TensorFlow Quantum: Marrying machine learning with quantum computing

TensorFlow Quantum: Marrying machine learning with quantum computing

Google open-sourced its new quantum computing machine learning library called TensorFlow Quantum (TFQ).

Introduction

On March 9th, 2020, Google AI , announced the release of TensorFlow Quantum (TFQ), an open-source library for the rapid prototyping of quantum machine learning models. They were not alone in this development effort, as they had help from the University of Waterloo and Volkswagen .

This new framework provides quantum computing (QC) researchers the software and design tools necessary for bringing the power and possibilities of machine learning to the realm of quantum computin g.

This is a natural progression of their offering in the quantum computing software platforms. In March 2018, Google announced Cirq , an open-source framework for noisy-intermediate-scale quantum ( NISQ )algorithms. At the most basic level, TFQ integrates Cirq with the widely-popular deep learning framework TensorFlow , and offers powerful classes, data structures, and methods for the design and implementation of hybrid quantum-classical models. On one hand, it features quantum computing abstractions that are compatible with the TensorFlow APIs , and on the other hand, boasts of ready-to-use, high-performance quantum circuit simulators.

But what kind of problems it will help solve and what does it focus on?

We will discuss that shortly. However, before progressing any further, here is an amazing video introducing the core concept of quantum computing at a high level and in an incredibly funny manner.

TFQ provides the tools necessary for bringing quantum computin g and machine learning research communities together to control and model natural or artificial quantum systems.

NISQ: A new frontier of physical science

John Preskil , Richard P. Feynman Professor of Theoretical Physics at the California Institute of Technology, in his influential paper “ Quantum Computing in the NISQ era and beyond ” states that, “ we are now in the early stages of exploring a new frontier of the physical sciences, what we might call the complexity frontier or the entanglement frontier ”.

He goes on to explain this further — “ Now, for the first time in human history, we are acquiring and perfecting the tools to build and precisely control very complex, highly entangled quantum states of many particles, states so complex that we can’t simulate them with our best digital computers or characterize them well using existing theoretical tools. This emerging capability will open the door to new discoveries ”.

Near-term (imperfect) quantum devices

Quantum computing and/or information processing systems have a long and complex history. Today, in 2020, it is almost beyond doubt that the commercialization of practical methods to fabricate the necessary number of high-quality qubits and quantum gates lies decades away. This is mostly due to the issue of noise that is an inherent feature of these systems. The noise makes the true scaling of quantum systems (to a larger number of gates) extremely difficult and renders quantum computing devices unstable and unreliable.

However, researchers are increasingly excited about robust usage and possibilities from near-term quantum devices. These modestly sized devices (50–100 qubits), called Noisy Intermediate-Scale Quantum (NISQ), are mostly designed for fixed applications and can live with the noisy imperfection that accompanies them.

This class of algorithms and devices provide the real possibility of simulating and analyzing the properties of systems that are intractable from a classical calculation point of view.

This sounds interesting! But, what are some examples of such systems whose working can be impossible to simulate using classical computers/algorithms but where this NISQ technology fits right in?

Niche but valuable application areas

For many applications of quantum devices, such as cryptography , the noise poses a serious limitation and can lead to unacceptable levels of error. NISQ devices are unlikely to be used for such applications anytime soon. Therefore, people in finance and business security systems need not be worried, just yet. The tried and tested algorithms like RSA-128 are safe, for now.

But cryptography is not the only possible application area for quantum computing. NISQ devices and algorithms make it possible to probe deep into the properties of complex molecules and advanced nano-materials. They also allow high-accuracy and high-fidelity simulation of extreme physics, such as particle-particle interaction, or high-energy atomic collisions, possible.

One can imagine a vast and diverse array of possible application areas with these abilities…

• drug discovery
• nanomaterial synthesis
• genetic engineering and personalized medicine
• chemical engineering and material science
• particle physics
• semiconductors and electronic devices
• optical fibers, photonics, and high-speed communication

Gradual and steady progress in the NISQ systems has already got scientists, researchers, and business executives excited about these applications (and many more). However, it is to be realized that with the release of toolkits such as TFQ, all of these applications, can now, take full advantage of the powerful machinery of deep neural networks, and merge the fundamental physics with the power of Big Data.

This class of algorithms and devices provide the real possibility of simulating and analyzing the properties of systems that are intractable from a classical calculation point of view.

What is TFQ focused on?

Open-source philosophy with a near-term target

The promise of quantum computing has a long and complex history . However, there is a growing call from researchers to focus on practical problems that can be solved in the 5-year time horizon. TFQ is released at an opportune moment to take advantage of this movement and research momentum.

