The qandle library supports hybrid quantum-classical machine learning workflows, particularly via integration with PyTorch. Currently, gradient computation (backward pass) is implemented in a way that leverages PyTorch's autodiff, but does not provide access to quantum-native differentiation methods.
To enhance flexibility and research opportunities, we should implement parameter-shift differentiation, which is a quantum-compatible method for computing gradients of expectation values with respect to gate parameters. This involves:
- Adding parameter-shift rules for supported gates (e.g., RX, RY, RZ, etc.).
- Integrating parameter-shift differentiation as an option for the backward pass in hybrid workflows, potentially allowing to disable PyTorch-based gradient tracking.
- Providing documentation and examples comparing parameter-shift and classical autodiff approaches.
- Benchmarking the performance and accuracy trade-offs for relevant quantum machine learning tasks.
This feature will support research into variational quantum algorithms and hybrid training strategies, making qandle more competitive with other quantum ML frameworks.
The qandle library supports hybrid quantum-classical machine learning workflows, particularly via integration with PyTorch. Currently, gradient computation (backward pass) is implemented in a way that leverages PyTorch's autodiff, but does not provide access to quantum-native differentiation methods.
To enhance flexibility and research opportunities, we should implement parameter-shift differentiation, which is a quantum-compatible method for computing gradients of expectation values with respect to gate parameters. This involves:
This feature will support research into variational quantum algorithms and hybrid training strategies, making qandle more competitive with other quantum ML frameworks.