Programming the Cutie Map: Big Data and Neural Nets · 3:36pm Feb 19th, 2017
This post is inspired by a comment by FanOfMostEverything on my last. It addresses the question of how The Map could function, or at least, how we would go about building such a powerful and useful tool in our world.
Twilight’s throne room centre-piece is evidently a lot more intelligent than your average coffee table. It is also imbued with wide-ranging senses. Somehow it can detect friendship problems across the kingdom and beyond, and then magically tell which pair of ponies is best qualified to take on the task of fixing it. It knows precisely what unique combination of skills will be required for a given mission, even before the ponies setting out know what they will encounter. It is as if it the strings are being pulled by some other-worldly intelligence, which somehow knows the end of the story before it has happened.
In the real world, this table is not something you can buy at IKEA, but it might be possible to build a similar indistinguishable-from-magic technology. There is a lot of interest in the idea of combining the data from social media streams, with the power of machine learning techniques, to identify similar problems. Take a look at this article: CIA using deep learning neural networks to predict social unrest five days before it happens. It describes how the CIA use algorithms to analyse data from social media posts, newspapers and web forums, to forecast incidents of social unrest. It’s not yet at the resolution of Twilight's map, but apparently it can “anticipate the development of social unrest and societal instability, some I think as near as three to five days out” according to a spokesman.
The idea of machine learning is to give a computer (or more often a farm of computers) a large amount of data, from which it can learn to make predictions by identifying patterns in the data leading up to past events. This is done using algorithms such as Artificial Neural Networks, modelled on the biological network of neurons inside our brains. Sophisticated deep learning methods use many hidden layers between the input and output. These can tackle some of the most formidable pattern recognition problems, which tax the most powerful computers, such as telling the difference between a picture of a dog and a pony. Given enough data to learn from, such networks can be very good at matching an input to an output. Unfortunately the process by which is does so is often so convoluted that it is impossible to tell how. All you can do is give it another set and see how well it predicts the outcome for that. If it gets it wrong, it can learn, and hopefully do better next time. To reach the point where it consistently performs well, you need serious amounts of Big Data.
To program a neural net to identify friendship problems, you could take a dataset of past Twitter and Facebook posts, with previous friendship problems flagged. You then let it learn a set of rules to link the two, then feed it a new set, and see how well it works.
But that’s only part of the job. The next task is to identify the best ponies to dispatch on the problem-resolution mission. Again this can be done with an artificial neural net. This time you need to add to the input a load of personal data on the aptitudes of the six ponies. This also has a real-world precedent, as it is a very common management problem to decide how to assign tasks to employees. Companies have developed software for this purpose. It is not usually clear how it is used or how effective it us. A popular strategy with many managers is to give the fun jobs to the staff members who say nice things about them, and give the tedious, demeaning work to any troublemakers. But when things go wrong, because people were assigned to tasks not fitting their skills and experience, they might get the blame. Better to have a computer to do the job, ideally one which a totally obfuscated logic, so they can ignore it when they want to, and if anything goes wrong, they can just blame it on the algorithm.
It would seem that Twilight’s map does a pretty good job at picking the right pair of ponies, so, if it is driven by such a machine learning approach, there must be a long story of it came by the data to learn how to do so. That seems perfectly credible given that it came out of the tree of harmony, which is somehow imbued with millennia of magic, tied to all the friendships in Equestria.
Thing is about computer based image recognition, is just how many people can describe an animal well enough to to have someone else draw it if they dont know what it is, besides drawing it themselves or recognising it when they see it.
That is, because it is so difficult, to me, its reasonable to assume that the way we actually recognise images etc, is almost totally different to teh way we wish and demand that we recognise images and so try and code computer the same way.
A favourite trick is that of speed artists. How many dots and lines are needed before an object is recognisable. Shouldnt that be the way a computer is trained to recognise images, recognising the important features and learning how to process out the 99.9999% of unneeded pixels through simple filtering?
