Speaker
Description
Current Vision-Language Models (VLMs) often fail at compositional reasoning, treating text inputs as bag of words. To address this without relying on large, classical negative-sampled datasets, we introduce QuCLIP, a language Informed quantum vision-language model.
QuCLIP maps the derivations of a linguistic framework with a transparent syntax/semantics interfact onto Variational Quantum Circuits (VQCs), treating function applications as tensor contraction executed via Bell measurements. To ensure near-term NISQ hardware viability, we utilize the map-state duality to minimize circuit depth, SWAP gate routing, and post selection overheads. For the visual modality, high-depth Amplitude Encoding is replaced by parameterized PCA and Multilayerd Perceptrion learnt quantum feature maps with data re-uploading to reduce the circuit size while mantaining accuracy.
Optimized using a Fubini-Study metric loss, QuCLIP operates with exceptional parameter efficiency, utilizing only tens of thousands of parameters. When evaluated on the ARO and SVO-Swap benchmarks, it consistently outperforms classical baselines. Evaluations across noiseless simulations and noisy physical hardware demonstrate that scaling execution environments from 100 to 1000 shots mitigates hardware noise and recovers classification accuracy or at least mantain correlation.
| Topic | Quantum algorithms for machine learning applications |
|---|---|
| Additional Topic | Quantum variational circuits and their applications |