Sparse coding on D-Wave hardware: structured dictionaries

The underlying problem we saw last time, that prevented us from using the hardware to compete with tabu on the cloud, was the mismatch of the connectivity of the problems sparse coding generates (which are fully connected) and the connectivity of the hardware.

The source of this mismatch is the quadratic term in the objective function, which is of the form . The coupling terms are proportional to the dot product of the dictionary atoms.

Here’s an idea. What if we demand that has to be zero for all pairs of variables and that are not connected in hardware? If we can achieve this structure in the dictionary, we get a very interesting result. Instead of being fully connected, the QUBOs with this restriction can be engineered to exactly match the underlying problem the hardware solves. If we can do this, we get closer to using the full power of the hardware.

L0-norm sparse coding with structured dictionaries

Here is the idea.

Given

1. A set of data objects , where each is a real valued vector with components;
2. A real valued matrix , where is the number of dictionary atoms we choose, and we define its column to be the vector ;
3. An binary valued matrix ;
4. And a real number , which is called the regularization parameter,

Find and that minimize

subject to the constraints that for all pairs that are not connected in the quantum chip being used.

The only difference here from what we did before is the last sentence, where we add a set of constraints on the dictionary atoms.

Solving the sparse coding problem using block coordinate descent

We’re going to use the same strategy for solving this as before, with a slight change. Here is the strategy we’ll use.

1. First, we generate a random dictionary , subject to meeting the orthogonality constraints we’ve imposed on the dictionary atoms.
2. Assuming these fixed dictionaries, we solve the optimization problem for the dictionary atoms . These optimization problems are now Chimera-structured QUBOs that fit exactly onto the hardware by construction.
3. Now we fix the weights to these values, and find the optimal dictionary , again subject to our constraints.

We then iterate steps 2 and 3 until converges to a minimum.

Now we’re in a different regime than before — step 2 requires the solution of a large number of chimera-structured QUBOs, not fully connected QUBOs. So that makes those problems better fits to the hardware. But now we have to do some new things to allow for both steps 1 and 3, and these initial steps have some cost.

The first of these is not too hard, and introduces a key concept we’ll use for Step 3 (which is harder). In this post I’ll go over how to do Step 1.

Step 1: Setting up an initial random dictionary that obeys our constraints

Alright so the first step we need to do is to figure out under what conditions we can achieve Step 1.

There is a very interesting result in a paper called Orthogonal Representations and Connectivity of Graphs. Here is a short explanation of the result.

Imagine you have a graph on vertices. In that graph, each vertex is connected to a bunch of others. Call the number corresponding to the connectivity of the least connected variable in the graph. Then this paper proves that you can define a set of real vectors in dimension where non-adjacent nodes in the graph can be assigned orthogonal vectors.

So what we want to do — find a random dictionary such that for all not connected in hardware — can be done if the length of the vectors is greater than .

For Vesuvius, the number is 512, and the lowest connectivity node in a Chimera graph is . So as long as the dimension of the dictionary atoms is greater than 512 – 5 = 507, we can always perform Step 1.

Here is a little more color on this very interesting result. Imagine you have to come up with two vectors and that are orthogonal (the dot product is zero). What’s the minimum dimension these vectors have to live in such that this can be done? Well imagine that they both live in one dimension — they are just numbers on a line. Then clearly you can’t do it. However if you have two dimensions, you can. Here’s an example: and . If you have more that two dimensions, you can also, and the choices you make in this case are not unique.

More generally, if you ask the question “how many orthogonal vectors can I draw in an -dimensional space?”, the answer is — one vector per dimension. So that is a key piece of the above result. If we had a graph with vertices where NONE of the vertices were connected to any others (minimum vertex connectivity ), and want to assign vectors to each vertex such that all of these vectors are orthogonal to all the others, that’s equivalent to asking “given a -dimensional space, what’s the minimum dimension of a set of vectors such that they are all orthogonal to each other?”, and the answer is .

Now imagine we start drawing edges between some of the vertices in the graph, and we don’t require that the vectors living on these vertices be orthogonal. Conceptually you can think of this as relaxing some constraints, and making it ‘easier’ to find the desired set of vectors — so the minimum dimension of the vectors required so that this will work is reduced as the graph gets more connected. The fascinating result here is the very simple way this works. Just find the lowest connectivity node in the graph, call its connectivity , and then ask “given a graph on vertices, where the minimum connectivity vertex has connectivity , what’s the minimum dimension of a set of vectors such that non-connected vertices in the graph are all assigned orthogonal vectors?”. The answer is .