As mentioned above, TFQ is built around Cirq , which is designed and optimized for probing whether medium-sized, noisy QC architectures are capable of addressing practical problems, which are otherwise intractable with classical digital computing machinery. Respecting the tradition of open-source philosophy, Cirq is released under the Apache 2 license. Future open-source QC frameworks can take advantage of and embed Cirq methods and tools without any copyright restriction.

Data structures for optimized quantum circuits

A wide array of circuit design primitives, flexible behavioral and functional descriptions of logic gates, access to highly optimized timing and layout design tools, are some of the desired features for any digital design framework.

Quantum logic design is no different in these aspects. And tools like Cirq provide researchers just with those. In Cirq, Google claims , data structures have been optimized for writing and compiling a multitude of quantum circuit descriptions to allow users to get the most out of modern NISQ architectures. As an open-source project, it is constantly evolving and adding new abstractions and tools. Native support exists for both local simulation and easy integration with future quantum hardware or larger simulators on the cloud.

Special focus on chemistry

At the same time Google announced Cirq (in 2018), they also announced the availability of OpenFermion and OpenFermion-Cirq.

Github repo for OpenFremion .

Github Repo for OpenFermion-Cirq .

As per this Google AI blog , “ OpenFermion is a platform for developing quantum algorithms for chemistry problems, and OpenFermion-Cirq is an open-source library that compiles quantum simulation algorithms to Cirq ”.

Google’s Quantum AI team has been experimenting with novel quantum algorithms for quite some time and in particular, they have explored the possibility of developing reduced-complexity algorithms for quick prototyping in the realm of quantum chemistry . OpenFermion builds on these advances to help users translate the details of a chemical problem (e.g. molecular kinetics or reaction energetics ) into optimized quantum circuits. These circuits can also be tweaked and customized to run on particular hardware (e.g. electronic, photonic, or molecular).

Data structures have been optimized for writing and compiling a multitude of quantum circuit descriptions to allow users to get the most out of modern NISQ architectures.

Disentanglement of quantum data

Quantum-mechanical data generated by NISQ processors are inherently noisy and are typically entangled (before the measurement). However, one of the greatest utilities of quantum ML techniques is their ability to be applied to this noisy and entangled data to facilitate the extraction of classical information, which can be processed with any digital computing architectures.

Building on these techniques, the TFQ library facilitates the development of powerful models and workflows that disentangle and generalize correlations in the quantum data. On one hand, this encourages researchers to enhance the quality and complexity of novel quantum architectures, and on the other hand, it fast-tracks the development of new quantum algorithms.

Hybrid quantum-classical workflow

Because of their fairly limited size (i.e. gate count) and noisy behavior, NISQ processors still need to work in a tight loop with classical processors to produce meaningful computing output. For ML workload, this means a loop employing a NISQ processor tightly coupled with a classical deep-learning model. A hybrid approach is essential.

Now, in the world of classical ML, TensorFlow is one of the most respected and widely used platforms. It supports heterogeneous and petascale computing across all kinds of computing hardware — CPUs, GPUs, TPUs, and even FPGAs or custom ASICs. It is, therefore, no surprise, that Google chose this framework to integrate with the Cirq platform thereby extending the reach of QC tools to the widest possible hardware architectures when it comes to the matter of hybridization.

It’s a wonderful world where compilers, schedulers, and optimizers merge seamlessly with qubits, quantum gates, circuits, and measurement operators . The supervised ML training can be done using standard Keras functions and the final outcome of quantum measurements, evolving into classical probability distributions, is obtained by standard TensorFlow Ops .

From Google’s white-paper , this is how the workflow looks,

For more details, refer to this comprehensive paper,

The supervised ML training can be done using standard Keras functions and the final outcome of quantum measurements, evolving into classical probability distributions, is obtained by standard TensorFlow Ops .

Summary

It is an exciting time.

On one hand, science, technology, business, and society, are on the cusp of transformative changes due to massive explosions of data-driven intelligence aided by machine learning and artificial intelligence tools and algorithms.

On the other hand, the traditional frontiers of physical science (particle physics and cosmology — the study of the ultra-small and of the ultra-large) are giving away to the study of ultra-complex in the form of quantum devices and information systems.

We can only hope that the merge of two such powerful and long-range technological streams bode well for the advancement of human society and life on earth in general.

Release of an open-source computing platform such as TFQ is only a small step in that journey, but certainly, it will not be the last one.