Lets start with polar coordinates and circular, spherical harmonics, complex and quaternion FFT. You might say but this is difficult and complicated maths, but if it reaches a solution in a couple of approximations, as opposed to megacode for so called simpler stuff, then the more complex option is the simpler option. The reason for quaternion FFT is rotating things in and out of image plane can be detected by change in contrast, data complexity in one axis relative to another, and the function following the rotation by only the phase altering and not the magnitudes.
There is a simple model that describes how hidden layer neural nets work, which is Beysean math, but its how its implemented thats the tricky bit. People who use Beys almost always use continous, dense sequencing, but this leads to exponentially exploding data requirements. Thing is, if all you are doing is predicting the next one or two items, then all the prior items, no matter how long the sequence, have already been defined. Therefore in a discontinous, sparse Beysean sequence, you only need to look at the prior linear number of elements for each of the predicted values. That is, a classic Baysean prediction on 3 elements increases as the cube. a Sparse Beysean prediction on 3 elements increases as the double of the square. This immideately leads to the same processing requirements as cube, leads to sequence length similar, or for 8 bit text, 256 length sequence instead of 3 sequence. Approx.
The next trick comes from taking the database of pair elements, and creating elements of collections of pairs to form new pattern recognition accross multiple bases. You have databases consisting of how often various pairs of letters occur, such as start and end letters, and start letter and length of word etc, and the word neuron will have all these various entries as input values, the strength of which is the probability of occuring.
If I got my act together back in 2003, Id not only have the short script, code for linking to, but analysis results of various bits of data.. but, Im useless at coding and computers keep changing, needing reinstalls, updates, and other distractions. Had the thing working once for scanning a couple year back, but computer went down again.
Some nice stuff coming along by brute force. Be very intresting when the efficint algorithms turn up.
Huh. I didn't expect to inspire this blog, but it is fascinating. Given the structure of Twilight's castle, I can't help but wonder if the entire thing is performing the predictive processing, an immense, property value-raising chunk of crystalline computronium. It'd certainly be appropriate if even Twilight's castle couldn't stop thinking.
Or it just read the script, but what fun is that explanation?
4428452 After all, it's Twilight's castle, not Pinkie's!
Hmm, interesting.
4428452
That's basically what I assume it is And it's an extension of the incarnation on this plane of Harmony, in the form of the Tree, which is itself a vast living crystalline computer. Its intelligence is vast and its purposes benign, but it thinks in a rather alien fashion to (say) incarnations such as Celestia or Twilight, who are basically just super-powerful and ultra-intelligent Ponies.
4428452 (spoilers for Friendship is Sufficiently Advanced, for those that care)
That's certainly the explanation I'm using, in my firm-ish SF AU. And for that matter, the Tree of Harmony itself in the Advancedverse is basically a millennia-old Precursor-race planetary-management-and-uplifted-species-experiment-supervisory computer...
4428804
Ill give you a quick guide to how powerful the Castle can be. Many years ago I decided to go OTP and design a magitech computer, using woven gem fibres and light. Each foot strip has a thousand fibres, therefore each cubic foot block has three million fibres and a billion junctions. The shed carried a thousand such blocks. A Trillion processing elements, Each module can clock at least light speed, so approx 1 Ghz, with thousand deep pipelines. Equivalent. But, the optical processing methos means that all the junctions are entangled, so thats a trillion qbits at a terrahertz, per frequency of light used. It uses a wide spectrum of densely modulated carriers.
If you use atomic formation, crystal growth methods, each module would be less than ten microns on a side, and the processing speed up would be proportionate due to light not having to propagate so far. So at least thrity thousand times more powerful.
Each cubic millimeter is a thousand of those.
Each cubic meter is a Billion of those.
How many cubic metres makes up the Crystal Castle.
Dont forget the holographic memory crystals dangling from the tree roots. Those re pony made and so less dense.
The Crystal Castle also acts as a Fractal Antenna. Which means its far more efficint for its size than a linear tower. The bigger the antenna, the further and more accurately it can reach. And make use of other crystals accross the world.
Should be a bit of global reach Nikolai Tesla tech in there, but I blame Elon Musk.