Null Space

Null Space is also an ASCII-based adventure game: https://students.digipen.edu/~tbrosman/null_space_download.html .

Now just knowing we can do it isn’t enough. But thankfully it’s not hard to think of a constructive procedure to do this. Here is one:

1. Generate a matrix where all entries are random numbers between +1 and -1.
2. Renormalize each column such that each column’s norm is one.
3. For each column in from the leftmost to the rightmost in order, compute the null space of that column, and then replace that column with a random column written in the null space basis.

If you do this you will get an initial random orthonormal basis as required in our new procedure.

By the way, here is some Python code for computing a null space basis for a matrix . It’s easy but there isn’t a native function in numpy or scipy that does it.

import numpy
from scipy.linalg import qr

def nullspace_qr(A):

A = numpy.atleast_2d(A)
Q, R = qr(A.T)
ns = Q[:, R.shape[1]:].conj()
return ns

OK so step 1 wasn’t too bad! Now we have to deal with step 3. This is a harder problem, which I’ll tackle in the next post.

Sparse coding on D-Wave hardware: things that don’t work

Ice, ice baby.

For Christmas this year, my dad bought me a book called Endurance: Shackleton’s Incredible Voyage, by Alfred Lansing.

It is a true story about folks who survive incredible hardship for a long time. You should read it.

Shackleton’s family motto was Fortitudine Vincimus — “by endurance we conquer”. I like this a lot.

On April 22nd, we celebrate the 14th anniversary of the incorporation of D-Wave. Over these past 14 years, nearly everything we’ve tried hasn’t worked. While we haven’t had to eat penguin (yet), and to my knowledge no amputations have been necessary, it hasn’t been a walk in the park. The first ten things you think of always turn out to be dead ends or won’t work for some reason or other.

Here I’m going to share an example of this with the sparse coding problem by describing two things we tried that didn’t work, and why.

Where we got to last time

In the last post, we boiled down the hardness of L0-norm sparse coding to the solution of a large number of QUBOs of the form

Find that minimizes

I then showed that using this form has advantages (at least for getting a maximally sparse encoding of MNIST) over the more typical L1-norm version of sparse coding.

I also mentioned that we used a variant of tabu search to solve these QUBOs. Here I’m going to outline two strategies we tried to use the hardware to beat tabu that ended up not working.

These QUBOs are fully connected, and the hardware isn’t

The terms in the QUBO that connect variables and are proportional to the dot product of the and dictionary atoms and . Because we haven’t added any restrictions on what these atoms need to look like, these dot products can all be non-zero (the dictionary atoms don’t need to be, and in general won’t be, orthogonal). This means that the problems generated by the procedure are all fully connected — each variable is influenced by every other variable.

Unfortunately, when you build a physical quantum computing chip, this full connectivity can’t be achieved. The chip you get to work with connects any given variable with only a small number of other variables.

There are two ways we know of to get around the mismatch of the connectivity of a problem we want to solve, and the connectivity of the hardware. The first is called embedding, and the second is by using the hardware to perform a type of large neighborhood local search as a component of a hybrid algorithm we call BlackBox.

Solving problems by embedding

In a quantum computer, qubits are physically connected to only some of the other qubits. In the most recent spin of our design, each qubit is connected to at most 6 other qubits in a specific pattern which we call a Chimera graph. In our first product chip, Rainier, there were 128 qubits. In the current processor, Vesuvius, there are 512.

Chimera graphs are a way to use a regular repeating pattern to tile out a processor. In Rainier, the processor graph was a four by four tiling of an eight qubit unit cell. For Vesuvius, the same unit cell was used, but with an eight by eight tiling.

For a detailed overview of the rationale behind embedding, and how it works in practice for Chimera graphs, see here and here, which discuss embedding into the 128-qubit Rainier graph (Vesuvius is the same, just more qubits).

The short version is that an embedding is a map from the variables of the problem you wish to solve to the physical qubits in a processor, where the map can be one-to-many (each variable can be mapped to many physical qubits). To preserve the problem structure we strongly ‘lock together’ qubits corresponding to the same variable.

In the case of fully connected QUBOs like the ones we have here, it is known that you can always embed a fully connected graph with vertices into a Chimera graph with physical qubits — Rainier can embed a fully connected 17 variable graph, while Vesuvius can embed a fully connected 33 variable graph. Shown to the right is an embedding from this paper into Rainier, for solving a problem that computes Ramsey numbers. The processor graph where qubits colored the same represent the same computational variable.

