101 Artificial Intelligence and Machine Learning

[intrinsical]

Many scientific fields have their own versions of the impossibility theorem. In Mathematics, there is Godel's Incompleteness Theorem which states roughly that any but the simplest toy mathematical system has facts that can never be proven or disproven. In Computer Science, Turing's Halting Theorem states that there are computer programs that will never stop running (In other words, are unable to solve certain tasks as they won't ever stop running or find a solution).

Just chiming in to point out that that's an incorrect statement of Turing's Halting Theorem. The Halting Theorem is a statement that a very specific kind of computer program does not exist, and the precise statement is perhaps somewhat technical. The Halting Theorem states that there is no Turing machine H operating on ordered pairs (F,X), such that for all inputs (F,X) we have that H(F,X) halts and computes whether the Turing machine encoded by F (under a fixed but standard encoding) halts on input X. The Halting Theorem (as the negative solution to the Halting Problem) is intimately related to Gödel's Incompleteness Theorem.

Yeah I was once a computer science undergraduate so Turing's Halting Theorem is something I have studied a long time ago. I was trying to provide the simplest possible explanation of the theorem with more focus on the theorem's implications rather than trying to explain the theorem's details, which would have had to gone into details about Turing Machines, Lambda Calculus and a bunch of first order predicate logic. As a result, I may have over simplified things and badly mangling the explanation. For that I apologize.

Let me try again... The simplest explanation I can give is probably this: In just about every field of science, we humans have proven that there bounds to what science and technology can achieve. There are some truths we can never know, some things we cannot do. It is a fundamental fact of nature that certain things contradict themselves and end up being impossible. And because science and technology is a study of nature, there is no way to achieve these things. This means in Math, there are certain things we cannot prove.

In Machine Learning, we can never come up with one single learning algorithm that is best at every single task we set it to accomplish. If you think about it, Machine Learning is all about learning from information. However, this information is always incomplete and we never have the complete set of information (ie, the true decision boundary). Frankly, if we had the complete set of information for a specific task, then we can already define a scientific theory (just like we have for Gravity, Quantum Mechanics, Natural Selection, etc) and can already make highly precise predictions without the need for Machine Learning. So in a way, every prediction a Machine Learning algorithm makes is a best guess. It is an educated guess based on what information it has available, but it is still only a guess. That is why we are talking about accuracy in percentages, because as every gambler knows guesses can sometimes be wrong.

This is part of why every single "intelligence", regardless how superior it is, can be beaten, because in most real world situations no one have the complete information and have to resort to making educated guesses. Even SkyNet has to guess. Sorry Singularity conspiracists, the technological singularity is just a purely theoretical conjecture that won't ever happen in real life.

 

[RAMA]

Ironically this interview with expert Ben Geortzel was reposted today and takes some wide ranging looks at AI, it's funding, who will get to strong AI first, and how fast it would be possible.

https://m.soundcloud.com/singularit...ingularity-1-on-1-the-future-is-ours-to-creat

 

[Mary Ann]

The link states that the track can't be found. Look into that please, RAMA.

 

[CorporalCaptain]

As a result, I may have over simplified things and badly mangling the explanation. For that I apologize.

No worries. I've done that, too.

Let me try again... The simplest explanation I can give is probably this: In just about every field of science, we humans have proven that there bounds to what science and technology can achieve. There are some truths we can never know, some things we cannot do. It is a fundamental fact of nature that certain things contradict themselves and end up being impossible. And because science and technology is a study of nature, there is no way to achieve these things. This means in Math, there are certain things we cannot prove.

For me, a straightforward intuitive understanding comes from considering what a diagonal argument is. Both Gödel's Incompleteness Theorem and Turing's Halting Theorem can be proven by diagonal arguments, so-called after Georg Cantor's diagonal method.

The basic intuitive idea in a diagonal argument is that the size of a set of codes to be put in correspondence with a set of objects is so much smaller than the set of objects, that no matter how you set up a correspondence between codes and objects, it will be possible to use the correspondence itself, in effect against itself, to prove that there exists an object that is not encoded. In this way, it is shown that no correspondence fully covers the objects with codes. There just simply "aren't enough" codes, given the size of the set of objects to be represented. (This is trivial when the sets are finite, so the intuition here is really relevant when the sets of codes and objects are both infinite.)

