Quantum computing’s potential to revolutionize AI depends on growth of a developer ecosystem in which suitable tools, skills, and platforms are in abundance. To be considered ready for enterprise production deployment, the quantum AI industry would have to, at the very least, reach the following key milestones:
- Find a compelling application for which quantum computing has a clear advantage over classical approaches to building and training AI.
- Converge on a widely adopted open source framework for building, training, and deploying quantum AI.
- Build a substantial, skilled developer ecosystem of quantum AI applications.
These milestones are all still at least a few years in the future. What follows is an analysis of the quantum AI industry’s maturity at the present time.
Lack of a compelling AI application for which quantum computing has a clear advantage
Quantum AI executes ML (machine learning), DL (deep learning), and other data-driven AI algorithms reasonably well.
As an approach, quantum AI has moved well beyond the proof-of-concept stage. However, that’s not the same as being able to claim that quantum approaches are superior to classical approaches for executing the matrix operations upon which AI’s inferencing and training workloads depend.
Where AI is concerned, the key criterion is whether quantum platforms can accelerate ML and DL workloads faster than computers built entirely on classical von Neumann architectures. So far there is no specific AI application that a quantum computer can perform better than any classical alternative. For us to declare quantum AI a mature enterprise technology, there would need to be at least a few AI applications for which it offers a clear advantage—speed, accuracy, efficiency—over classical approaches to processing these workloads.
Nevertheless, pioneers of quantum AI have aligned its functional processing algorithms with the mathematical properties of quantum computing architectures. Currently, the chief algorithmic approaches for quantum AI include:
- Amplitude encoding: This associates quantum-state amplitudes with the inputs and outputs of computations performed by ML and DL algorithms. Amplitude encoding allows for statistical algorithms that support exponentially compact representation of complex multidimensional variables. It supports matrix inversions in which the training of statistical ML models reduces to solving linear systems of equations, such as those in least-squares linear regressions, least-squares version of support vector machines, and Gaussian processes. It often requires the developer to initialize a quantum system in a state whose amplitudes reflect the features of the entire data set.
- Amplitude amplification: This uses an algorithm that finds with high probability the unique input to a black box function that produces a particular output value. Amplitude amplification is suitable for those ML algorithms that can be translated into an unstructured search task, such as k-medians and k-nearest neighbors. It can be accelerated through random walk algorithms where randomness comes from stochastic transitions between states, such as in that inherent to quantum superposition of states and the collapse of wave functions due to state measurements.
- Quantum annealing: This determines the local minima and maxima of a machine-learning function over a given set of candidate functions. It starts from a superposition of all possible, equally weighted states of a quantum ML system. It then applies a linear, partial differential equation to guide the time evolution of the quantum-mechanical system. It eventually yields an instantaneous operator, known as the Hamiltonian, that corresponds to the sum of the kinetic energies plus the potential energies associated with the quantum system’s ground state.
Leveraging these techniques, some current AI implementations use quantum platforms as coprocessors on select calculation workloads, such as autoencoders, GANs (generative adversarial networks), and reinforcement learning agents.
As quantum AI matures, we should expect that these and other algorithmic approaches will show a clear advantage when applied to AI grand challenges that involve complex probabilistic calculations operating over highly multidimensional problem domains and multimodal data sets. Examples of heretofore intractable AI challenges that may yield to quantum-enhanced approaches include neuromorphic cognitive models, reasoning under uncertainty, representation of complex systems, collaborative problem solving, adaptive machine learning, and training parallelization.
But even as quantum libraries, platforms, and tools prove themselves out for these specific challenges, they will still rely on classical AI algorithms and functions within end-to-end machine learning pipelines.
Lack of a widely adopted open source modeling and training framework
For quantum AI to mature into a robust enterprise technology, there will need to be a dominant framework for developing, training, and deploying these applications. Google’s TensorFlow Quantum is an odds-on favorite in that regard. Announced this past March, TensorFlow Quantum is a new software-only stack that extends the widely adopted TensorFlow open source AI library and modeling framework.
TensorFlow Quantum brings support for a wide range of quantum computing platforms into one of the dominant modeling frameworks used by today’s AI professionals. Developed by Google’s X R&D unit, it enables data scientists to use Python code to develop quantum ML and DL models through standard Keras functions. It also provides a library of quantum circuit simulators and quantum computing primitives that are compatible with existing TensorFlow APIs.