So one way we could use Vesuvius to solve the sparse coding QUBOs is to restrict to be 33 or less and embed these problems. However this is unsatisfactory for two (related) reasons. The first is that 33 dictionary atoms isn’t enough for what we ultimately want to do (sparse coding on big data sets). The second is that QUBOs generated by the procedure I’ve described are really easy for tabu search at that scale. For problems this small, tabu gives excellent performance with a per problem timeout of about 10 milliseconds (about the same as the runtime for a single problem on Vesuvius), and since it can be run in the cloud, we can take advantage of massive parallellism as well. So even though on a problem by problem basis, Vesuvius is competitive at this scale, when you gang up say 1,000 cores against it, Vesuvius loses (because there aren’t a thousand of them available… yet ).

So this option, while we can do it, is out. At the stage we’re at now this approach can’t compete with cloud-enabled tabu. Maybe when we have a lot more qubits.

Solving sparse coding QUBOs using BlackBox

BlackBox is an algorithm developed at D-Wave. Here is a high level introduction to how it works. It is designed to solve problems where all we’re given is a black box that converts possible answers to binary optimization problems into real numbers denoting how good those possible answers are. For example, the configuration of an airplane wing could be specified as a bit string, and to know how ‘good’ that configuration was, we might need to actually construct that example and put it in a wind tunnel and measure it. Or maybe just doing a large-scale supercomputer simulation is enough. But the relationship between the settings of the binary variables and the quality of the answer in problems like this is not easily specified in a closed form, like we were able to do with the sparse coding QUBOs.

BlackBox is based on tabu search, but uses the hardware to generate a model of the objective function around each search point that expands possibilities for next moves beyond single bit flips. This modelling and sampling from hardware at each tabu step increases the time per step, but decreases the number of steps required to reach some target value of the objective function. As the cost of evaluating the objective function goes up, the gain in making fewer ‘steps’ by making better moves at each tabu step goes up. However if the objective function can be very quickly evaluated, tabu generally beats BlackBox because it can make many more guesses per unit time because of the additional cost of the BlackBox modeling and hardware sampling step.

BlackBox can be applied to arbitrary sized fully connected QUBOs, and because of this is better than embedding because we lose the restriction to small numbers of dictionary atoms. With BlackBox we can try any size problem and see how it does.

We did this, and unfortunately BlackBox on Vesuvius is not competitive with cloud-enabled tabu search for any of the problem sizes we tried (which were, admittedly, still pretty small — up to 50 variables). I suspect that this will continue to hold, no matter how large these problems get, for the following reasons:

1. The inherently parallel nature of the sparse coding problem ( independent QUBOs) means that we will always be up against multiple cores vs. a small number of Vesuvius processors. This factor can be significant — for a large problem with millions of data objects, this factor can easily be in the thousands or tens of thousands.
2. BlackBox is designed for objective functions that are really black boxes, so that there is no obvious way to attack the structure of the problem directly, and where it is very expensive to evaluate the objective function. This is not the case for these problems — they are QUBOs and this means that attacks can be made directly based on this known fact. For these problems, the current version of BlackBox, while it can certainly be used, is not in its sweet spot, and wouldn’t be expected to be competitive with tabu in the cloud.

And this is exactly what we find — BlackBox on Vesuvius is not competitive with tabu on the cloud for any of the problem sizes we tried. Note that there is a small caveat here — it is possible (although I think unlikely) that for very large numbers of atoms (say low thousands) this could change, and BlackBox could start winning. However for both of the reasons listed above I would bet against this.

What to do, what to do

We tried both obvious tactics for using our gear to solve these problems, and both lost to a superior classical approach. So do we give up and go home? Of course not!

We shall go on to the end… we shall never surrender!!!

We just need to do some mental gymnastics here and be creative.

In both of the approaches above, we tried to shoehorn the problem our application generates into the hardware. Neither solution was effective.

So let’s look at this from a different perspective. Is it possible to restrict the problems generated by sparse coding so that they exactly fit in hardware — so that we require the problems generated to exactly match the hardware graph? If we can achieve this, we may be able to beat the classical competition, as we know that Vesuvius is many orders of magnitude faster than anything that exists on earth for the native problems it’s solving.

Sparse coding on D-Wave hardware: setting up the problem

Sparse coding is a very interesting idea that we’ve been experimenting with. It is a way to find ‘maximally repeating patterns’ in data, and use these as a basis for representing that data.

Some of what’s going to follow here is quite technical. However these are beautiful ideas. They are probably related to how human perception and cognition functions. I think sparse coding is much more interesting and important than, say, the Higgs Boson.

Everything starts from data

Twenty five data objects. Each is a 28×28 pixel greyscale image.