It's like you are trying to describe something that is vastly more complex than the language you have available to describe it. No matter how you define your terms, you will always fail to capture something. That's what I think is a good intuitive way to approach these particular topics, and I think it has merit even beyond situations in which diagonal methods directly apply.

 

[intrinsical]

Why Handwriting Recognition is (Relatively) Easy Compared To Face Recognition

Earlier, I alluded to the fact that handwriting recognition is much easier than other image recognition tasks such as face recognition. The main reason for this has to do with the decision space of the numbers. For example the number 3 looks very different from the number 4. The same is true when you compare the 784 input values. The 784 numbers representing an image of 3 is very different from the 784 numbers representing an image of 4. No one has actually measured the amount of distance, but I'ld imagine the amount of space is probably like the amount of space between Earth and Mars. This is why even an old classifier like Neural Networks from the 1980s could easily find a decision boundary to classify the images of handwritten numbers and characters.

I found a blog post that took the 784 dimensional datapoints and compressed them down to 2 dimensions so that we can have a rough look at how the datapoints for one number relates to datapoints for a different number. As you can see, there is almost no overlap in most number pairs. For example, 0s and 1s are clearly quite separate. The only number pairs with significant overlaps are 3s and 5s, 4s and 9s, 5s and 8s. This is to be expected as these numbers look visually similar.

When we tried the same approach of just using raw greyscale pixel values for other image recognition tasks (eg, face recognition), they never worked. This is because the decision space separating faces and non-faces is a lot narrower. If you examined the headshots of Kiera Knightly and a Cyberman, you can see there's roughly two dark circular patches where each eye should be and a thin horizontal dark patch where the mouth should be. To an algorithm that only know of pixel vales, these two headshots are nearly identical. And since Machine Learning algorithms works best when there is a clearly separable decision boundary, image recognition failed miserably despite many advances over the years.

Finding Better Image Features
Clearly if we wanted to do image recognition, we cannot simply use pixel values. Since the 1970s and 1980s, researchers have spent over 3 decades devising various ways of representing images as numeric features that offered more separation than just using pixel values would. Examples of applications include distinguishing faces from non-faces, humans from non-humans, cars from non-cars, planes, street signs, cats, dogs, etc...).

There are probably hundreds, perhaps even thousands of image derived features that have tried over the decades. Here are two significant techniques that had varying degrees of success in image recognition tasks.

Image Segmentation
An image can often be chopped up into different parts, such as the background as well as various foreground objects. The idea behind image segmentation is to, well, (A) segment an image into its constituent parts and (B) measure each segment's image properties such as the color, brightness, whether the texture is smooth or rough and so on. These properties are then feed these to a machine learning algorithm.

Scale Invariant Feature Transform (SIFT)
http://en.wikipedia.org/wiki/Scale-invariant_feature_transform
Invented in around 2004, SIFT was the first image feature that allowed us to reliably recognize objects in images. The concept behind SIFT is similar to that of fingerprint matching. To match fingerprints, unique points such as loops, arches, whorls are first identified. By comparing the angles and distances between these unique points, we can very reliably tell if two fingerprints belong to the same person.

SIFT does the same thing by identifying hundreds of interesting points in an image. Typically, these are edges and corners. By comparing the SIFT points in one image with SIFT points in another image, we can tell if the same object is in both images.

SIFT was quite a huge breakthrough that enabled a whole host of image applications that wasn't possible in the past. These applications include image stitching (eg, to create 360 degree panoramas from a series of separate shots) and augmented reality (SIFT was used to identify glyphs).

[yt]https://www.youtube.com/watch?v=YHT46f2244E[/yt]

[yt]https://www.youtube.com/watch?v=caFHvamMUTw[/yt]

Since 2004, there has been an explosion of similar image features with weird abbreviations such as SURF, GLOH and HoG. These newer techniques are either faster than SIFT or offered even better recognition rates.

These relatively recent advances in image features allowed us to achieve better accuracy in a wide range of image recognition tasks. Here are a few examples that power autonomous driving vehicles.

Recognizing road lanes and potential obstructions ahead of the vehicle.
[yt]https://www.youtube.com/watch?v=JmxDIuCIIcg[/yt]

Recognizing road signs
[yt]https://www.youtube.com/watch?v=hU7yHQkg-7U[/yt]

Is Image Recognition a Solved Problem?
From the advances I have talked about in the prior post, it does look like we have cracked the image recognition code. Or have we? Very early in this thread, I mentioned that the goal of Artificial Intelligence is to get the machine and not humans to do the work. Yet, it took two generations of computer vision and artificial intelligence researchers 30 years to "crack" image recognition. The same is true of in many other domains.

