QuIC: A Quantum-Inspired Interaction Classifier for Revitalizing Shallow CNNs in Fine-Grained Recognition

TL;DR

Fine-grained visual classification on resource-constrained devices struggles with subtle inter-class differences when using shallow backbones. The authors introduce QuIC, a quantum-inspired interaction classifier that adds a per-class symmetric interaction matrix to capture second-order feature covariances while preserving a linear path for efficiency. They formalize the approach with $E = \langle \psi | \mathcal{M} | \psi \rangle$ and $y_k = \text{BN}( \mathbf{w}_k^\top \mathbf{z} + \mathbf{z}^\top \mathbf{M}_k \mathbf{z}) + b_k$, enabling stable single-stage training without dimensionality explosion. Empirically, QuIC boosts shallow backbones such as VGG16 by nearly 20 percentage points and outperforms SE blocks on ResNet18, with Grad-CAM and t-SNE analyses confirming localized discriminative attention and compact feature clustering. This physics-inspired, plug-and-play module offers an efficient path to high FGVC performance on edge devices.

Abstract

Deploying deep learning models for Fine-Grained Visual Classification (FGVC) on resource-constrained edge devices remains a significant challenge. While deep architectures achieve high accuracy on benchmarks like CUB-200-2011, their computational cost is often prohibitive. Conversely, shallow networks (e.g., AlexNet, VGG) offer efficiency but fail to distinguish visually similar sub-categories. This is because standard Global Average Pooling (GAP) heads capture only first-order statistics, missing the subtle high-order feature interactions required for FGVC. While Bilinear CNNs address this, they suffer from high feature dimensionality and instability during training. To bridge this gap, we propose the Quantum-inspired Interaction Classifier (QuIC). Drawing inspiration from quantum mechanics, QuIC models feature channels as interacting quantum states and captures second-order feature covariance via a learnable observable operator. Designed as a lightweight, plug-and-play module, QuIC supports stable, single-stage end-to-end training without exploding feature dimensions. Experimental results demonstrate that QuIC significantly revitalizes shallow backbones: it boosts the Top-1 accuracy of VGG16 by nearly 20% and outperforms state-of-the-art attention mechanisms (SE-Block) on ResNet18. Qualitative analysis, including t-SNE visualization, further confirms that QuIC resolves ambiguous cases by explicitly attending to fine-grained discriminative features and enforcing compact intra-class clustering.

Figures (4)

• Figure 1: Efficiency Analysis: Accuracy vs. Training Time. QuIC achieves the best trade-off, offering superior accuracy with minimal computational overhead compared to attention-based baselines.
• Figure 2: Comparison of misclassification counts for the top confused pairs using the ResNet18 backbone. QuIC (Green) demonstrates superior discriminative ability on highly similar species, significantly reducing errors on challenging pairs such as "American Crow vs. Common Raven", "Arctic Tern vs. Common Tern", and "Chuck-will's-widow vs. Common Poorwill" compared to GAP (Red) and SE (Blue) baselines.
• Figure 3: Visualization of "Exclusive Wins" for the (Top) American Crow and (Bottom) Common Raven. While baselines misclassify these samples (e.g., mistaking the Crow for a Cowbird and the Raven for a Crow) due to diffuse attention, QuIC successfully localizes fine-grained discriminative parts—specifically the beak shape and throat hackles—leading to correct predictions.
• Figure 4: t-SNE visualization of feature spaces on the CUB-200-2011 test set using the VGG16 backbone. Compared to the diffuse, overlapping clusters in (a) FC and (b) GAP, and the ambiguous boundaries in (c) SE-Block, (d) QuIC produces significantly more compact clusters with clear inter-class margins. This demonstrates QuIC's ability to revitalize the shallow backbone by enforcing highly discriminative feature learning.