Data Science and Learning Models

Data science and learning models are rapidly creating a gateway to a different reality. Today, you can get 92% accuracy with voice recognition. Learning models can recognize images, which means they can look for sunrise and find you pictures that have sunrises inside them. A learning model can watch videos and find specific words spoken within them. Learning models within Siri, Alexa and google assistant can have natural conversations with you. They can translate over a 100 different languages for you. A learning model can play you songs you’ve never heard before .. but those that you will really like. A learning model on your watch can recognize what exercise you are doing. A learning model can anticipate what time you need to wake up to minimize traffic on your route to work that morning and a learning model can recommend you your perfect match. All of these learning models are algorithms. Not the algorithms that most of us grew up writing, but those that learn and predict. The algorithms of data science and machine learning.

At its core, data science is about predictions using data. And these predictions can be within a variety of contexts. For example, predicting whether a piece of content is relevant to a person is a context that’s different from predicting how much a driverless car should slow down on that bumpy road. And because the success of prediction depends on knowing as much as possible, in data science we often deal with large amounts of data – popularly known as Big Data. Today, predictions using Big Data have become so successful that many subfields of artificial intelligence are leveraging these techniques. In the past, AI was originally about writing rules that mimicked the human brain, and they were faced with limitations of the combinatorial explosion. The reason why AI has taken a renewed uprising is because machine learning models with big data are creating similar outcomes that were originally envisioned in AI.

Data science is an emerging, interdisciplinary field, so the term could seem ambiguous and change based on the context it’s being applied.. There are many names linked with “data science”.. such as: Data Science, AI, machine learning, pattern recognition, statistical modeling, data mining, knowledge discovery, predictive analytics, adaptive systems, self organizing systems, etc. All these words lead to the same thing : predictions using data. Going forward, in this article I will use the term learning models as the vehicle that’s used to make predictions in all these areas.

Our society is changing, one learning model at a time. Learning models are remaking science, technology, business, politics and even war. Satellites, DNA sequencers, and particle accelerators explore nature in deep detail, and learning models turn the deluge of data into new scientific knowledge. Driverless vehicles pilot themselves across land, sea and air and the stock market is no longer a human endeavor, it’s mostly learning models that taught themselves to trade stocks with other learning models, and as a result, the floor of the stock exchange isn’t filled with traders doing their day jobs anymore, it’s largely a TV set.

With such a large impact, there is a need for all technologists to understand the concepts behind learning models: how they work and how they are built. However, the thought of getting into the details of learning models could seem a little daunting. At least it was to me when I first started this journey. However, just like any other technology paradigm, this field is also made up of a few building blocks. Understanding those building blocks could help us wrap our heads around the basics, and I have tried to pull the building blocks together from three excellent books:

• The Master Algorithm, by Pedro Domingos. Dr Pedro Domingos, is the co-founder of the International Machine Learning Society, the author or co-author of over 200 technical publications in machine learning, data mining, and other areas. He is the winner of the SIGKDD Innovation Award, the highest honor in data science.
• Data Science for Business, by Foster Provost. Professor Provost recently retired as Editor-in-Chief of the journal Machine Learning after 6 years. He is a member of the editorial boards of the Journal of Machine Learning Research (JMLR) and the journal Data Mining and Knowledge Discovery. He is also the co-founder of the International Machine Learning Society.
• Python Machine Learning, by Sebastian Rashcka. Sebastian Raschka is a PhD student at Michigan State University, who develops new computational methods in the field of computational biology. He has been ranked as the number one most influential data scientist on GitHub by Analytics Vidhya.

My goal is for you to be able to have a general understanding of learning models, and as a result, the core of data science in the next 60 minutes. Hang tight and keep reading!

Most of the knowledge in the world in the future is going to be extracted by machines and will reside in machinesYann LeCun, Director of AI Research, Facebook

Where does knowledge come from? Until recently it came from 3 sources:
Evolution: that’s the knowledge that’s encoded in your DNA
Experience: that’s the knowledge that’s encoded in your neurons, and
Culture: which is the knowledge you acquire by talking with other people, reading books, and from the internet. Everything that we do comes from these three sources of knowledge.

