New Research Shows How Quantum Generative Models Can Outperform Classical Models
New Research Shows How Quantum Generative Models Can Outperform Classical Models
We’ve said before that generative modeling — at the heart of buzzworthy generative AI tools like ChatGPT — will likely be the first place we see a practical advantage with quantum computers. In short, this stems from quantum computers’ unique advantages when it comes to encoding and sampling from complex probability distributions — the crux of generative modeling. For more details, see our blog post on the subject from last year.
But until now, we haven’t been able to demonstrate that quantum computers can outperform existing classical state-of-the-art generative models. That all changed with new research published last month. For the first time, we identified scenarios in which quantum generative models can outperform classical models. This research is a breakthrough moment for the field of Quantum AI.
For the first time, we identified scenarios in which quantum generative models can outperform classical models. This research is a breakthrough moment for the field of Quantum AI.
In the research, we establish the first framework for quantitatively comparing quantum and classical generative models, namely Quantum Circuit Born Machines (QCBMs) and Transformers (TFs), Recurrent Neural Networks (RNNs), Variational Autoencoders (VAEs), and Wasserstein Generative Adversarial Networks (WGANs), respectively. We choose architectures that are the basis of models successfully used for many applications. For example, transformers are widely used today in tools like OpenAI’s ChatGPT and DALL-E.
The study builds on previous work by the Zapata Quantum AI team to develop metrics for evaluating practical generalization in quantum generative models. Generalization is one of the most desirable qualities of a generative model: in a practical context, it measures a model’s ability to generate diverse, high-quality samples that go beyond the data it was trained on.
For an example of why this is relevant, imagine a generative model that was designed to generate new molecules that could be used to make pharmaceutical drugs. This model wouldn’t be very useful if it only generated the same known, existing molecules that were used to train the model. Nor would it be useful if it generated molecules that didn’t have the drug’s desired effect. A model would need good generalization to generate new, desirable molecules.
Generalization is one of the most desirable qualities of a generative model: in a practical context, it measures a model’s ability to generate diverse, high-quality samples that go beyond the data it was trained on.
In our latest research, we used our generalization metrics to compare the performance of quantum and classical generative models in the hopes of finding a potential practical quantum advantage (PQA). In other words, we were looking for a scenario in which a quantum algorithm had an advantage over the best known available classical algorithms using our best efforts.
You can think of it like a track and field meet, where each model is a different competitor racing in different hurdles races, for example the 100m hurdles or 400m hurdles. Each different race would be akin to a different real-world task in our framework, for example a learning task with a limited sampling budget, or one where the cost function is expensive to evaluate. In each race, the models run the same race but on different tracks, with the hurdles differing according to each model’s strengths and weaknesses (namely, training efficiency, sampling efficiency and expressive power). This allows us to identify the strengths of quantum generative models and the kinds of applications where they may be most valuable.
In all the “races,” the quantum generative models were competitive with the state-of-the-art solvers. Crucially, the quantum models had better generalization than the classical models in one key race: when data was scarce.
Such a feature is highly desirable in a wide range of real-world applications where the available data is scarce. For example, let’s return to the drug discovery example from earlier. For many medical conditions, there are few existing drugs that could be used to train a generative model for proposing new molecules. Having a generative model that could generate new, working drugs using that limited dataset of existing drugs would be a game-changer.
Well… no. To be clear, this research does not demonstrate practical quantum advantage in and of itself. Due to the limitations of existing quantum devices, our tasks were designed to only have 20 variables, which could then be encoded with 20 qubits. In most generative modeling tasks of practical value, there would be hundreds if not thousands of variables.
We also didn’t use real quantum devices, but rather simulations of quantum devices — without the noise that is unavoidable in real quantum devices available today. An interesting future research direction would be to either include noise in the simulations or to use real quantum hardware.
We also did not compare our quantum model to the best classical generative models available today, as required by the definition of PQA. For example, while we compared our quantum model with transformers, they weren’t on the same level as GPT-4. However, any other classical model could hypothetically be compared in this framework, and we encourage any readers to use the framework to compare their own models.
The real outcome of the research is that we created a series of races that can be used to compare the best quantum and classical models in the future, and identified the race where a quantum model would have the greatest potential to deliver a PQA: in the scarce data regime.
So yes, the research proves the generalization power of quantum generative models and shows the types of generative modeling problems that quantum models excel at. But the capabilities of quantum computers will need to scale up significantly before these quantum generative models will be able to deliver an advantage for real-world problems.
The real outcome of the research is that we created a series of races that can be used to compare the best quantum and classical models in the future, and identified the race where a quantum model would have the greatest potential to deliver a PQA: in the scarce data regime.
But that doesn’t mean there isn’t value to be gained today. In the meantime, we can achieve an advantage with generative models based on tensor networks, a quantum-inspired approach. Tensor networks (TN) first gained popularity among quantum physicists to model quantum states on classical computers. They can also be used to simulate quantum circuits.
As our previous research has shown, this means TN-based generative models are forward-compatible with fully quantum generative models. Similar to a relay race, we can start training a generative model with TNs, train it to the limits of what’s possible with classical computers, and then map the model onto a quantum circuit to include deeper correlations that are only possible with a quantum computer. Thus, TNs can be thought of as the bridge to practical quantum advantage in generative modeling.
One area where we’ve already shown an advantage with TN-based generative models is in generating solutions to complex optimization problems. These problems abound in the business world — from optimizing transportation networks to network architectures to warehouse designs and workforce scheduling. Using our generator-enhanced optimization (GEO) technique, we can train generative models on the solutions generated by state-of-the-art classical optimizers and generate new solutions.
The promising generalization performance of the quantum generative models in our recent research paper gives us optimism that as quantum computers become more powerful, they will be able to propose higher-quality solutions to optimization problems
In recent work with BMW to optimize their vehicle production scheduling, we showed that GEO using a TN-based generative model tied or outperformed state-of-the art solvers in 71% of problem configurations. GEO worked particularly well for configurations of the problem where we had a large space of possible solutions, showing its promise for more complex optimization problems that traditional solvers may struggle with.
The promising generalization performance of the quantum generative models in our recent research paper gives us optimism that as quantum computers become more powerful, they will be able to propose higher-quality solutions to optimization problems in the GEO framework.
Best of all, organizations that already have GEO in operation using TNs will be able to easily map their TN models onto these more powerful quantum circuits — setting themselves up for a potentially greater advantage.
While there is still much work to be done, this achievement demonstrates the incredible potential of quantum computing and sets the stage for further breakthroughs in the future.