# How does QC help apply machine learning techniques?

Google and NASA have set up an Quantum Artificial Intelligence laboratory. In December of 2015, they announced a groundbreaking result. What makes QC particularly attractive to machine learning researchers?

One of the major advantages to considering quantum computing – specifically quantum annealing for machine learning is that the solutions can be viewed as samples from a Boltzmann distribution. This is a very difficult computational process, that is typically approximated by a Markov Chain Monte Carlo method. If one can effectively sample from a Boltzmann distribution, then building probabilistic models based on the joint-probability distribution is in reach. This opens the doors to generative learning algorithms that can provide a much richer statistical framework for guiding experiments, performing model selection, model averaging etc. In short, with a quantum computer, one can integrate principles of probabilistic uncertainty into machine learning algorithms, as opposed to purely deterministic models.

To answer this question at a reasonably high level as a Machine Learning Researcher, I’ll say:

Every Machine Learning procedure is an optimization problem at core (generally NP-Hard) and because Quantum Computing promises to solve such problems with better accuracy (preferably optimally) and in relatively less time – it gives us hope and sounds attractive. Currently the machine learning community relies on heuristics and sub-optimal solutions.

– JasOberoi

I cannot speak for NASA or Google ( and I rather believe in some reputable proceedings from a scientific discourse, than to resign on one’s own critical thinking and remain to rely on any form of PR-advertorials, as an illustrative example of such ref.: >>> a story about Google’s unfulfilled self-promotion ) so let’s agree to put any similar groundbreakings announced a-priori aside and rather focus on relevant facts.

A better point of view, so as to attack the core of your question, could be to ask “**What does Machine Learning rely on?** ( Where QC can remarkably help )”

ML builds on three cornerstones ( and forgive me this intentional simplification )

- Decide what
**[ LEARNER ]**-process would best fit the real-world problem ( the formulation thereof ) - Find set of
**[ HYPERPARAMETER ]**s that would allow the [ LEARNER ]**best & fast meet**a criterial function’s**{ local | global***( Hic Sunt Leones )***} extreme**( .min() for penalty-function, .max() for utility-function ) - Carefully sanitize [
**DataSET ]**s so that the actual content, presented to the [ LEARNER ]-process, does not spoil the otherwise good choice of 1 + 2.

While there are remarkable research efforts spent on more specialised, advanced & sophisticated [LEARNER]s,

the first & undoubtedly a **sure shot** for QC aims on [2], where the worst beasts from complexity-ZOO do live.

There ( in [2] ) any **QC**-approach has a big chance to beat and outperform a way more, than by a “length of nose” a classical seq-en-ti-al-ly-bas-ed-com-pu-tin-g **both in achievable time & in precission.**

( Not speaking about a courage to indeed expand the sizes of [HYPERPARAMETER]s’-search-space into such non-prohibitive scales no one would ever dare to ask for on the currently available computing facilities )