I’ve recently been inspired to read around the subject of artificial intelligence (thanks to a certain someone, who I hope won’t find anything wrong with what I’m about to write, but let me apologise now in case I mess anything up…). It’s been an incredibly fascinating journey so far. I should probably start off by saying that I really had no idea so much already existed in the AI field. It’s all pretty amazing. As a self confessed sci-fi geek, everything I knew about AI came from novels and films. This meant that I knew next to nothing about the real world implications. In a recent article I read (on my quest to rectify my ignorance), I learnt about some of the new impressive capabilities of computers.
In December 2015, a new algorithm was developed that allowed computers to ‘learn’ new concepts in a way that was (scarily) close to humans. These computers could engage in creative tasks and learn things much more quickly thanks to this algorithm.
For us humans, learning a new concept, such as a new dance move or a new letter in an unfamiliar alphabet, is relatively easy. After a few examples, we’re able to understand the concept which means that we can recognise it and use it later on. Machines, on the the other hand, need thousands of examples to be able to perform the same concept with similar accuracy as a human. Researchers from MIT, New York University, and the University of Toronto, have developed a new way in which machines could shorten this learning process. They developed a Bayesian Programme Learning (BPL) framework which allows concepts to be represented as simple computer programmes.
Take the letter ‘A’ for example. Through using the BPL framework, you could represent the letter ‘A’ in the form of a computer code which would generate examples of the ‘A’ when the programme is run. The code looks like it’s been written by an actual computer programmer, but it hasn’t! In actual fact, the algorithm programmes itself by constructing code to produce the letter it sees (in this case, an ‘A’). The code is even more impressive because it is able to produce different outputs every time, unlike a more traditional code which only produces the same output every time it’s run.
Using the BPL framework to create algorithms is very different from using the standard pattern recognition algorithms. These standard algorithms represent concepts as configurations of pixels, whereas BPL works by ‘explaining’ the input data to the algorithm so that it can ‘learn’ faster and better. In the letter ‘A’ example, BPL is designed to capture the causal and the compositional properties of writing and recognising the letter, which allows the algorithm to use data more efficiently. The algorithm also allows the machine to use knowledge from previous concepts (that it’s already learnt) to learn new concepts a lot faster. For example, if the machine had already learnt the Latin alphabet, it could use this knowledge to learn the Greek alphabet in an accelerated time frame.
In order to test the algorithm’s ability to learn and recognise new concepts, the researchers used a ‘visual Turing test’. They showed some humans and some computers a single example of a character and asked them to reproduce it, or to create new characters in the same style. They then asked human judges to take a look at the original character as well as the character produced by the person and by the computer. The judges had to identify which of the characters was created by the machine. The results showed that less than 25% of the human judges were able to correctly ascertain which characters were created by a fellow human, and which by a computer.
Even though we still have some way to go before machines are as smart as humans, the development of this algorithm is a pretty big deal. It’s the first time that a machine has been able to learn and use real-world concepts in ways that are hard to tell apart from humans. This means that machines will soon be able to learn things much more quickly and at the same standard as a human, really emphasising the ‘intelligence’ in ‘artificial intelligence’!
I think for me, personally, the science behind how these machines work is super interesting, but it’s something I don’t think I’ll ever be able to fully wrap my head around (too much complex mathematics is involved…and unfortunately I suck at maths). Instead, I much prefer thinking about the questions that have been brought to light since the reality of living with artificially intelligent life forms (e.g computers and robots) is fast approaching. Questions like: How will society change when AI becomes more mainstream? To what extent can we regard AI life forms as individuals? Do they even have feelings? Do they truly understand the significance of things they are learning?
I don’t know how long it’ll take before we start living in a world where AI life forms are a constant part of our lives, but from what I’ve learnt and seen so far, I think we’re slowly getting there. And the question of whether or not AI will enhance our lives is something we’ll probably be able to answer in the next fifty years or so. I, for one, am hoping that the answer is: “it will” (enhance our lives I mean) and not: “Shit! Robots are taking over the world!” But, we’ll see…