Seminar: Quantum and Quantum-Inspired Machine Learning in Practice
By Snehal Raj

Snehal Raj will give a seminar on Thursday 15 October at 2pm in room 5A126.

Quantum and Quantum-Inspired Machine Learning in Practice

Quantum computing offers new ways to represent data and construct learning models. Their value depends on both computational cost and performance on the task. This talk presents three studies that connect model design with empirical evaluation.

First, I will introduce QuIC, a method for fine-tuning large language models with few trainable parameters. Inspired by quantum circuits, it uses compound matrices to construct large orthogonal transformations from small trainable matrices. I will discuss the resulting trade-offs between parameter count and task performance. Next, I will present quantum graph neural networks that implement message passing within quantum circuits. I will explain how their structure affects graph discrimination and examine numerical results on graph-learning tasks. Finally, I will examine the “train classical, deploy quantum” approach to generative modelling. Our benchmarks show that similar values of a moment-matching loss can accompany very different coverage of unseen valid samples. This motivates direct tests of generated samples alongside the training objective.

The talk will introduce the quantum concepts as needed and discuss the evidence and limits of each approach.

Papers covered:

1. QuIC: Quantum-Inspired Compound Adapters for Parameter Efficient Fine-Tuning

2. Scalable Message-Passing Quantum Graph Neural Networks in the Weisfeiler–Leman Hierarchy

3. Train classical, deploy quantum” requires rethinking generalization