Quantum tunneling search could accelerate optimisation on photonic chips for machine learning

Quantum Tunneling Search (QTS) was experimentally demonstrated on a photonic chip; utilising femtosecond direct laser writing they mapped non-convex functions onto a waveguide Hamiltonian. This allowed successful identification of both local and global minima within complex landscapes, a key challenge in machine learning optimisation. By introducing a varying Hamiltonian approach, the team further improved search efficiency over classical methods like stochastic gradient descent, verifying QTS demonstrates quantum speedups for these problems.

The team experimentally demonstrated Quantum Tunneling Search (QTS) using light on a chip; they mapped complex problems onto a system mimicking quantum behaviour within tiny optical pathways called waveguides. This new approach successfully identified both low points and the absolute lowest point, the global minimum, within challenging landscapes often encountered in machine learning optimisation tasks.

Imagine trying to find the lowest point in a hilly terrain with many valleys, it’s easy for traditional search methods to get stuck in a small valley instead of finding the deepest one. QTS uses ‘quantum tunneling’, where particles can pass through barriers than going over them, allowing faster exploration of these challenging landscapes often found in machine learning tasks.

The team used femtosecond direct laser writing, like using an incredibly precise 3D printer at extremely short timescales, to craft this network of light paths and verify quantum speedups compared to conventional approaches; but will this technology scale beyond proof-of-concept experiments.

Photonic Waveguide Arrays Enable High Precision Global Minimum Searches

Quantum Tunneling Search (QTS) achieved an average probability distribution highlighting local minima across both one- and two-dimensional fields utilising fabricated waveguide arrays spaced precisely 10 micrometres apart. Previously unattainable precision stemmed from accurately mapping complex functions onto physical systems, a feat now realised. Adjusting potential well depths during quantum evolution, a varying Hamiltonian approach, further improved search efficiency by enabling identification of global minima within these non-convex fields; this is a task classical stochastic gradient descent struggles with.

The arrays were arranged as 29×1 for 1D searches and 21×9 for 2D optimisation problems. Mapping an objective function onto the chip’s potential was achieved via femtosecond direct laser writing, controlling refractive index with remarkable precision. Further refinement involved ‘local annealing’, where initial energy states spread within isolated wells before being used as input for full field exploration. This resulted in more efficient searching than classical stochastic gradient descent (SGD). Experiments using 10μm-spaced 29×1 and 21×9 waveguide arrays showed average probability distributions highlighting local minima of various 1D and 2D non-convex landscapes.

Analogue quantum simulation successfully addresses optimisation despite lacking clear advantages

This demonstration of Quantum Tunneling Search offers exciting potential as an alternative to gradient descent, yet important details remain elusive; notably, the abstract avoids quantifying how much faster QTS is compared to established methods like Stochastic Gradient Descent (SGD). Assessing scalability beyond these initial experiments proves difficult without concrete benchmarks detailing performance gains across specific problem instances. Still, even absent definitive speed comparisons against advanced classical methods, this demonstration matters sharply.

It confirms complex optimisation problems can be tackled using near-term analogue quantum simulators. These devices manipulate light rather than individual atoms and deliver experimental evidence supporting theoretical predictions about how quantum mechanics could outperform conventional computing.

This physical implementation involved mapping objective functions onto tiny optical pathways called waveguides through femtosecond direct laser writing; akin to highly precise 3D printing at incredibly short timescales, it overcame previous limitations hindering accurate function-to-system mappings. Consequently, this approach opens avenues for exploring more complex optimisation landscapes beyond those accessible with classical methods, a significant step forward in computational problem solving.

This research demonstrated Quantum Tunneling Search (QTS) using photonic waveguide arrays, specifically 29x 1 and 21x 9 arrangements created with femtosecond direct laser writing, as an alternative method for non-convex optimization problems. The experiments showed QTS successfully identified local minima within one- and two-dimensional objective functions, and further refinement allowed the algorithm to locate global minimums as well.

This work experimentally confirms that QTS can function as an efficient solver, offering performance gains over Stochastic Gradient Descent. It suggests near-term analog quantum simulators have a practical role in machine learning by providing a physical system to explore complex computational landscapes.

More like this

Quantum Computing NewsTechnischen Hochschule Deggendorf’s Institute Future Technologies leads quantum computing consortium

Quantum Machine LearningQuantum control schemes could boost fragile qubits with new algorithms

Artificial IntelligenceMultiverse Computing tests quantum neural networks in a 30B LLM

Quantum AlgorithmsQuantum neural networks may need fewer cloud jobs with new method