For example in Natural Language Processing tasks such as translating sentences from one language to another language, we first have to generate a whole host features. W have to identify the grammar of each word (is it a noun, verb, pronoun, etc...) in the sentence. We have to figure out if a word or a string of word is a name. We have to incorporate synonyms , hyponyms and hypernyms. We have to figure out "he" in this sentence refers to "Tom" in the previous sentence. We have to figure out the structure of a sentence. Only after having created all these as input features can we actually start applying the machine learning algorithm to translate a sentence in one language to another language.

To many researchers, spending decades doing "input feature engineering" makes modern Machine Learning no different from the old logic-centric knowledge based systems of the 1970s and 1980s. We are still wasting a lot of human effort and time in creating these systems. We have completely failed in the goal of getting the machine to do all the work.

in 2006, we asked a big question. Instead of us doing all the work thinking up good features, can we get the machine to do it for us? Thus Deep Learning was born and Neural Networks saw a sudden revival.

 

[Dryson]

intrinsical

Time for a pop quiz.

The image below shows a simple dataset with two classes, depicted by red and green crosses. Also in the image are two decision boundaries, one with a blue line and the other with a purple line. Both decision boundaries accurately split the dataset into two parts with 100% accuracy, but one decision boundary is better than the other. Can you tell me which decision boundary, blue or purple, is the better decision boundary? And why is it better?

It is possible to dataset with class values (x)between opposing sets where each right triangle could be used to create a class value of distance between each class of each data set to determine a rate of electrical energy consumption and allocation of energy consumption so that when each class of each data set has been figured out to be true or false and residing along a similar path of thought, each path of thought would have its class pathway or line based on the class types below, of being assumptively incorrect (base of theory not pertaining to any known proven theories) partially incorrect (not enough data to prove correct, more incorrect data than correct), incorrect / assumptively correct (base of theory pertaining to any known proven theory but not proven by said theory), partially correct (note enough data to prove correct, more correct data than incorrect data), correct so that a pattern of logical and illogical thoughts could be created for the AI to then figure out?

We also know that a simple logic gate has two values 0 means OFF or the gate is open (kinda odd when you think about a gate being open and allowing something through) and 1 means ON or the gate is closed (kinda odd when you think about a gate being closed and not allowing something through)

There should be sub-logic gates such as 01(sup.) for assumptively incorrect, 0 for incorrect and 01(sub.) for partially incorrect. The same values would be assigned for the 1 as well. See image:

https://www.mediafire.com/convkey/d653/48fhmqmr4q10bmf6g.jpg

The classes would be 01(sup.) - The grass is green. 0 - The grass is green because I believe it to be green. 01(sub.) - The grass is green because I see the color green. 1- (sup.) The grass is green because of water. 1 -The grass is green because it has lots of chlorophyll, which is needed for photosynthesis. 1(sub.) The grass is green because it needs water.

 

[Amaris]

The big yellow one is the Sun.
The big yellow one is the Sun.

 

[intrinsical]

Shallow Learning vs Deep Learning
As I have alluded in my last post, whether a specific Machine Learning task is solvable depends heavily on the input data. If there is a high degree of correlation (PS: I am not referring to the mathematical/statistical concept of correlation) between the input data and the classification label, then a straight forward application of any number of Machine Learning algorithms I have described so far would very likely provide you with a good to high accuracy solution. If however, the input data does not have a clear correlation with the classification label, then we as human beings will have to somehow transform the input data into better input features before we can apply Machine Learning algorithms. These days, we refer to this type of learning task where there is a direct correlation between input data and the classification label as "Shallow Learning". The word "Shallow" implies that this type of problem is simple and easily solvable by modern Machine Learning techniques.

Conversely there are classification tasks where the input data has some connection to the classification label but there is no direct correlation. This often means that it may be difficult (but not impossible) to find a good decision boundary to separate the classes. Most image recognition (eg, face recogntion, object recognition) tasks fall under this category. In such cases, we have two options. We can hand-craft some kind of data transformation step to turn the original source data (eg, raw pixel values in images) into some better set of input features (color, changes in brightness, texture properties, SIFT, etc..). In 2006, several prominent AI researchers suggested a radically different approach that is now collectively referred to as "Deep Learning". Basically with the Deep Learning approach, we let the computer automatically figure out the data transformation step rather than manually hand-craft the data transformation from raw pixels into input features.