Sparse coding requires data. You can think of ‘data’ as a (usually large) set of objects, where the objects each can be represented by a list of real numbers. As as example, we’ll use the somewhat pathological MNIST handwritten digits data set. But you can use any dataset you can imagine. Each data object in this set is a small (28×28 pixel) greyscale image of a handwritten digit. A 28×28 pixel greyscale image can be represented by 28×28 = 784 numbers, each in the range 0..255. The training set has 60,000 of these.

We can represent the entire data set using a two-dimensional array, where there are 60,000 columns (one per image), and 784 rows (one for each pixel). Let’s call this array .

Technical detail: this thing is bigger than it has to be

What the first few MNIST images look like, keeping an increasingly large number of SVD modes.

One thing you may notice about MNIST is that each image kind of looks mostly similar. In fact we can exploit this to get a quick compression of the data. The trick we will use is called Singular Value Decomposition (SVD). SVD quickly finds a representation of the data that allows you to reduce its dimensionality. In the case of MNIST, it turns out that instead of using 784 pixel values, we can get away with using around 30 SVD modes instead, with only a small degradation in image quality. If we perform an SVD on and keep only the parts corresponding to the largest 30 singular values, we get a new matrix which still has 60,000 columns, but only 30 rows. Let’s call the column vector , which stores the compressed image. This trick, and others related to it, can always be used to pre-process any raw data we have.

The Dictionary

Let’s now create a small number — let’s say 32 — of very special images. The exact number of these actually matters quite a bit (it’s an example of something called a hyperparameter, many of which lurk in machine learning algorithms and are generally a giant pain in the ass), but the most important thing is that it should be larger than the dimension of the data (which in this case is 30). When this is the case, it’s called an overcomplete basis, which is good for reasons you can read about here.

These images will be in the exact same format as the data we’re learning over. We’ll put these in a new two dimensional array, which will have 32 columns (one for each image) and 30 rows (the same as the post-SVD compressed images above). We’ll call this array the dictionary. Its columns are dictionary atoms. The dictionary atom is the column vector . These dictionary atoms will store the ‘maximally repeating patterns’ in our data that are what we’re looking for.

At this stage, we don’t know what they are — we need to learn them from the data.

The sparse coding problem

Now that we have a data array, and a placeholder array for our dictionary, we are ready to state the sparse coding problem.

Now to see how all this works, imagine we want to reconstruct an image (say ) with our dictionary atoms. Imagine we can only either include or exclude each atom, and the reconstruction is a linear combination of the atoms. Furthermore, we want the reconstruction to be sparse, which means we only want to turn on a small number of our atoms. We can formalize this by asking for the solution of an optimization problem.

L0-norm sparse coding

Define a vector of binary (0/1) numbers of length 32. Now solve this problem:

Find and (remember the vectors are column vectors in the matrix ) that minimize

The real number is called a regularization parameter (another one of those hyperparameters). The larger this number is, the bigger the penalty for adding more dictionary atoms to the reconstruction — the rightmost term counts the number of atoms used in the reconstruction. The first term is a measure of the reconstruction error. Minimizing this means minimizing the distance between the data (sometimes called the ground truth) and the reconstruction of the image .

[Note that this prescription for sparse coding is different that the one typically used, where the weights are real valued and the regularization term is of the L1 (sum over absolute values of the weights) form.]

You may see a simple way to globally optimize this. All you have to do is set , for all the other k, and for all the other k — in other words, store the image in one of the dictionary atoms and only turn that one on to reconstruct. Then the reconstruction is perfect, and you only used one dictionary atom to do it!

OK well that’s useless. But now say we sum over all the images in our data set (in the case of MNIST this is 60,000 images). Now instead of the number of images being less than the number of dictionary atoms, it’s much, much larger. The trick of just ‘memorizing’ the images in the dictionary won’t work any more.

Now over all the data

Let’s call the total number of data objects . In our MNIST case . We now need to find (this is a 32xS array) and that minimize

This is the full sparse coding prescription — in the next post I’ll describe how to solve it, what the results look like, and introduce a bunch of ways to make good use of the dictionary we’ve just learned!

Quantum computing for learning: Shaping reality both figuratively and literally

I’d like to document in snippets and thought-trains a little more of the story behind how my co-workers and I are trying to apply quantum computing to the field of intelligence and learning. I honestly think that this is the most fascinating and cool job in the world. The field of Artificial Intelligence (AI) – after a period of slowdown – is now once again exploding with possibility. Big data, large-scale learning, deep networks, high performance computing, bio-inspired architectures… There have been so many advancements lately that it’s kind of hard to keep up! Similarly, the work being done on quantum information processing here at D-Wave is ushering in a new computational revolution. So being a multi-disciplinary type and somewhat masochistic, I find it exciting to explore just how far we can take the union of these two fields.