Developers can use TensorFlow Quantum for supervised learning on such AI use cases as quantum classification, quantum control, and quantum approximate optimization. They can execute advanced quantum learning tasks such as meta-learning, Hamiltonian learning, and sampling thermal states. They can use the framework to train hybrid quantum/classical models to handle both the discriminative and generative workloads at the heart of the GANs used in deep fakes, 3D printing, and other advanced AI applications.
Recognizing that quantum computing is not yet mature enough to process the full range of AI workloads with sufficient accuracy, Google designed the framework to support the many AI use cases with one foot in traditional computing architectures. TensorFlow Quantum enables developers to rapidly prototype ML and DL models that hybridize the execution of quantum and classic processors in parallel on learning tasks. Using the tool, developers can build both classical and quantum datasets, with the classical data natively processed by TensorFlow and the quantum extensions processing quantum data, which consists of both quantum circuits and quantum operators.
Google designed TensorFlow Quantum to support advanced research into alternative quantum computing architectures and algorithms for processing ML models. This makes the new offering suitable for computer scientists who are experimenting with different quantum and hybrid processing architectures optimized for ML workloads.
To this end, TensorFlow Quantum incorporates Cirq, an open source Python library for programming quantum computers. It supports programmatic creation, editing, and invoking of the quantum gates that constitute the Noisy Intermediate Scale Quantum (NISQ) circuits characteristic of today’s quantum systems. Cirq enables developer-specified quantum computations to be executed in simulations or on real hardware. It does this by converting quantum computations to tensors for use inside TensorFlow computational graphs. As an integral component of TensorFlow Quantum, Cirq enables quantum circuit simulation and batched circuit execution, as well as estimation of automated expectation and quantum gradients. It also enables developers to build efficient compilers, schedulers, and other algorithms for NISQ machines.
In addition to providing a full AI software stack into which quantum processing can now be hybridized, Google is looking to expand the range of more traditional chip architectures on which TensorFlow Quantum can simulate quantum ML. Google also announced plans to expand the range of custom quantum-simulation hardware platforms supported by the tool to include graphics processing units from various vendors as well as its own Tensor Processing Unit AI-accelerator hardware platforms.
Google’s latest announcement lands in a fast-moving but still immature quantum computing marketplace. By extending the most popular open source AI development framework, Google will almost certainly catalyze use of TensorFlow Quantum in a wide range of AI-related initiatives.
However, TensorFlow Quantum comes into a market that already has several open source quantum-AI development and training tools. Unlike Google’s offering, these rival quantum AI tools come as parts of larger packages of development environments, cloud services, and consulting for standing up full working applications. Here are three full-stack quantum AI offerings:
- Azure Quantum, announced in November 2019, is a quantum-computing cloud service. Currently in private preview and due for general availability later this year, Azure Quantum comes with a Microsoft open-sourced Quantum Development Kit for the Microsoft-developed quantum-oriented Q# language as well as Python, C#, and other languages. The kit includes libraries for development of quantum apps in ML, cryptography, optimization, and other domains.
- Amazon Braket, announced in December 2019 and still in preview, is a fully managed AWS service. It provides a single development environment to build quantum algorithms, including ML, and test them on simulated hybrid quantum/classical computers. It enables developers to run ML and other quantum programs on a range of different hardware architectures. Developers craft quantum algorithms using the Amazon Braket developer toolkit and use familiar tools such as Jupyter notebooks.
- IBM Quantum Experience is a free, publicly available, cloud-based environment for team exploration of quantum applications. It provides developers with access to advanced quantum computers for learning, developing, training, and running AI and other quantum programs. It includes IBM Qiskit, an open source developer tool with a library of cross-domain quantum algorithms for experimenting with AI, simulation, optimization, and finance applications for quantum computers.
TensorFlow Quantum’s adoption depends on the extent to which these and other quantum AI full-stack vendors incorporate it into their solution portfolios. That seems likely, given the extent to which all these cloud vendors already support TensorFlow in their respective AI stacks.
TensorFlow Quantum won’t necessarily have the quantum AI SDK field all to itself going forward. Other open source AI frameworks—most notably, the Facebook-developed PyTorch—are contending with TensorFlow for the hearts and minds of working data scientists. One expects that rival framework to be extended with quantum AI libraries and tools during the coming 12 to 18 months.
Read more on the next page...