Clifford QNN for MNIST-2 classification

Authors: Hanrui Wang, Gokul Subramanian Ravi

The model contains 16 encoder gates (4RY, 4RZ, 4RZ, 4RY) to encoder 16 pixels on 4 qubits. The encoder part is still non-clifford. The trainable part contains 5 blocks. In each block, we have 4 RX, 4 RY, 4 RZ and then 4 CNOT with ring connections.

The advantange of using Clifford QNN is that the simulation of Clifford circuits are exponentially more efficient than simulation of general quantum circuits.

TODOs

• [ ] Clifford encoder

• [ ] Improve quantization aware finetuning accuracy

Train the model in floating and then perform static quantization:

python mnist_clifford_qnn.py

Train for 20 epochs. Test results:

| Floating | Accuracy | Loss | | ———– | ———– | ——— | | Floating point | 0.868 | 0.378 | | Clifford | 0.660 | 0.648 |

Train the model in floating and then perform quantization-aware finetuning:

Using the straight-through estimation (SSE) of gradients.

python mnist_clifford_qnn.py --finetune

Train for 20 epochs and then finetune 20 epochs. Test results:

| Floating | Accuracy | Loss | | ———– | ———– | ——— | | Floating point | 0.868 | 0.378 | | Clifford | 0.722 | 0.582 |