You are invited to the wedding of Quantum Computing and TensorFlow

The quantum computerization of TensorFlow (TF) is a quixotic dream that no doubt has crossed the minds of many, both technically and not technically savvy, people. Here at artiste-qb.net, we are very committed and well underway to achieving this goal. Since our company is fully committed to open source, it doesn’t really matter if we achieve this goal before anyone else. If someone else beats us to it, we will learn from their code and vice versa. That is, as long as they too deliver open source to the world. And if they don’t, we think that their software is doomed…quantum open source rules! How did Google vanquish, or at least de-fang, the Microsoft monopoly? To a large extent, by using Open Source. Open source rules AI, the cloud and mobile.

So, let me tell you how we are using TF for quantum computing.

When Google first open sourced TF a mere 2 years ago, I wrote a blog post to mark the occasion. In that post, I called TF a platform for Fakesian networks instead of Bayesian networks. My beef was that TF had only deterministic nodes, which are standard in the field of artificial neural nets. It had no probabilistic nodes, which are standard in the 2 fields of classical Bayesian networks (BNets) and Hierarchical models (HM). But this past year, the open source community has fallen into the breach, filled in the gap, with a software library called Edward built on top of TF, that adds probabilistic nodes (the buzz word is “Probabilistic Deep Thinking”) to TF. And Edwards has been approved for integration into TF, so soon it will be seamless integrated into TF. Thus, soon, TF will combine artificial neural nets and BNets seamlessly. It will have superpowers!

Of course, in quantum mechanics, one must use complex amplitudes instead of probabilities for the nodes, and one must use an L2 norm instead of an L1 one with those amplitudes, so you can’t use Edward to do quantum mechanics just yet. Edward will have to be made “quantum ready”. By that we mean that one will have to rewrite parts of Edward so that it has a “quantum switch”, i.e. a parameter called ‘is_quantum’ which when True gives a quantum BNet and when False gives a classical BNet. That is precisely what artiste-qb.net’s open source program Quantum Fog already does, so our company is uniquely placed to make a quantum version of Edward.

Another obstacle to marrying TF and quantum computers is that the quantum BNets will have to be compiled into a sequence of elementary operations (SEO) such as control nots and single qubit rotations. Once again, our company artiste-qb.net is uniquely placed to accomplish this task. Our open source quantum simulator Qubiter is the only one in the business that includes a “quantum csd complier”, which is a tool that will help express quantum BNets as a SEO.

HMs are really a subset of BNets. However, the BNet and HM communities have historically grown somewhat independently. The BNet community is centered around pioneers like Judea Pearl (at UCLA), inventor of some of the most important BNet methods, whereas the HM community is centered around pioneers like Andrew Gelman (at Columbia), author of many great books and a great blog in the HM field. The HM tribe only uses continuous distributions for node probabilities, and they are very keen on MCMC (Markov Chain Monte Carlo). The BNet community uses both discrete (expressed as matrices or tensors) and continuous distributions for their node probabilities, and they use MCMC and other methods too, like the junction tree method, to do inferences.

Edward has a distinguished pedigree in both the BNet and HM communities. Edward originated in Columbia Univ. One of its main and original authors is Dustin Tran, currently a PhD student at Columbia. So you can be sure that the Edward people are in close communication and receive useful feedback from the Gelman tribe. Another distinguished author of Edward is Kevin Murphy, who has been working on BNets for more than a decade. Murphy wrote the oldie but goodie Bayes Net toolbox for Matlab. He has also written several books on bnets and machine learning. He previously worked as a prof at the Univ. of British Columbia but he now works at Google. He is one of the main organizers of the young (2 year old) Bayesian Deep Learning conference, which, by the way, will have its annual meeting in less than a week (Dec. 9, 2017).

Classical BNets are a very active field, both in academic research and in commerce. Judea Pearl won a Turing award for them. BNets are very popular in bioinformatics, for example. Whereas no qc company has yet broken-even financially, there are classical BNet companies that have lasted and been profitable for almost 2 decades, such as Bayesia, Hugin and Norsys/Netica.

Oh, and one last thing. It’s called TensorFlow, not TensorNetwork, for a very good reason. If you try to use TF to implement the “tensor networks” used in quantum computing, you will fail, unless you start using BNets instead of Tensor Networks and pretend these 2 are the same thing, which is probably what the Tensor Networks people will do. In TF (and BNets), the lines emanating out of a node carry in them a full tensor that they pass along to other nodes. In a Tensor Network, a Tensor does not Flow into the arrows emanating out of its node. The tensor just sits in the node. For more discussion about the important differences between a quantum BNet and a Tensor Network, see this blog post.

Tensor Networks versus Quantum Bayesian Networks: And the winner is…