I want to emphasise that while the end goal of Deep Learning is still on improving our ability to do the Machine Learning task of classification with higher accuracy, its focus is not on finding decision boundaries. Instead, its focus is on the transformation step of turning raw data into something better, something would result in a decision space where it is easier to separate the classes (eg, faces from non-faces). In fact, a lot of Deep Learning systems still use algorithms like Support Vector Machines to do the final step of classification.

Face Recognition with Deep Learning

Being a science fiction forum, I am sure many of you have seen this picture compositing the faces of all twelve actors of The Doctor from Doctor Who into a single image. In some sense, this composite is the "average" face of The Doctor. If we wanted a Machine Learning algorithm to detect The Doctor's face, this image would be an excellent input feature. All we have to do is compare the pixel values of our source image against this composite image of all 12 Doctors.

And this is the type of feature that Deep Learning ends up with when tasked with doing face recognition, a set of prototypical faces. As you can probably imagine, there would be lots of matches if we took an image of a face (eg, Kiera Knightly's face from my last post) and compare it against our set of prototypical faces. This implies that using these prototypical faces as features, we should easily find a simple and separable decision boundary using SVM or even just a simple linear classifier. Now the question is, how do we end up with this set of prototypical faces?

Stacked Sparse Overcomplete Ensembles
Deep Learning not a single technique. It is more of a general methodology that combine four concepts into a single workable and highly succesful idea.

Sparsity in mathematics refers to the idea that data can be broken down more atomic parts. For example, images of faces can be broken down into eyes, nose, mouth, hair and general facial structure. Or at a lower level, any image can be broken down into various kinds of shading and textures.

Overcompleteness is another mathematical concept that in inprecise but simple English means having lots of lots of examples. So in face recognition, this means having lots of different varieties of eyes, noses, mouths, hairs. Typically, Deep Learning systems will try to extract anywhere from 10,000 to 500,000 such examples.

Ensemble Learning is a Machine Learning concept that I have touched upon in an earlier post. If you recall, it is about leveraging the wisdom of a crowd. By having a large and diverse selection of eyes, noses, mouths and so on, we can essentially do a mathematical mix-and-match to assemble a large variety of prototypical faces.

Stacking is an Machine Learning concept that has its roots in Multi-Layered Neural Networks from the 1980s. In the old days, it was discovered that Neural Networks which had multiple layers of neurons sometimes performed better than having the same number of neurons arranged in a single flat layer. With Deep Learning, Stacking refers to repeatedly transforming the data. In face recognition, the first layer in the stack transforms raw pixels into basic shading, lines and edges. The second layer takes these lines, edges and shading and assemble them into commonly seen variety of eyes, noses, mouths and ears. Finally, the third layer in the stack takes the eyes, noses, mouths and assemble them into commonly seen faces.

In terms of actual mathematical techniques that combine these four concepts into a successful Deep Learner, there are actually several different variants that are all considered Deep Learning algorithms. There are Stacked Autoencoders, Convolutional Neural Networks, Restricted Boltzmann Machines. The mathematics involved is somewhat different for each of the techniques I have mentioned, but the general strategy for all of them are the same - Stacked Sparse Overcomplete Ensembles. .

The approach of Deep Learning works very well on a variety of image and audio related tasks. In 2012, Google took 10 million snapshots taken from youtube videos and thrown them at at 16,000 servers that were configured to use Deep Learning training. The end result is a system that could identify different types of objects including humans, animals, furniture, vehicles, etc... It wasn't very accurate, but this was the first system that could classify 10,000 different types of objects. In 2014, Facebook trained a face recognition system that combined Deep Learning, Ensemble Learning and Support Vector Machines to achieve a whopping 97.35% accuracy. Compared with the human benchmark of 97.53% accuracy, I would say that Machine Learning has achieved near human level of performance in face recognition.

 

[intrinsical]

Other than myself, it has been kind of quiet in this thread. So I'm starting to wonder if anyone is following this or if I should stop?

 

[CorporalCaptain]

Other than myself, it has been kind of quiet in this thread. So I'm starting to wonder if anyone is following this or if I should stop?

Well, I've substantively participated in the thread. Sorry if my participation hasn't been of a kind and quantity to warrant mention.

