Hi everyone,
I’m working on a hybrid quantum-classical model where I integrate a Variational Quantum Circuit (VQC) into a CNN (e.g., ResNet or VGG) for image classification tasks. Most examples I’ve seen replace the final fully connected (FC) layer with a quantum circuit, but I’m curious about the implications of placing the VQC at different positions within the CNN backbone.
Specifically, I’m exploring options like:
- Inserting the VQC after
layer2
or layer3
- Using it before the final pooling layer
- Replacing the FC head as usual
My question is:
How does the position of the VQC in the CNN affect model performance, feature representation, and gradient flow? Are there any insights from research or practical experiments on the impact of early vs late insertion?
I’d appreciate references, examples, or guidance on what others have found effective.
Thanks in advance!
Hi @quantum_expert_123 !
I want to start by saying that I am not aware of particular research addressing your question.
It seems to me like the optimal choice would depend on the specific characteristics of your classification task.
I did a quick literature review and mostly found what you said about replacing the final FC with a quantum circuit.
This paper on Quantum Convolutional Neural Networks for Multi-Channel Supervised Learning has a nice literature review. I found two other papers relevant in this context: Lean classical-quantum hybrid neural network model for image classification and Hybrid Quantum Neural Network in High-dimensional Data Classification
You could always try and experiment with the position of the VQC for your particular case and come to new conclusions!
Good luck!
1 Like
Hi @daniela.murcillo, thank you for your help! My research involves trying to apply a hybrid quantum classical CNN to medical images to see if there are any improvements over classical CNNs. Sure, I will experiment with this then, thanks again for your help!
1 Like