Only recently, there’s a fourth source of knowledge on the planet. Computers. There’s more and more knowledge now that comes from computers, which is discovered by computers.

Notice, also, that each of these forms of knowledge discovery is orders of magnitude faster than the previous one and discovers orders of magnitude more knowledge. And indeed, the same thing is true of computers. Computers can discover knowledge orders of magnitude faster than any of these things that went before and that co-exist with them and orders of magnitude more knowledge in the same amount of time. So how do computers discover new knowledge?

Pedro Domingos has classified 5 ways of in which computers discovered knowledge. We will cover each of these and their applications in more detail :

1. The first way computers discover knowledge is by filling gaps in existing knowledge. Pretty much the same way that scientists work, right? You make observations, you hypothesize theories to explain them, and then you see where they fall short. And then you adapt them, or throw them away and try new ones, and so on. This is based on logic and is the closest to the algorithms that many of us grew up writing. The core algorithms used in this form of knowledge discovery are related to inverse deduction.
2. The second way computers discover knowledge is by emulating the brain, the greatest learning machine on earth. Through the understanding of the brain from neuroscience and reverse engineering it. The core algorithms used in this form of knowledge discovery are related to backpropagation.
3. The third one is by simulating evolution. Evolution, by some standards, is actually an even greater learning algorithm than your brain is, because, first of all, it made your brain. It also made your body. And it also made every other life form on Earth. So this is based on evolutionary biology. The core algorithms used in this form of knowledge discovery are related to genetic programming.
4. The fourth one is based on the fact that all knowledge is uncertain. When something is induced from data,you’re never quite sure about it.So the way to learn is to quantify that uncertainty usingprobability.And then as you see more evidence,the probability of different hypotheses evolves. The core algorithms used in this form of knowledge discovery are related to the Bayes Theorem.
5. The last approach, is through reasoning by analogy. Humans do this all the time. You’re faced with a new situation, you try to find a matching situation in your experience, and then you transfer the solution from the situation that you already know to the new situation that you’re faced with. The core algorithms used in this form of knowledge discovery are related to Support Vector machines.

#1 Learning Models based on Logic

Deduction means going from general rules to specific facts. The opposite of that, hypothesizing general rules from specific facts is called induction and is the essence of learning models based on logic.

Let’s take an example of 6 data points

1. Star wars is a science fiction movie
2. You like Star Wars
3. Star Trek is a science fiction movie
4. You like Star trek
5. Avatar is a science fiction movie
6. You like Avatar

The generic hypothesis derived from these facts (i.e. induction) is : You like science fiction movies.

Once rules like this are derived from data, they can be combined in all sorts of different ways to answer questions that have never been thought of. Now these examples are in English. So to induce rules from structured and unstructured data, computers use first order logic (predicate calculus) to represent facts and rules. The gaps in knowledge are replaced by general principles which are then validated with data and readjusted. So as an example, if there were these new data points

1. Aliens is a science fiction movie
2. You don’t like Aliens

Here the “gap” is something about the movie Aliens (maybe because it’s also in the horror genre). So the generic hypothesis of You like science fiction movies is refined. This continues till as many gaps as possible are filled based on new instances of data to get to the most accurate hypothesis as possible to predict a particular outcome.

One of the amazing examples of inverse deduction is a Robot scientist, which is a complete automated biologist, It starts out with basic knowledge of molecular biology, DNA, proteins, RNA, and all of that stuff. And then what it actually does is it formulates hypotheses using inverse deduction.It designs experiments to test this hypothesis using things like DNA sequences etc. It physically carries out the experiments with no human help. And then, given the results, it refines the hypotheses,or comes up with new ones, and so on. Now there’s only one of these robots in the world today, its name is Eve. And in 2014 Eve discovered a malaria drug. And the thing that’s amazing about this is that, well, once you’ve made one robot scientist like this,there’s nothing stopping you from making a million. And if we have a million scientists working on a problem that only a few were working on before, it could really speed up the progress of science.

Some algorithms are based on this concept are decision trees, random forests, boosted trees, rotation forests, etc.

While learning models based on logic work well to solve some problems, in many cases they might be too abstract and rigid. If you take an example such as image recognition, there could be billions of variations of the same (let’s say a cat’s) image and a rigid method that uses induction based on logic will likely fail in those predictions.