 

[Mary Ann]

I haven't yet posted in this thread because this is well outside my personal knowledge, so I have nothing to add. However I have learned a lot, and find the subject fascinating. I think I speak for several other people when I say please don't take our silence as a sign of our not being interested. Quite the opposite; you explain the subject so well there is hardly any need for clarifications. Please do continue!

 

[Robert Maxwell]

I'm still reading, I just haven't been around much. I hope you keep it up!

Maybe you need to offer another exercise.

 

[Dryson]

I think its a rather interesting topic.

The only question is how do you create a machine brain that focuses on many aspects of its environment where one time the machine brain encounters an environment such as a tree, a sandbox, a swing set and a teeter tauter with one set of actors and emotions that when the same environment is encountered again with the same actors but with different emotions that the machine brain is able to learn from past experiences to form a different thought and not walk off towards the woods because it anticipates something happening in the second set that occurred in the first set?

 

[Robert Maxwell]

I think its a rather interesting topic.

The only question is how do you create a machine brain that focuses on many aspects of its environment where one time the machine brain encounters an environment such as a tree, a sandbox, a swing set and a teeter tauter with one set of actors and emotions that when the same environment is encountered again with the same actors but with different emotions that the machine brain is able to learn from past experiences to form a different thought and not walk off towards the woods because it anticipates something happening in the second set that occurred in the first set?

What???

 

[intrinsical]

Thanks for the replies, I really appreciate it. Dryson, I don't understand your question so I'm afraid I cannot answer it.

As you may have noticed, I have already covered the development of Machine Learning from 1970s all the way to the 2010s. The next post will be more of a personal opinion piece on the short term future of Machine Learning. I don't really have any plans beyond that point. So unless you have any questions, general or specific, this thread may be winding down soon.

As a side note, given the things I have worked on in the past and my personal interest areas, I can probably offer more insights on how these things work:
- Search Engines
- IBM's Watson/Apple's Siri
- Autonomous Vehicles/Drones
- Stock Markets predictions (not high frequency trading)

 

[Tim Walker]

Article regarding unconventional computing.

 

[intrinsical]

I realized that while I have described what Deep Learning is trying to achieve, I have not described HOW it does its magic. So consider this as Deep Learning, Part 2. Before we begin, here's a couple of reminders.

Lookback #1: Images are Datapoints
This is really just a reminder that all images can be thought of as datapoints in a high dimensional space of all images. Each example image of a handwritten digit, face, car, elephant or chair is a single point in this image space. Each feature image generated by Deep Learning algorithms such as lines, edges, eyes, noses, mouths, prototypical faces, prototypical cars are also datapoints in the image space.

The Space of All (Greyscale) Images

And this space is a standard Euclidean Space. The only significant difference is that instead of the usual 2D or 3D space that most people think about, this space has hundreds of dimensions (basically one dimension for each pixel value). Don't let this scare you as all the standard 2D/3D geometry math we learnt in high school works exactly the same way in this high dimensional space. We can still calculate Euclidean Distance between two images or calculate the angle between two images. More importantly for Machine Learning algorithms, we can continue to define decision boundaries around groups of images.

Lookback #2: Supervised Learning
Other than Deep Learning algorithms, every single Machine Learning algorithm (Logistic Regression, Decision Trees, NeuralNets, SVM) I have talked about so far fall under the category of Supervised Learning. Specifically, the algorithms I have talked about all supervised Classification algorithms that takes two inputs:

(A) A set of datapoints (which can be images)
(B) Labels for each datapoint (Yes/No, True/False, Face/Not-Face)

And output a decision boundary.

The reason why these algorithms are called Supervised Learning algorithms is because of (B), the labels we provide for every datapoint. By providing the labels, we humans are basically acting as a teacher who provides the student algorithm with all the correct answers and hoping that it will learn from the examples we have provided. There are Machine Learning algorithms that do not fall under the category of Supervised Learning.

Unsupervised Learning
In Unsupervised Learning we only provide the algorithm with datapoints, but WITHOUT any labels.

Unlike Supervised Learning whose goal is to find a good decision boundary, Unsupervised Learning is more about learning the "structure" of the data. For example, by finding "clusters"; regions where the datapoints are clumped together densely. One of the oldest and most commonly used Unsupervised Learning algorithm is called K-Means.

