Google's quantum supremacy experiment used 53 noisy qubits to demonstrate it could perform a calculation in 200 seconds on a quantum computer that would take 10,000 years on the largest classical computer using existing algorithms. This marks the beginning of the Noisy Intermediate-Scale Quantum (NISQ) computing era. In the coming years, quantum devices with tens-to-hundreds of noisy qubits are expected to become a reality.
Quantum computing relies on properties of quantum mechanics to compute problems that would be out of reach for classical computers. A quantum computer uses qubits. Qubits are like regular bits in a computer, but with the added ability to be put into a superposition and share entanglement with one another.
Classical computers perform deterministic classical operations or can emulate probabilistic processes using sampling methods. By harnessing superposition and entanglement, quantum computers can perform quantum operations that are difficult to emulate at scale with classical computers. Ideas for leveraging NISQ quantum computing include optimization, quantum simulation, cryptography, and machine learning.
Quantum machine learning
Quantum machine learning (QML) is built on two concepts: quantum data and hybrid quantum-classical models.
Quantum data is any data source that occurs in a natural or artificial quantum system. This can be data generated by a quantum computer, like the samples gathered from the Sycamore processor for Google’s demonstration of quantum supremacy. Quantum data exhibits superposition and entanglement, leading to joint probability distributions that could require an exponential amount of classical computational resources to represent or store. The quantum supremacy experiment showed it is possible to sample from an extremely complex joint probability distribution of 2^53 Hilbert space.
The quantum data generated by NISQ processors are noisy and typically entangled just before the measurement occurs. Heuristic machine learning techniques can create models that maximize extraction of useful classical information from noisy entangled data. The TensorFlow Quantum (TFQ) library provides primitives to develop models that disentangle and generalize correlations in quantum data—opening up opportunities to improve existing quantum algorithms or discover new quantum algorithms.
The following are examples of quantum data that can be generated or simulated on a quantum device:
- Chemical simulation —Extract information about chemical structures and dynamics with potential applications to material science, computational chemistry, computational biology, and drug discovery.
- Quantum matter simulation —Model and design high temperature superconductivity or other exotic states of matter which exhibits many-body quantum effects.
- Quantum control —Hybrid quantum-classical models can be variationally trained to perform optimal open or closed-loop control, calibration, and error mitigation. This includes error detection and correction strategies for quantum devices and quantum processors.
- Quantum communication networks —Use machine learning to discriminate among non-orthogonal quantum states, with application to design and construction of structured quantum repeaters, quantum receivers, and purification units.
- Quantum metrology —Quantum-enhanced high precision measurements such as quantum sensing and quantum imaging are inherently done on probes that are small-scale quantum devices and could be designed or improved by variational quantum models.
Hybrid quantum-classical models
A quantum model can represent and generalize data with a quantum mechanical origin. Because near-term quantum processors are still fairly small and noisy, quantum models cannot generalize quantum data using quantum processors alone. NISQ processors must work in concert with classical co-processors to become effective. Since TensorFlow already supports heterogeneous computing across CPUs, GPUs, and TPUs, it is used as the base platform to experiment with hybrid quantum-classical algorithms.
A quantum neural network (QNN) is used to describe a parameterized quantum computational model that is best executed on a quantum computer. This term is often interchangeable with parameterized quantum circuit (PQC).
A goal of TensorFlow Quantum is to help discover algorithms for the NISQ-era, with particular interest in:
- Use classical machine learning to enhance NISQ algorithms. The hope is that techniques from classical machine learning can enhance our understanding of quantum computing. In meta-learning for quantum neural networks via classical recurrent neural networks, a recurrent neural network (RNN) is used to discover that optimization of the control parameters for algorithms like the QAOA and VQE are more efficient than simple off the shelf optimizers. And machine learning for quantum control uses reinforcement learning to help mitigate errors and produce higher quality quantum gates.
- Model quantum data with quantum circuits. Classically modeling quantum data is possible if you have an exact description of the datasource—but sometimes this isn’t possible. To solve this problem, you can try modeling on the quantum computer itself and measure/observe the important statistics. Quantum convolutional neural networks shows a quantum circuit designed with a structure analogous to a convolutional neural network (CNN) to detect different topological phases of matter. The quantum computer holds the data and the model. The classical processor sees only measurement samples from the model output and never the data itself. In Robust entanglement renormalization on a noisy quantum computer, the authors learn to compress information about quantum many-body systems using a DMERA model.
Other areas of interest in quantum machine learning include:
- Modeling purely classical data on quantum computers.
- Quantum-inspired classical algorithms.
- Supervised learning with quantum classifiers.
- Adaptive layer-wise learning for quantum neural network.
- Quantum dynamics learning.
- Generative modeling of mixed quantum states .
- Classification with quantum neural networks on near term processors.