# Shadows of quantum machine learning

**Authors:** Sofiene Jerbi, Casper Gyurik, Simon C. Marshall, Riccardo Molteni, Vedran Dunjko  
**arXiv:** [2306.00061](https://arxiv.org/abs/2306.00061)  
**Date:** 2023-05-31  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2306.00061v1/

## Key points

- The field of quantum machine learning holds great potential for solving practical problems
- A major challenge in implementing these models is the reliance on a quantum computer during the evaluation phase
- The authors propose the concept of "classical shadows" as a solution to this issue
- Classical shadows involve generating a computable approximation of the learned function from a trained model, eliminating the need for a quantum computer
- Previous works have explored this concept and suggested approaches to construct shadow models
- Training completely classical models is also considered as an alternative solution
- The authors define shadow models based on the frameworks of cryptography
- Shadow models can solve certain learning tasks that are intractable for fully classical models under widely-believed cryptography assumptions
- The likelihood or unlikelihood of all scenarios is discussed by the authors, providing valuable insights into the limitations and possibilities of using classical shadows in practice
- This research contributes to advancing our understanding of how to overcome reliance on quantum computers when evaluating trained quantum machine learning models
- The concept of classical shadows offers a promising avenue for making these models more accessible and applicable in real-world scenarios.

## Summary

The field of  holds great potential for solving practical problems using  However, a major challenge in implementing these models is the reliance on a  during the evaluation phase. To address this issue, the authors propose the concept of "classical shadows" for  This involves generating a computable approximation of the learned function from a trained model, eliminating the need for a 

While previous works have explored this concept and suggested approaches to construct shadow models, they also raise the possibility of training completely classical models as an alternative solution. In this study, the authors take a novel approach by defining shadow models based on the frameworks of  Through their research, they demonstrate that certain learning tasks that are intractable for fully classical models can be solved by shadow models under widely-believed cryptography assumptions. Furthermore, the authors discuss the likelihood or unlikelihood of all This analysis provides valuable insights into the potential limitations and possibilities of using classical shadows in practice. Overall, this research contributes to advancing our understanding of how to overcome reliance on  when evaluating trained <kd> quantum machine learning models.</ kd>The concept of classical shadows offers a promising avenue for making these models more accessible and applicable in real-world scenarios.