Mitigating Measurement-Induced Training Instability in Hybrid Quantum Neural Networks for Protein Classification

Research Question and Approach

Hybrid quantum neural networks (QNNs) typically use the expectation values of quantum measurement operators as logits for classical loss functions. The authors observe that because these expectation values are constrained to the interval [-1, 1], they create a structural mismatch with standard softmax-based cross-entropy losses, which are designed for unbounded inputs. This mismatch leads to 'measurement-induced logit contraction,' a phenomenon where the loss function loses sensitivity to logit differences, causing suppressed gradients and unstable training.

To mitigate this, the authors introduce Quantum Measurement Temperature (QMT), a learnable scalar parameter that rescales quantum measurement outputs before they are passed to the loss function. By treating the temperature as a learnable variable, the model can dynamically adjust the effective range of its logits, thereby amplifying gradient signals and improving the model's ability to distinguish between classes.

The authors demonstrate that QMT acts as a geometric decompression operator. Mathematically, they show that QMT scales the gradient magnitude by 1/T and the gradient variance by 1/T², effectively pushing the loss function out of the low-sensitivity regime. Their experiments on fluorescence microscopy protein images and multi-class variants of Fashion MNIST show that QMT consistently enhances logit separation, stabilizes training across different random initializations, and improves overall classification accuracy compared to unscaled readouts. Crucially, this approach is architecture-agnostic, meaning it can be applied to any existing hybrid QNN without requiring changes to the quantum ansatz or circuit depth.

This work addresses a fundamental, previously overlooked bottleneck in the practical deployment of hybrid quantum-classical models. By identifying that training instability often stems from the interface between quantum measurements and classical loss functions rather than just the quantum circuit itself, the authors provide a simple, plug-and-play solution. This enables more reliable training of QNNs for real-world applications, such as high-resolution protein classification in fluorescence microscopy, where data heterogeneity and noise typically exacerbate gradient suppression.