Artificial Intelligence Today

Artificial Intelligence has been around since the beginning of computing and was founded as an academic discipline in 1956. Ever since then, the scientific community has thought that they are close to a real breakthrough every decade or so. Already in the 40’s research on neurology found out that the human brain is a network of neurons that fires in all or nothing pulses, and Alan Turing suggested that it might be possible to build an electronic brain. In 1951, a 24 years old graduate student, Marvin Minsky built the first neural net machine, also known as SNARC, which is referred to as the first artificial self-learning machine.

But if they already had the right ideas back then, why has AI got so much traction only in recent years? It turns out that programs like this require incredible amounts of computing power and early logic relies on traversing trees, where the complexity increases exponentially. In mathematics this is called a combinatorial explosion, which is described as a function that grows very rapidly as a result of combinatorial considerations. Therefore a lot of effort has been put in heuristics, finding techniques for solving problems more quickly by trading completeness, accuracy or precision for speed. Heuristics help to reduce the number of alternatives from an exponential number to a polynomial number. In a general sense, the term heuristic is used for any advice that is often effective, but is not guaranteed to work in every case.

The Dangers of Artificial Intelligence

There has been quite a lot media attention on the advances of AI in recent years, in particular for unarmed drones and weapons technology. And rightfully so, the US military is funding more research than ever on autonomous and self-aware robots. Some of our heavy weight leaders in the scientific community, the physicist Stephen Hawking, the entrepreneur Elon Musk and the multi billionaire, philanthropist Bill Gates in particular, recently issued an open letter about the dangers of AI if we don’t take extreme caution. They agree that success in creating AI would be the biggest event in human history, but unfortunately possibly the last unless we learn how to avoid the risks.

So how afraid should we be? I think that there is some substance to their concerns, but in reality the vast majority of AI:s you and me will encounter are referred to as weak AI:s. For instance, a real time GPS Map application that routes your pathway across traffic to find the most efficient way from point A to point B. The routing software can get really good at that, thousands of times better and faster than a human actually, but would you say that it was thinking? No, and that’s the rub!

They warn us about what they call Strong Artificial Intelligence. AI that is capable of thinking by itself, perhaps with a consciousness and ability to make decisions. Some scientists think that in 2025 we will have an AI that is equally as intelligent as a human being and by 2040 one single AI instance will be as intelligent as the entire human race collectively. The AI problem can be divided into 2 sub problems, the first one is to figure out how to create intelligence, the second one is how to give the AI safe constraints or the same values as humans. I say that if we figure out problem 2 first, we’re safe. But what happens is just speculation, but I agree that the unknown can be daunting.

“Massive transformation is on the horizon. Jobs, companies, and entire industries will be reimagined.”

• Paul Roetzer, CEO @PR 20/20

Artificial Intelligence Today

The Silicon Valley giants like Google and Facebook are currently riding on top of the AI wave of popularity. Google bought the secretive AI company DeepMind for 400 million dollars and are currently experimenting with the tech. Facebook uses AI to recognize peoples’ faces in pictures.¹ Then we all know about Google’s fully automated self-driving cars capable of driving a million miles without an accident. Your Android device knows when you’re on your way to work and informs you if there’s a traffic jam on the way or if your bus is late. If I search the net for a weekend trip to London both Facebook and the rest of the net gets filled with travelling ads. AI technology is literally everywhere and I bet that tomorrows fortune 500 companies follow the formula:

Take X, add AI.

But that is only the beginning. All the major cloud companies, plus dozens of startups, are in a mad rush to launch a Watson-like cognitive services, speech recognition software and basically anything AI related. According to quantitative analysis firm Quid, AI has attracted more than $17 billion in investments since 2009. In the year 2014 alone more than $2 billion was invested in 322 companies with AI-like technology.²

Facebook and Google have recruited researchers to join their in-house AI research teams. Yahoo, Intel, Dropbox, LinkedIn, Pinterest, and Twitter have all purchased AI companies since last year. Private investment in the AI sector has been expanding 62% a year on average for the past four years, a rate that is expected to continue.