Approximately 5 years ago, while working on my PhD at University, I started to have an interest in whether or not quantum computing could be used to process information in a brain-like way. I’m trying to remember where this crazy obsession came from. I’d been interested in brains and artificial intelligence for a while, and done a bit of reading on neural nets. Probably because one of my friends was doing a PhD that involved robots and I always thought it sounded super-cool. But I think that it was thinking about Josephson junctions that really got me wondering. Josephson junctions are basically switches. But they kind of have unusual ways of switching (sometimes when you don’t want them to). And because of this, I always thought that Josephson junctions are a bit like neurons. So I started searching the literature to find ways in which researchers had used these little artificial neurons to compute. And surprisingly, I found very little. There were some papers about how networks of Josephson junctions could be used to simulate neurons, but no-one had actually built anything substantial. I wrote a bit about this in a couple of old posts (from Physics and Cake blog):

• Quantum Neural Networks 1: The Superconducting Neuron Model
• Josephson Junction Neurons

I’d read about the D-Wave architecture and I’d been following the company’s progress for some time. After reading a little about the promise of Josephson junction networks, and the pitfalls of the endeavour (mostly because making the circuits reproducible is extremely difficult), I then began wondering whether or not the D-Wave processor could be used in this way. It’s a network of qubits made from Josephson junctions after all, and they’re connected together so that they talk to each other. Yeah, kind of like neurons really. Isn’t that funny. And hey, those D-Wave types have spent 8 years getting that network of Josephson junctions to behave itself. Getting it to be programmable, addressable, robust, and scalable. Hmm, scalable…. I particularly like that last one. Brains are like, big. Lotsa connections. And also, I thought to myself (probably over tea and cake), if the neurons are qubits, doesn’t that mean you can put them in superposition and entangled states? What would that even mean? Boy, that sounds cool. Maybe they would process information differently, and maybe they could even learn faster if they could be in combinations of states at the same time and … could you build a small one and try it out?

The train of thought continued.

From quantum physics to quantum brains

That was before I joined D-Wave. Upon joining the company, I got to work applying some of my physics knowledge to helping build and test the processors themselves. However there was a little part of me that still wanted to actually find ways to use them. Not too long after I had joined the company there happened to be a competition run internally at D-Wave known as ‘Apps Day’, open to everyone in the company, where people were encouraged to try to write an app for the quantum computer. Each candidate got to give a short presentation describing their app, and there were prizes at stake.

I decided to try and write an app that would allow the quantum computer to learn how to play the board game Go. It was called QUAGGA, named after an extinct species of zebra. As with similar attempts involving the ill-fated zebra, I too might one day try to resurrect my genetically-inferior code. Of course this depends on whether or not I ever understand the rules of Go well enough to program it properly Anyway… back to Apps Day. There were several entries and I won a runner-up prize (my QUAGGA app idea was good even though I hadn’t actually finished coding it or run it on the hardware). But the experience got me excited and I wanted to find out more about how I could apply quantum processing to applications, especially those in the area of machine learning and AI.

That’s why I moved from physics into applications development.

Since then the team I joined has been looking into applying quantum technology to various areas of machine learning, in a bid to unite two fields which I have a really strong feeling are made for each other. I’ve tried to analyse where this hunch originates from. The best way to describe it is that I really want to create models of machine intelligence and creativity that are bio-inspired. To do that I believe that you have to take inspiration from the mammalian brain, such as its highly parallel, hierarchical arrangement of substructures. And I just couldn’t help but keep thinking: D-Wave’s processors are highly parallel systems with qubits that can be in one of two states (similar to firing or not firing neurons) with connections between them that can be inhibitory or excitory. Moreover, like the brain, these systems, are INCREDIBLY energy efficient because they are designed to do parallel processing. Modern CPUs are not – hence why brain simulations and machine learning programs take so much energy and require huge computer clusters to run. I believe we need to explore many different hardware and software architectures if we want to get smarter about intelligent computing and closer to the way our own minds work. Quantum circuits are a great candidate in that hunt for cool brain-like processing in silicon.

So what on earth happened here? I’d actually found a link between my two areas of obsession interest and ended up working on some strange joint project that combined the best of both worlds. Could this be for real? I kept thinking that maybe I wanted to believe so badly I was just seeing the machine-learning messiah in a piece of quantum toast. However, even when I strive to be truly objective, I still find a lot of evidence that the results of this endeavour could be very fruitful.