Refer to this post for a deeper dive into logic based learning models

#2 Learning Models based on reverse engineering the brain

Learning models such as neural networks, deep learning all rely on the paradigm of reverse engineering the brain. So how does this work?

The brain is made of neurons and in order to reverse engineer it what we’re going to do is we’re going to build a model, a mathematical model of how a single neuron works. We’re going to make it as simple as we can, provided it’s enough to learn and to do the inferences that we need it to do. And then we’re going to put these models of neurons together into big networks.And then we’re going to train those networks. And at the end of the day, what we have is, in some sense, a miniature brain that’s much simpler than the real one, but hopefully with some of the same properties.

Now, a neuron is a very interesting kind of cell. It’s a cell that actually looks like a tree (see figure A). There’s the cell body, and then there’s the trunk of the tree is what is called the axon, and then the branches are called dendrites. But where neurons get very different from trees is that the branches of one neuron actually connect back with the roots of many others. And that’s how you get a big network of neurons.

Where the dendrites of one neuron join the dendrites of another, that’s called the synapse (see figure C). And to the best of our knowledge, the way humans learn everything we know is encoded in the strings of the synapses between our neurons. Neurons communicate via electric discharges down their axons. They’re literally an electric discharge called an action potential. And what happens is that if you look at all the charge coming into a neuron through its various synapses, if it exceeds a certain threshold, then that neuron itself fires. And of course, then it sends currents to the neurons downstream. The synaptic process itself involves chemistry and whatnot, but those details are not important for us. So learning basically happens when a neuron helps to make another neuron fire. And then the strength of the connection goes up. This is how all our knowledge is encoded, it’s in how strong the synapses are. If the neurons fire together a lot, then the synapses become stronger and it becomes easier for the first neuron to fire the second one.

Now what we have to do is, first of all, turn it into a model. What our neuron’s going to do is it’s going to do a weighted combination of the inputs. Let’s suppose, for example, that the figure B was representing a neuron in a retina. And let’s assume each of the inputs x₁, x₂.. is a pixel. Each input gets multiplied by a weight that corresponds to the strength of the synapse to the neuron. And if that weighted sum exceeds a threshold computed by f(x) then we get 1 as the output, otherwise we get 0. So, for example this neuron is trying to recognize a cat, if this is a cat then hopefully what happens is that this weighted sum will be high enough that this will fire and the neuron will say yes, this is a cat.

Now, a single neuron cannot perform a complex task like recognize a cat, what we need is a network of neurons. Now how do you train a whole network of these neurons (figure D)? The inputs go to one set of neurons, each of which computes a function, and then those go to another layer and many many layers until finally you get the output.

But what happens if there was an error? A neuron should have been firing, but wasn’t. The key question is, what do I change in that whole big, messy network to try to make it give the right answer tomorrow?
There is no obvious answer to this question. Because, think of one neuron somewhere in the middle of the network. How is it responsible for the error at the output? The error at the output could have come from an infinitude of different places. This is called the credit assignment problem. And this is the problem that backpropagation solves.

How does backpropagation work? Well, let’s think of the difference between our actual output and our desired output. Let’s say the picture was that of a cat and the neural net identified it as a dog. We can call the numeric difference between a cat(1) and a dog(0) as a delta. This is the error. The output should have been 1 (cat), but let’s say it was 0.2. So it needs to go up. What can we tweak in these weights to make the value go up from 0.2 to 1? Well, at that last layer, the neurons with the highest weights are the ones most responsible for the result. And if one neuron is saying 0 but the answer is 1, well, then its weight needs to go up, because it’s preventing it from firing. And if another one was saying 1 but is not firing, then its weight needs to go down. So what we do is compute the derivative of this error with respect to the weights in the last layer.
So now we know an error signal at the neurons at this layer. And I can keep doing the same thing all the way back to the input. And this is why it’s called backpropagation. Because what we are doing is propagating back the errors and then updating the weights, changing the weights in order to make that error as small as possible. This is the backpropagation algorithm. It’s what’s at the heart of deep learning. And these days, this is used for just about everything on Earth.

Very early on, people used backpropagation to do things like predict the stock market. These days, you use it for search, ad placement, video recognition, speech recognition, simultaneous translations and several applications of neural nets are being introduced everyday.

