Quantum Transfer Learning
Quantum transfer learning with quantum pooling layer
Model : Quantum transfer learning with quantum pooling layer
The idea of transfer leanring is to feed the data through pre-trained feature extraction networks first, and train only a small size feed forward netwrok after it to fine tune the moedel with respect to a specific data set.
Since CIFAR100 image data are too large to directly encode in quantum circuits today, we here rely on a “imagenet” pre-trained ResNet18 as feature extraction layer. The image feature is reduced to 4 dimension through this network, and encoded into a 4 qubit circuit network.
We then utilize the idea of quantum pooling layer[2] to further reduce the quantum curcuit to a single qubit. A single qubit is sufficient for binary classification by chooing the eigenstate with higher probability.
