I’m working on a hybrid quantum-classical neural network that augments a classical CNN with a variational quantum circuit (VQC) for image classification for a medical image dataset.
Model architecture (high-level):
- Input image is first processed by a classical CNN (convolution +
pooling layers). - Features are passed through fully connected layers to reduce
dimensionality. - Instead of going directly to a classical output layer, the feature
vector is encoded into a quantum state and passed to a parameterized
quantum circuit (VQC). - The VQC performs single-qubit rotations and entangling operations,
and the measurement outcomes (expectation values) are used as the
final feature vector for classification.
Problem:
Despite experimenting with different configurations, the model always predicts only a single class (e.g., always class 0), regardless of the input image. This happens consistently during training and evaluation.
What I’ve tried:
- Verified that class imbalance is not the issue, the dataset is relatively balanced.
Varying:
-
Number of qubits (4–8)
-
Circuit depth (reps=1, 2)
-
Entanglement pattern (full, circular, linear)
-
Training with different learning rates, and batch sizes.
My question:
What could be causing this hybrid model to collapse to a single-class prediction, and what strategies can I try to debug or stabilize training in this hybrid quantum-classical setup?
Are there known pitfalls in training VQC-based classifiers when integrated into classical CNNs (e.g., vanishing gradients, poor initialization, circuit expressivity issues)?