Some algorithms are based on this concept are static neural nets, dynamic neural nets, recurring neural nets, memory networks, deep belief networks, etc

While neural nets may make several types of predictions, but what about things like commonsense reasoning? It involves information that have never been seen together. If there’s a pothole, should a car drive over it? If a self driving car was only trained with backpropagation, then it would only be able to respond to situations it encountered before. If it was trained on a highway then it would run over Manhattan potholes all the time!

#3 Learning Models based on evolution

So where did that brain come from? That brain was produced by evolution. So maybe a powerful learning model can be created by simulating evolution. These are called genetic algorithms. How do these work?

With evolution you have a population of individuals, each of which is described by its genome. And then each of these individuals gets to go out in the world and be evaluated. It gets evaluated at the task that it’s supposed to be doing. And the individuals that do better will have a higher fitness and will therefore have a higher chance of being the parents of the next generation. You get two very fit parents and you cross over their genomes. And so now you have a child genome that is partly the genome of one parent and partly the genome of the other parent. And then you also have random mutation. Some bits just get randomly mutated because of copying errors, etc which also creates a new population.

We could actually start out with a population of simple “components” that is essentially random. And when these components randomly try to work with other components to produce a particular outcome, they could have various degrees of success or failure. After some number of generations of this, we actually have things that are doing a lot of non-trivial functions. For example, we can evolve circuits that start completely random and end up with radio receivers and amps and things like that in just this way. And they often work better than the ones that are designed by humans. In fact there are a whole bunch of patents where the designs were actually invented by genetic algorithms and not by humans.

So let’s try to apply this to computer programs. A program is really a tree of subroutine calls all the way down to simple operations like additions, multiplications, ANDs and ORs. Let’s represent the problems as trees.
Then in order to cross over two programs (figure a) , what we’re going to do is we’re going to randomly pick a node in each of the trees and then we’re going to swap the sub-trees at those nodes. And we can then evaluate the “fitness” of this new program in producing a result. Various techniques like this can be applied, while re-evaluating the fitness, keeping the combinations that are most fit and discarding the others. And then further mutating or crossing over to ultimately lead to a program best suited to produce a result.

In fact, these days, what the genetic folks are having a lot of fun with is something that’s exciting and scary at the same time. They’re not just doing this with software anymore. They’re not doing it as a simulation inside the computer,but actually doing it out in the physical, real world with robots. You literally start with robots that are random piles of parts (in software). And then once those robots are good enough, they actually get printed by a 3-D printer. And then they start crawling, and walking, and doing things in the real world– seeing how fast they can crawl, trying to recover from injury, and so on and so forth. Then what happens is that in each generation, the fittest robots get to program the 3-D printer to produce the next generation of robots!

While learning models based on evolution can be very powerful it has an inherent problem, it’s too slow. If you think about biological evolution, it took billions of years. But learning algorithms should be able to learn in minutes or seconds. Also many of the products of evolution have many obvious faults. For example the brain has many constraints that computers don’t, like limited short term memory. There is no reason to stay within these constraints.

#4 Learning Models based on uncertainty

The basic idea here is that everything that we learn is uncertain. So what we have to do is compute the probability of each one of our hypotheses and then update it as new evidence comes in. And the way we do that is with Bayes’ theorem. So how does Bayes’ theorem work?

The idea of the Bayes’ theorem is this. Let’s suppose that you have all your hypotheses. You define your space of hypotheses in some way : it could be a neural net or a decision tree or any other mechanism. And now the first thing that you’re going to have is the prior probability for each hypothesis. This is how much you believe in that hypothesis before you’ve even seen any data. So the prior is how much you believe in each hypothesis before you see the evidence. But then what happens is that as the evidence comes in, you update the probability of each hypothesis. A hypothesis that is consistent with the data will see its probability go up. A hypothesis that is inconsistent with the data will see its probability go down.
The consistency of the hypothesis of data is measured by what’s called the likelihood function, which is the probability of seeing the data if the hypothesis is true. And this theory is saying that if your hypothesis makes what you’re seeing likely, then conversely, what you’re seeing makes your hypothesis likely. And the Bayesians incorporate that in the likelihood. And the product of the likelihood and the prior is just the posterior, which is how much you believe the hypothesis after you’ve seen the evidence. So as you see more evidence, the probabilities evolve. And hopefully, at the end of the day, one hypothesis will come out as clearly better than the others. There’s also the marginal, which is just something that you have to divide by to make sure that the probabilities add up to 1.

