Week 7: The Noise Sweep

And here we are with a new entry in the series, right back from my Spring Break hiatus! The looming deadline for the final exhibition of this senior project looms far closer than I would like to admit at the moment. After trying to implement my very own fancy quantum circuit right up until the spring break started, and failing miserably, I decided to take a step back and start experimenting with putting external stress on my QNN instead of tinkering with the circuit design. So far, I’ve only been running simulations where the QNN and fair classical model compete in the face of a single level of AWGN noise. While that was enough to convince me that my hypothesis was correct, the true proof would come from watching the classical system fail as the noise increases.

For this week’s simulation run, I created a noise sweep test on Amazon SageMaker that progressively adds noise multipliers to the image of my traffic lights, starting from the slightest amount of fuzz all the way up to a full-blown whiteout of static on each of the images in my dataset, and then I ran those images through both networks. And guess what? Just as I hoped, the results showed a clear superiority of the quantum model even without any additional fancy modifications. At low noise levels, the QNN performed about the same as the fair CNN. However, once the noise multiplier grew above 3 or so, the performance of the classical model tanked, as the algorithm could no longer tell apart red from green because the pixel core got so corrupted by the added noise. On the contrary, the QNN degraded gracefully thanks to the fact that, since the qubits are entangled, they seem to evaluate the connection between all four pixels within the traffic light in their classification decision.

With this great progress made, my next task will be learning how to distill all these quantum physics equations into plain text for the discussion section of my research paper. Until next time!

Patrick Zhou, signing off.