdom by showcasing that a few layers may be sufficient for effective neural network training
- Represents a paradigm shift from traditional backpropagation methods towards more biologically inspired and accurate alternatives

## Summary

In their paper titled "Hebbian Deep Learning Without Feedback," authors Adrien Journé, Hector Garcia Rodriguez, Qinghai Guo, and Timoleon Moraitis address the limitations of backpropagation (BP) in neural network training. Recent approximations have improved computational efficiency and biological compatibility but significant drawbacks remain. The authors propose a novel approach called multilayer SoftHebb to overcome these challenges. This algorithm trains deep neural networks without relying on feedback, target, or error signals by leveraging Hebbian learning principles in soft winner-take-all networks. SoftHebb eliminates weight transport, non-local plasticity, time-locking of layer updates, iterative equilibria, and supervisory feedback signals present in traditional methods to achieve efficiency. The key innovation of SoftHebb lies in its ability to maintain high accuracy levels while enhancing efficiency and biological plausibility. It demonstrates superior performance on popular datasets such as MNIST, CIFAR-10, STL-10 and ImageNet with accuracies reaching 99.4%, 80.3%, 76.2%, and 27.3% respectively when incorporating up to five hidden layers and a linear classifier. This groundbreaking approach challenges the conventional wisdom surrounding deep learning by showcasing that a few layers may be sufficient for effective neural network training. Overall, SoftHebb represents a paradigm shift from traditional backpropagation methods by offering a more biologically inspired and accurate alternative for machine learning tasks. The research conducted by Journé et al sheds light on the potential of Hebbian learning principles to revolutionize deep learning practices and pave the way for more efficient and effective neural network training strategies in the future.