A lot of great things have been done with Bayesian learning. For example, self-driving cars have Bayesian learning in their brains.

One application of Bayesian learning is medical diagnosis. And the way it works there is the following. The hypothesis, before I’ve seen evidence, is that the disease is a flu or that the disease is not flu. And the prior probability is like your prior probability of a disease being flu in the flu season: 90%, 99%, or anything else, take your pick. And then the evidence is the actual symptoms. So for example, if the symptom is “headache” that probably makes it more likely to be flu. If one of the symptoms is “weakness”, that makes it even more likely to be the flu. On the other hand, if there is “no fever”, that actually makes it less likely to be the flu. And so what the naive Bayes classifier does is it incorporates that evidence. And at the end of the day, it computes a probability that the disease is the flu or not the flu taking all of that into account. And then based on that probability, you can decide whether or not to get further medical tests done.

Some algorithms are based on this concept are Naive Bayes, Gaussian Bayes, Independence bayes etc

With all the great applications of Bayesian learning, it’s not without challenges. For one, because it’s all based on probability, bayesians are never 100% certain. Also, bayesian learning works on a single table of data, where each column represents a variable. But if we have more than one table, bayesian learning is stuck. As an example Bayesian learning cannot directly combine variables of a car, with variables of the environment, both of which are needed to be learned operate a self driving car in certain situations. So, it needs to be augmented with other machine learning techniques to solve such problems.

#5 Learning Models based on Reasoning by Analogy

The basic idea here is that everything that we do, everything that we learn is reasoning by analogy. It’s looking at similarities between the new situation that we need to make a decision in and the situations that we’re already familiar with.

So let’s see how this works with this small puzzle.
Here’s a map of two kingdoms out of the seven kingdoms of Westeros in the show The Game of Thrones. Westerlands, whose capital is Casterly rock and The Reach, whose capital is Highgarden. And we don’t know where the frontier between the two kingdoms is. We just know where the major cities are. Let’s say we know the major cities of Westerlands are represented by a plus(+) and the major cities of The Reach are represented by a minus (-). If this is all we know, how can we determine the border between the two kingdoms?

Now the nearest neighbor algorithm has a very simple answer to this question. It just says, I’m going to assume that a point on the map is in Westerlands if it’s closer to a city marked with +, than to a city marked with -. And the effect that this has is to divide the map into the neighborhood of each city. And then Westerlands is just the union of the neighborhoods of the + cities. The neighborhood of a city is the points that are closer to it than to any other city. And then as a result, you get this jagged straight line frontier.

As you can see, the nearest neighbor algorithm is very simple, with enough data the points converge very closely and it can form very complicated frontiers, converging to the best possible hypothesis.

If you look at this map closely, you could actually throw away some of the cities, like Lannisport, or Shield islands and the frontier doesn’t change at all. The only cities you need to keep are the ones defining the frontier, which are called support vectors, because they support the frontier where it is. So often, you can throw away the great majority of your data and it doesn’t change anything.

You have now have a method of classification where if you identify a new city in The Game of Thrones , it can be slotted into belonging to either the Westerlands or The Reach. This similar mechanism can be applied to a completely different application: recommender systems. Let’s say I want to figure out what movies to recommend to you. We can do that by collaborative filtering: find people who have similar tastes as you (as in they rated a movie 5* and you rated the same movie 5* ). And now if they give 5* to a movie you haven’t seen, then I’m going to hypothesize that, by analogy you are “within the same frontier” and you will also like that movie and so I will recommend it to you.

Some algorithms are based on this concept are Linear regression, Logistic regression, K-means clustering and Support Vector machines.

Conclusion

These are the 5 main types of learning models. There are hundreds of learning models and data mining techniques and most of them fall into one of these types, or a combination of these types.
Is there a problem that these learning models can solve in your area of interest? If so, which of these models do you think can help solve that problem. Please post in the comments below.