SP-5: Quanvolution

Summary: This architectural solution replaces the first several layers in classical convolutional neural networks by parameterised quantum circuits operating in parallel or sequentially. The circuit has a relatively small number of qubits and operates as a convolutional filter.

Figure 1. Graphical representation of the quanvolution pattern.

Problem: The challenge is to adapt the quantum feature extraction pattern (as detailed in the previous pattern) for high-dimensional data, given the constraints of NISQ-era quantum computers.

Solution: A small quantum circuit with trainable parameters is used as a convolutional filter, replacing the first several layers of a convolutional neural network. The quantum circuit sequentially scans over the input data tensor, processing a small data window at a time.

Benefits:

  • Scalability for NISQ computers. Regardless of the dimensionality of the input data, implementations of this pattern require shallow quantum circuits with a small number of qubits, making the system compatible with most existing NISQ computers. This pattern enables the system to process high-resolution images on NISQ devices.
  • Applicable to spatial multi-dimensional data The quanvolutional filter processes data locally while preserving information on spatial order and dimensionality, even when quantum circuits with 1D input are used (Domingo et al. 2023, Onuoha et al. 2022).
  • Reduced number of parameters and space complexity. Replacing the first several layers of a convolutional neural network by quantum circuits reduces the number of parameters without compromising the expressibility of the model (Domingo et al. 2023, Hai and Bao 2023; Chao et al. 2022).
  • Distributability This pattern allows for parallelization, as the quanvolution filters can operate simultaneously, share the same parameters, and are interconnected via classical channels.
  • Trainability There were no indications of the Barren Plateau issue for this pattern, as the quantum circuits used are mostly shallow and have few trainable parameters.

Drawbacks:

  • Portability and deployability This pattern requires frequent data exchange transactions between classical and quantum components. As a result, the system can only be deployed on computers where quantum processing units, CPUs and/or GPU are closely integrated. If the components are designed to exchange data over a network, a high-bandwidth, fast connection must be established. Therefore, a highly specialised hardware environment must be established for this pattern to function effectively.
  • Efficiency Using a quantum circuit as a convolutional filter requires multiple encodings and measurements, which can become a performance bottleneck.

Known uses:

  • (Prabhu et al. 2023) have applied the quanvolutional pattern to perform multi-class classification for cardiovascular diseases.
  • (Domingo et al. 2023) utilised the quanvolutional pattern as a part of a quantum-classical convolutional neural network developed for predicting binding affinity in drug design.
  • (Baek et al. 2023) applied this pattern for point cloud data processing in classification applications.