Our deep and ever-increasing understanding of physics (including quantum mechanics) is allowing us to harness and shape the reality of the universe to create new types of brains. This is super-cool. However, the thing I find even cooler is that if you work hard enough at something, you may discover that several fascinating areas are related in a deeper way than you previously understood. Using this knowledge, you can shape the reality of your own life to create a new, hybrid project idea to work on; one which combines all the things you love doing.

It’s like the quantum computer is playing 20 questions…

I’ve been thinking about the BlackBox compiler recently and came up with a very interesting analogy to the way it works. There are actually lots of different ways to think about how BlackBox works, and we’ll post more of them over time, but here is a very high level and fun one.

The main way that you use BlackBox is to supply it with a classical function which computes the “goodness” of a given bitstring by returning a real number (the lower this number, the better the bitstring was).

Whatever your optimization problem is, you need to write a function that encodes your problem into a series of bits (x1, x2, x3…. xN) to be discovered, and which also computes how “good” a given bitstring (e.g. 0,1,1…0) is. When you pass such a function to Blackbox, the quantum compiler then repeatedly comes up with ideas for bitstrings, and using the information that your function supplies about how good its “guesses” are, it quickly converges on the best bitstring possible.

So using this approach the quantum processor behaves as a co-processor to a classical computing resource. The classical computing resources handles one part of the problem (computing the goodness of a given bitstring), and the quantum computer handles the other (suggesting bitstrings). I realized that this is described very nicely by the two computers playing 20 questions with one another.

The quantum computer suggests creative solutions to a problem, and then the classical computer is used to give feedback on how good the suggested solution is. Using this feedback, BlackBox will intelligently suggest a new solution. So in the example above, Blackbox knows NOT to make the next question “Is it a carrot?”

There is actually a deep philosophical point here. One of the pieces that is missing in the puzzle of artificial intelligence is how to make algorithms and programs more creative. I have always been an advocate of using quantum computing to power AI, but we now start to see concrete ways in which it could really start to address some of the elusive problems that crop up when trying to build intelligent machines.

At D-Wave, we have been starting some initial explorations in the areas of machine creativity and machine dreams, but it is early days and the pieces are only just starting to fall into place.

I was wondering if you could use the QC to actually play 20 questions for real. This is quite a fun application idea. If anyone has any suggestions for how to craft 20 questions into an objective function, let me know. My first two thoughts were to do something with Wordnet and NLTK. You could try either a pattern matching or a machine learning version of ‘mining’ wordnet for the right answer. This project would be a little Watson-esque in flavour.

The dreams of spiritual machines

When I was in middle school, every year we had to select a project to work on. These projects came from a list of acceptable projects. The projects were typical science-ish projects you’d expect a seventh grader to take on. One year my project was about whooping cranes. Not sure why I picked that one. Maybe I thought it might be related to whooping cough.

One year the subject I picked was dreams. What were they? How did they come about? What, if anything did they tell us about our waking life? I remember being intensely fascinated by the topic at the time, feeling that the answers I was getting to my questions from grown-ups and the encyclopedias checked out from the school library (there was no internet back then, at least in a form I could access) were not satisfactory at all. This was one of my earliest realizations that there were questions no-one yet knew the answers to.

The subject of dreams has come up in my adult life several times, and each time the same questions about them bubble up from my early encounter with them. An acquaintance of mine went through a period of having night terrors, where she would scream so loud that it would wake people in neighboring houses. She described them as being a sense of horror and dread of the most intense and indescribable kind, with sure knowledge that it would never end. This led to multiple 911 calls over periods of years. Several trips to specialists and tests revealed nothing out of the ordinary. Then one day they suddenly stopped. To this day no one has a good explanation for why they started, or why they stopped.

One of my friends has multiple vivid, realistic dreams every night, and he remembers them. They are also often terrifying. I on the other hand rarely dream, or if I do, I don’t remember them.

Recently I have been thinking of dreams again, and I have four computer scientists to thank. One of them is Bill Macready, who is my friend and colleague at D-Wave, and inventor of the framework I’ll introduce shortly. The second is Douglas Hofstadter. The third is Geoff Hinton. The fourth is David Gelertner.

Gelertner is a very interesting guy. Not only is he a rock star computer scientist (Bill Joy called him “one of the most brilliant and visionary computer scientists of our time”), he is also an artist, entrepreneur and a writer with an MA is classical literature. He was injured badly opening a package from the Unabomber in 1993. He is the author of several books, but the one I want to focus on now is The Muse in the Machine, which  is must-read material for anyone interested in artificial intelligence.

