Search Talks

Training, validation and testing datasets; methods that prevent overfitting; using supervised neural networks to classify phases of matter (classical Ising model, classical Ising gauge theory)

From earliest infancy, we live in and learn to function in a world of causes and effects. Yet science has had an ambivalent, even hostile attitude toward causation for more than a century. Statistics courses teach us that “correlation is not causation,” yet they are strangely silent about what is causation.

A central reason for this silence is that causation does not reside in data alone, but in the process that generates the data. In order to answer causal questions, like “What would happen if we lowered the price of toothpaste?” or “Should I brake for this object?” we need a model of causes and effects. Judea Pearl has developed a simple calculus for expressing our cause-effect knowledge in a diagram and using that diagram to tell us how to interpret the data we gather from the real world. His methods are already transforming the practice of statistics and could equip future artificial intelligences with causal reasoning abilities they currently lack.

This talk is largely based on Mackenzie’s book co-written with Pearl, The Book of Why.

Most machine learning algorithms involve optimizing a single set of parameters to decrease a single cost function. In adversarial machine learning, two or more "players" each adapt their own parameters to decrease their own cost, in competition with the other players. In some adversarial machine learning algorithms, the algorithm designer contrives this competition between two machine learning models in order to produce a beneficial side effect. For example, the generative adversarial networks framework involves a contrived conflict between a generator network and a discriminator network that results in the generator learning to produce realistic data samples. In other contexts, adversarial machine learning models a real conflict, for example, between spam detectors and spammers. In general, moving machine learning from optimization and a single cost to game theory and multiple costs has led to new insights in many application areas.

I will discuss the status and future of numerical lattice Quantum Chromodynamics (QCD) calculations for nuclear physics. With advances in supercomputing, we are beginning to quantitatively understand nuclear structure and interactions directly from the fundamental quark and gluon degrees of freedom of the Standard Model. Recent studies provide insight into the neutrino-nucleus interactions relevant to long-baseline neutrino experiments, double beta decay, and nuclear sigma terms needed for theory predictions of dark matter cross-sections at underground detectors. I will also address new work constraining `exotic glue’ in nuclei, which will be measurable for the first time at a future electron-ion collider, and explain how machine learning tools are providing new possibilities in this field.

We propose to generalise classical maximum likelihood learning to density matrices. As the objective function, we propose a quantum likelihood that is related to the cross entropy between density matrices. We apply this learning criterion to the quantum Boltzmann machine (QBM), previously proposed by Amin et al. We demonstrate for the first time learning a quantum Hamiltonian from quantum statistics using this approach. For the anti-ferromagnetic Heisenberg and XYZ model we recover the true ground state wave function and Hamiltonian. The second contribution is to apply quantum learning to learn from classical data. Quantum learning uses in addition to the classical statistics also quantum statistics for learning. These statistics may violate the Bell inequality, as in the quantum case. Maximizing the quantum likelihood yields results that are significantly more accurate than the classical maximum likelihood approach in several cases. We give an example how the QBM can learn a strongly non-linear problem such as the parity problem. The solution shows entanglement, quantified by the entanglement entropy.

Convex optimization, linear and semidefinite programming in particular, has been a standard tool in quantum information theory, giving certificates of local and quantum correlations, contextuality, and more. Increasingly, similar methods are making headways in quantum many-body physics, giving lower bounds -- and thus certificates -- on the ground state energy. The disadvantage of such methods is that they do not scale well to large system sizes, whether those systems are multiparty Bell scenarios or lattice models of numerous sites. Machine
learning is entering the field as the latest buzzword. While it provides a more scalable alternative to convex programming and enables forming new conjectures, the outcome of learning methods remains uncertified. In this talk, I introduce the most important paradigms in machine learning for quantum information theory, give an overview of some earlier work in the field, argue for the importance of certifiable predictions of learning algorithms, and present some of our preliminary results.

Short-depth algorithms are crucial for reducing computational error on near-term quantum computers, for which decoherence and gate infidelity remain important issues. Here we present a machine-learning inspired approach for discovering such algorithms. We apply our method to a ubiquitous primitive: computing the overlap Tr(rho*sigma) between two quantum states rho and sigma. The standard algorithm for this task, known as the Swap Test, is used in many applications such as quantum support vector machines, and, when specialized to rho=sigma, quantifies the Renyi entanglement. Here, we find algorithms that have shorter depths than the Swap Test, including one that has constant depth (independent of problem size). Furthermore, we apply our approach to the hardware-specific connectivity and gate alphabets used by Rigetti's and IBM's quantum computers and demonstrate that the shorter algorithms that we derive significantly reduce the error - compared to the Swap Test - on these computers.

Over the last decade, there have been enormous gains in machine learning technology primarily driven by neural networks. A major reason neural networks have outperformed older techniques is that the cost of optimizing them scales well with the size of the training dataset. But neural networks have the drawback that they are not very well understood theoretically.

Recent work by several groups has explored an alternative approach to creating machine learning model functions based on tensor networks, which are a technique for parameterizing many-body quantum wavefunctions. The cost of training tensor network models scales similarly to the cost of training neural networks. In addition, their relatively simple, linear structure has provided good theoretical understanding of their properties, and underpins many powerful techniques to optimize and manipulate them. 

After introducing tensor network machine learning models, I will discuss some of the techniques to optimize them and results for supervised and generative machine learning tasks. These techniques can automatically adapt the number of parameters and suggest interesting interpretations and extensions for 'deep' tensor network variants. I will conclude by discussing a recent tensor network based proposal to formulate hybrid quantum-classical algorithms for machine learning with quantum computers.

Can computers think? They can certainly calculate - with staggering speed and ever-increasing power - and they have driven scientific and technological advances that would have been impossible without them. Even so, we would like to believe that, for some puzzles, there's no substitute for old-fashioned human intuition. But this view may be changing.

A new breed of machine learning algorithms have begun knocking down cognitive milestones that, until recently, scientists believed were still decades away. Major advances are being made in computer vision, language translation, autonomous robotic action, and other complex applications. At the same time, these new algorithms are helping scientists accelerate discovery in physics.

This stunning progress poses as many questions as answers: What are the fundamental possibilities and limits of machine learning? Can we create true human-level artificial intelligence, and how might its thoughts differ from our own? What new breeds of computer will fuel artificial intelligence and, conversely, how will artificial intelligence enable new forms of computing?

In his public lecture at Perimeter Institute, Roger Melko will explore how computers have helped humanity solve increasingly complex puzzles, and ask which challenges, if any, only human intuition is equipped to tackle.