So what does all this mean for the rest of us? Do we sit here and wait until the Silicon Valley companies develop their technology and get all the market share for themselves? Well, to be honest the answer is both yes and no and frankly depends on what you want to do. The thing is that creating a truly intelligent AI requires incredible amounts of resources and data, and naturally an AI gets smarter the more data it gets which puts certain companies at a competitive advantage. Literally everyone is using Google’s services so they have arguably the best data source for their research. So in the end I think there will be just 2-3 companies with competing AI’s. But in the same way, the internet is a result of the cold war, we get a lot of technology from this research as well. We will be able to use their AI services, and we can implement our own frameworks with components of AI, like Machine Learning or Deep Neural Networks.

Deep Neural Networks

“Neural networks and deep learning currently provide the best solutions to many problems in image recognition, speech recognition, and natural language processing. - Michael Nielsen³”

Artificial neural networks are trained by showing it a great amount of training examples and gradually adjusting the network parameters until it gives satisfactory classifications. A typical network is constructed using 10-30 stacked layers of artificial neurons. When for instance feeding images to a Neural Network the input layer does some low level processing and then outputs the result to the next layer, until the final layer is reached which then determines the networks interpretation of the image.

“One of the challenges of neural networks is understanding what exactly goes on at each layer. We know that after training, each layer progressively extracts higher and higher-level features of the image, until the final layer essentially makes a decision on what the image shows. For example, the first layer maybe looks for edges or corners. Intermediate layers interpret the basic features to look for overall shapes or components, like a door or a leaf. The final few layers assemble those into complete interpretations, these neurons activate in response to very complex things such as entire buildings or trees. - Alexander Mordvintsev”⁴

Machine Learning

Deep neural networks is a subfield of Deep Learning which in turn is a subfield of Machine Learning. Which is a subfield of Artificial Intelligence. Machine Learning in general is the most utilized and useful aspect of AI at the moment. By definition, Machine learning explores the study and construction of algorithms that can learn from and make predictions on data.⁵ Machine Learning algorithms make data-driven predictions or decisions based on a model of example inputs instead of following conventional strict logic, typical in programming. These tasks are typically classified into three broad categories, depending on the nature of the learning input, these are:

Supervised Learning: The learning algorithm is fed with desirable labeled inputs and their desired outputs, in order to teach general rules which map inputs to outputs.

Unsupervised Learning: The learning algorithm is fed with unlabeled data letting it make its own independent interpretations. However, you can feed it with grouped data in order to help it reach meaningful conclusions.

Reinforcement Learning: Inspired from how humans learn, a learning algorithm is given a goal and by interacting with a dynamic environment, without any input about if it has come close to its goal or not, it also learns about the consequences of its actions.

The thing is that Machine Learning does not have to be that complicated. A really simple but surprisingly efficient example is by using Linear Regression. The purpose of this article is not to teach specific implementations, but in short you take a bunch of data points and try to find a linear approximation which fits the data best. You can read more about Linear Regression on Wikipedia. My point is that as long as we are creative we can find use cases for AI in basically any sort of application we are developing. For instance when looking for deviant behavior in fraud detection or credit risks, or dynamic pricing and product recommendations in eCommerce applications, or just take a look at your Spotify account and see how they recommend music based on what you are listening to. They actually analyze the music itself with deep neural networks and are able to recommend music based on how it sounds, so it is based on a lot more than just metadata!

”It will be more about the sound. You’ll be able to search by the content of the music instead of just text information associated with it.” – Nicola Montecchio, head of Deep Learning @Spotify

Conclusion

Personally I think that a lot of these concepts are easy enough to understand and for us to stay relevant as developers we need to start investing in these skill sets. We are moving more and more towards a Big Data world and effective analysis techniques are essential in order to stay competitive, and imagine what user experiences we can create in our applications using this technology!

I will give a real world example of something I have done. I worked for one of Sweden’s biggest datacom and telecom companies, their billing and customer data was in an economy system and their associated costs for their broadband network was in another specialized system. However, they had no connection between each customer and the associated cost for the connection and to make matters worse, there were tens of thousands of entities to be matched, so it could not be done manually. No conventional script could do it either because most of the data was written by hand and therefore full of spelling errors and no specific patterns to match since they adjusted the data based on customer billing needs. So me and a former colleague made an AI that made an educated guess for each entity. The program could match about 90% in one go and the remaining 10% was associated with some sort of error, maybe a resource was unmatchable because there was no customer utilizing it!

So in the end we used AI to solve a real business problem, a problem which had no simple solution and would have been extremely hard to solve using a conventional algorithm. We simply need AI as a tool in our toolbox and the businesses that realize it first are going to have a huge edge amongst their competition.