In this book, Gelertner presents a compelling theory of cognition that includes emotion, creativity and dreams as a central critically important aspect of the creation of machines that think, feel and act as we do. In this theory, emotion, creativity, analogical thought and even spirituality are viewed as being essential to the creation of machines that behave as humans do. I can’t do the book justice in a short post – you should read it.

I am going to pull one quote out of the book though, but before I do I want to briefly touch on what Geoff Hinton has to do with all of this. Hinton is also a rock star in the world of artificial intelligence, and in particular in machine learning. He was one of the inventors of back propagation, and a pioneer in deep belief nets and unsupervised learning. A fascinating demo I really like starts around the 20:00 mark of this video. In this demo, he runs a deep learning system ‘in reverse’, in generative mode. Hinton refers to this process as the system “fantasizing” about the images it’s generating; however Hinton’s fantasizing can also be thought of as the system hallucinating, or even dreaming, about the subjects it has learned. Systems such as these exhibit what I believe to be clear instances of creativity – generating instances of objects that have never existed in the world before, but share some underlying property. In Hinton’s demo, this property is “two-ness”.

Alright so back to Gelertner, and the quote from The Muse in the Machine:

A computer that never hallucinates cannot possibly aspire to artificial thought.

While Gelertner speaks a somewhat different language than Hinton, I believe that the property of a machine that he is referring to here – the ability to hallucinate, fantasize or dream – is exactly the sort of thing Hinton is doing with his generative digit model. When you run that model I would argue that you are seeing the faintest wisps of the beginning of true cognition in a machine.

Douglas Hofstadter is probably the most famous of the four computer scientists I’ve been thinking about recently. He is of course the author of Godel, Escher, Bach, which every self-respecting technophile has read, but more importantly he has been a proponent for the need to think about cognition from a very different perspective than most computer scientists. For Hofstadter, creativity and analogical reasoning are the key points of interest he feels we need to understand in order to understand our own cognition. Here he is in the “Pattern-finding as the Core of Intelligence” introduction to his Fluid Analogies book:

In 1977, I began my new career as a professor of computer science, aiming to specialize in the field of artificial intelligence. My goals were modest, at least in number: first, to uncover the secrets of creativity, and second, to uncover the secrets of consciousness, by modeling both phenomena on a computer. Good goals. Not easy.

All four of these folks share a perspective that understanding how analogical thinking and creativity work is an important and under-studied part of building machines like us.

Recently we’ve been working on a series of projects that are aligned with this sort of program. The basic framework is introduced here, in an introductory tutorial.

This basic introduction is extended here.

One of the by-products of this work is a computing system that generates vivid dreamscapes. You can look at one of these by clicking on the candle photograph above, or by following through the Temporal QUFL tutorial, or by clicking on the direct link below.

The technical part of how these dreamscapes are generated is described in these tutorials.  I believe these ideas are important. These dreamscapes remind me of H.P. Lovecraft’s Dreamlands, and this from Celephais:

There are not many persons who know what wonders are opened to them in the stories and visions of their youth; for when as children we learn and dream, we think but half-formed thoughts, and when as men we try to remember, we are dulled and prosaic with the poison of life. But some of us awake in the night with strange phantasms of enchanted hills and gardens, of fountains that sing in the sun, of golden cliffs overhanging murmuring seas, of plains that stretch down to sleeping cities of bronze and stone, and of shadowy companies of heroes that ride caparisoned white horses along the edges of thick forests; and then we know that we have looked back through the ivory gates into that world of wonder which was ours before we were wise and unhappy.

I hope you like them.

Quantum computing and light switches

So as part of learning how to become a quantum ninja and program the D-Wave One, it is important to understand the problem that the machine is designed to solve. The D-Wave machine is designed to find the minimum value of a particular mathematical expression which I can write down in one line:

As people tend to be put off by mathematical equations in blogposts, I decided to augment it with a picture of a cute cat. However, unless you are very mathematically inclined (like kitty), it might not be intuitive what minimizing this expression actually means, why it is important, or how quantum computing helps. So I’m going to try to answer those three questions in this post.

.
1.) What does the cat’s expression mean?

The machine is designed to solve discrete optimization problems. What is a discrete optimization problem? It is one where you are trying to find the best settings for a bunch of switches. Here’s a graphical example of what is going on. Let’s imagine that our switches are light switches which each have a ‘bias value’ (a number) associated with them, and they can each be set either ON or OFF:

The light switch game

The game that we must play is to set all the switches into the right configuration. What is the right configuration? It is the one where when we set each of the switches to either ON or OFF (where ON = +1 and OFF = -1) and then we add up all the switches’ bias values multiplied by their settings, we get the lowest answer. This is where the first term in the cat’s expression comes from. The bias values are called h’s and the switch settings are called s’s.

