Trainability of Dissipative Perceptron-Based Quantum Neural Networks

Area of Science:

• Quantum Computing
• Machine Learning
• Artificial Intelligence

Background:

• Quantum neural networks (QNNs) are being developed for efficient machine learning on quantum data.
• Understanding the scalability and trainability of different QNN architectures is crucial for practical applications.

Purpose of the Study:

• To rigorously analyze the gradient scaling and trainability of dissipative quantum neural networks (DQNNs).
• To investigate the potential for barren plateaus in DQNNs and provide quantitative bounds on gradient scaling.

Main Methods:

• Analysis of gradient scaling in DQNNs, a specific QNN architecture where input qubits are discarded.
• Quantitative bounding of gradient scaling under varying conditions, including different cost functions and circuit depths.

Main Results:

• DQNNs can exhibit barren plateaus, characterized by gradients vanishing exponentially with the number of qubits.
• The trainability of DQNNs is not universally guaranteed and depends on factors like cost function and circuit depth.

Conclusions:

• This work provides the first rigorous scalability analysis for a perceptron-based QNN architecture.
• The findings highlight potential limitations in the trainability of DQNNs, necessitating careful consideration for large-scale implementation.