# Opening the black box of deep learning

**Authors:** Dian Lei, Xiaoxiao Chen, Jianfei Zhao  
**arXiv:** [1805.08355](https://arxiv.org/abs/1805.08355)  
**Date:** 2018-05-22  
**Summary page:** https://www.summarizepaper.com/en/paper/1805.08355v1/

## Key points

- The paper "Opening the Black Box of Deep Learning" explores the internal mechanisms of deep learning and proposes that neural networks can be viewed as physical systems.
- The authors examine deep learning from three different perspectives: microscopic, macroscopic and physical world views.
- Principles from physics such as quantum mechanics and statistical physics are used to answer multiple theoretical puzzles in deep learning.
- Calculation methods for convolution calculation, pooling, normalization and Restricted Boltzmann Machine are presented from a physics perspective.
- The paper explains why deep learning must be deep, what characteristics are learned in deep learning, why Convolutional Neural Networks do not have to be trained layer by layer and the limitations of deep learning.
- The authors propose a theoretical direction and basis for the further development of deep learning now and in the future.
- Understanding its internal mechanism also provides meaningful insights into understanding human brain mechanism.
- This paper sheds light on the internal workings of deep learning through a unique perspective that combines both computer science and physics principles.

## Summary

The paper "Opening the Black Box of Deep Learning" by Dian Lei, Xiaoxiao Chen and Jianfei Zhao explores the internal mechanisms of deep learning and proposes that neural networks can be viewed as physical systems. The authors examine deep learning from three different perspectives: microscopic, macroscopic and physical world views. By using principles from physics such as quantum mechanics and statistical physics they answer multiple theoretical puzzles in deep learning. The paper presents calculation methods for convolution calculation, pooling, normalization and Restricted Boltzmann Machine from a physics perspective. It also explains why deep learning must be deep, what characteristics are learned in deep learning, why Convolutional Neural Networks do not have to be trained layer by layer and the limitations of deep learning. The authors propose a theoretical direction and basis for the further development of deep learning now and in the future. The success of deep learning has shown that its technology contains profound truth with important implications for its development and effective application in various fields. Understanding its internal mechanism also provides meaningful insights into understanding human brain mechanism. However most theoretical research on deep learning is based on mathematics. This paper proposes a new approach to understanding neural networks by viewing them as physical systems based on scientific theory from physics. Overall this paper sheds light on the internal workings of deep learning through a unique perspective that combines both computer science and physics principles. It offers valuable insights into how we can continue to develop this technology in the future while also providing a better understanding of how our own brains work.