Quantum Transfer Learning

Quantum Transfer Learning

The Hybrid Approach

Transfer learning uses a pre-trained classical model as a feature extractor and attaches a small trainable layer on top. In quantum transfer learning, that trainable layer is a parameterized quantum circuit (PQC).

The idea: classical networks (ResNet, BERT, etc.) learn rich feature representations from large datasets. These features can be re-used for new tasks by attaching a small task-specific head. Replacing this head with a quantum circuit is quantum transfer learning.

The Architecture

Classical Transfer Learning: Pre-trained CNN -> Feature vector (512D) -> Classical dense layer -> Output

Quantum Transfer Learning: Pre-trained CNN -> Feature vector (512D) -> Dimensionality reduction -> PQC (n qubits) -> Measurement -> Output

The PQC has n qubits, so we reduce the feature vector to n values (e.g., via PCA or just taking the first n features). These n values are encoded as rotation angles.

Encoding the Classical Features

Given n classical features $f_0,f_1,\dots ,f_{n-1}$ from the pre-trained model:

Only the variational layer angles ${\theta }_i$ are updated during training. The encoding angles come from the frozen pre-trained model.

Why This Might Help

The PQC variational layer operates in a $2^n$-dimensional Hilbert space. For n=4 qubits, that is 16 dimensions. The hypothesis: this larger feature space might allow better separation of classes even with few training examples.

Experimental evidence is mixed. On small datasets, quantum transfer learning sometimes matches classical dense layers. On large datasets, classical layers typically win due to hardware limitations.

Practical Considerations