So depending upon which switches we set to +1 and which we set to -1, we will get a different score overall. You can try this game. Hopefully you’ll find it easy because there’s a simple rule to winning:

We find that if we set all the switches with positive biases to OFF and all the switches with negative biases to ON and add up the result then we get the lowest overall value. Easy, right? I can give you as many switches as I want with many different bias values and you just look at each one in turn and flip it either ON or OFF accordingly.

OK, let’s make it harder. So now imagine that many of the pairs of switches have an additional rule, one which involves considering PAIRS of switches in addition to just individual switches… we add a new bias value (called J) which we multiply by BOTH the switch settings that connect to it, and we add the resulting value we get from each pair of switches to our overall number too. Still, all we have to do is decide whether each switch should be ON or OFF subject to this new rule.

But now it is much, much harder to decide whether a switch should be ON or OFF, because its neighbours affect it. Even with the simple example shown with 2 switches in the figure above, you can’t just follow the rule of setting them to be the opposite sign to their bias value anymore (try it!). With a complex web of switches having many neighbours, it quickly becomes very frustrating to try and find the right combination to give you the lowest value overall.

.

2.) It’s a math expression – who cares?

We didn’t build a machine to play a strange masochistic light switch game. The concept of finding a good configuration of binary variables (switches) in this way lies at the heart of many problems that are encountered in everyday applications. A few are shown in figure below (click to expand):

Even the idea of doing science itself is an optimization problem (you are trying to find the best ‘configuration’ of terms contributing to a scientific equation which matches our real world observations).

.
3.) How does quantum mechanics help?

With a couple of switches you can just try every combination of ON’s and OFF’s, there are only four possibilities: [ON ON], [ON OFF], [OFF ON] or [OFF OFF]. But as you add more and more switches, the number of possible ways that the switches can be set grows exponentially:

You can start to see why the game isn’t much fun anymore. In fact it is even difficult for our most powerful supercomputers. Being able to store all those possible configurations in memory, and moving them around inside conventional processors to calculate if our guess is right takes a very, very long time. With only 500 switches, there isn’t enough time in the Universe to check all the configurations.

Quantum mechanics can give us a helping hand with this problem. The fundamental power of a quantum computer comes from the idea that you can put bits of information into a superposition of states. Which means that using a quantum computer, our light switches can be ON and OFF at the same time:

Now lets consider the same bunch of switches as before, but now held in a quantum computer’s memory:

Because all the light switches are on and off at the same time, we know that the correct answer (correct ON/OFF settings for each switch) is represented in there somewhere… it is just currently hidden from us.

What the D-Wave quantum computer allows you to do is take this ‘quantum representation’ of your switches and extract the configuration of ONs and OFFs with the lowest value.
Here’s how you do this:

You start with the system in its quantum superposition as described above, and you slowly adjust the quantum computer to turn off the quantum superposition effect. At the same time, you slowly turn up all those bias values (the h and J’s from earlier). As this is performed, you allow the switches to slowly drop out of the superposition and choose one classical state, either ON or OFF. At the end, each switch MUST have chosen to be either ON or OFF. The quantum mechanics working inside the computer helps the light switches settle into the right states to give the lowest overall value when you add them all up at the end. Even though there are 2^N possible configurations it could have ended up in, it finds the lowest one, winning the light switch game.

The Developer Portal

Keen-eyed readers may have noticed a new section on the D-Wave website entitled ‘developer portal’. Currently the devPortal is being tested within D-Wave, however we are hoping to open it up to many developers in a staged way within the next year.

We’ve been getting a fair amount of interest from developers around the world already, and we’re anxious to open up the portal so that everyone can have access to the tools needed to start programming quantum computers! However given that this way of programming is so new we are also cautious about carefully testing everything before doing so. In short, it is coming, but you will have to wait just a little longer to get access!

A few tutorials are already available for everyone on the portal. These are intended to give a simple background to programming the quantum systems in advance of the tools coming online. New tutorials will be added to this list over time. If you’d like to have a look you can find them here: DEVELOPER TUTORIALS

In the future we hope that we will be able to grow the community to include competitions and prizes, programming challenges, and large open source projects for people who are itching to make a contribution to the fun world of quantum computer programming.

IEEE talk at Johns Hopkins University

I gave a talk at Johns Hopkins University on Monday entitled ‘Why the world needs quantum computing’. The video was for a general audience (no specific quantum physics background required). Here is a link to the talk. The video is available for download from this site!