K-Means works by starting with K randomly selected points in the decision space and pulling these randomly selected points towards the datapoints. You can think of it as each datapoint acting as if it had gravity. If we do this an infinite number of times, these randomly selected points will naturally be pulled to the center of the denseest clusters. In practice, we can't pull the K points an infinite number of times since that would take forever. However we can just keep repeating the pulling until the K points don't change too much, which imply we're very close to the centers of the dense clusters. If for example K=2, we will end up with two clusters that look something like this.

Prior to 2006 and the introduction of Deep Learning, that's pretty much where Unsupervised Learning stopped. We could find dense clusters, which is well and nice. It was deemed important enough to be taught to undergraduates but in practice are rarely used. The reason is because knowing the location of big dense clusters of datapoints does not result in state of the art, highly accurate systems. The biggest reason is simply that Supervised Learning with the advantage of knowing the label of each datapoint always signficantly outperform all Unsupervised Learning learning systems that do not have the label. So Unsupervised Learning is only used in rare cases when either there are no labels or when the cost of collecting the labels for each datapoint is prohibitively expensive.

Another more mathematical reason is that algorithms like K-Means could only find big dense clusters but not the fine microstructure of the data itself. For example, many of the older Unsupervised Learning algorithms like K-Means cannot correctly cluster the "two-moons" and "three-spirals" datapoints.

 

[Tim Walker]

It is expected that Moore's Law will end due to practical problems-we won't be able to make computer chips any smaller. We will then reach a technological plateau.

A few years ago there was an article in Scientific American....

In an experiment, DNA computing solved what was described as the traveling salesmen problem. The basic computing element-the DNA molecule-is at a molecular scale, permitting massive parallelism.

 

[Robert Maxwell]

It is expected that Moore's Law will end due to practical problems-we won't be able to make computer chips any smaller. We will then reach a technological plateau.

A few years ago there was an article in Scientific American....

In an experiment, DNA computing solved what was described as the traveling salesmen problem. The basic computing element-the DNA molecule-is at a molecular scale, permitting massive parallelism.

You can read about the latter here, which is fairly interesting.

It's worth noting that their approach involved a brute force search, so it isn't an algorithmic improvement so much as it is a problem to which genetic programming (on actual DNA) is well-suited. Errors become more and more frequent as complexity is added in a DNA-based solution, so brute force remains the most expedient approach using that method. It must be pointed out that brute forcing every possible solution to a given problem is basically the opposite of artificial intelligence, though.

Incidentally, rapid exhaustive space searches are the main reason why quantum computing is expected to ruin traditional cryptography (and why quantum cryptography is already in development.)

 

[intrinsical]

There are lots of algorithms that solve the Traveling Salesman Problem (TSP). In fact, it is the single most examined problem in Computer Science. And I have seen solutions that use not just DNA, but also water, light and even quantum mechanical effects.

The problem is that TSP falls under a class of problems that Computer Scientists refer to as NP-Complete (Nondeterministic Polynomial Time problems). Actually, I have understate the case for TSP as it is not just NP-Complete, it is NP-Hard. That is, the amount of time required to solve a TSP grows exponentially as we increase the number of cities. This means that if, for example, our largest and fastest computer farm running our fastest known TSP algorithm can find a solution for 100,000 cities in a year, then it may take a million years years to find a solution for 100,001 cities. ie, all our solutions cannot scale up. So even if the DNA-based algorithm is massively scalable, it still faces the exact same scalability constraints as traditional Von Neumann machines (aka computers).

As for quantum computing, I have been keeping an eye on it for some time as it CAN potentially solve TSP without facing the scalability issue. However, I'm not holding too much hope of seeing anything other than baby/faux quantum computing machines in my lifetime. What we have now are quantum cryptographic (leveraging quantum entanglement "spooky action at a distance" effects) and quantum annealing (leveraging quantum tunneling effects) machines that use quantum mechanical effects to achieve quite amazing feats in their own right. However, those machines probably cannot be converted into a quantum computing machines as such a machine probably has to rely on quantum superposition effects.

This doesn't mean all is lost as Quantum Annealing Machines are of particular interest to AI researchers, because they can potentially speed up a lot of our AI algorithms such as neural networks training. For example, D-Wave's CTO had posted this blog article that contains the quantum mechanical math required to potentially solve Deep Learning algorithms. The problem with D-Wave is that its quantum annealing machine is currently only as fast as a standard computer at solving some simpler AI problems. Still, there's enough potential for one to be hopeful. PS: This is probably the reason why Google is so intensely interested in D-Wave's machine.