Original article was published on Artificial Intelligence on Medium
Reinforcement learning is defined by an “agent” that can choose from several actions or outputs. This agent is given feedback as it makes choices and is rewarded for making “good” choices (by some definition that we specify in advance). The agent is initially catastrophically bad at its job, but will quickly get better as it collects more and more feedback on its actions. Reinforcement learning can be used to train manufacturing robots — for example, a mechanical arm on a production line. The robot arm is rewarded for performing the desired action — whether it’s moving a component from one place to another, or building a certain car part — and punished for irrelevant or negative actions (eg dropping things, making a mess, or becoming self-aware and attempting to destroy all humans. Bad robot).
What’s all this I hear about Deep Learning?
Deep learning is a subfield of machine learning that concerns itself with models made up of so-called “artificial neural networks”. These are made up of layered algorithms that can progressively pick up on more and more sophisticated features within input data. For example, a low-level and unsophisticated feature of an image of a human face would be where the edges of the face are. A more sophisticated feature could be whether the person looks happy or sad — which isn’t as straightforward. These algorithms are made up of layers of “neurons” that are loosely analogous to those of the human brain, because given a certain input, some neurons in the model will activate and others won’t. The neurons then work together to come up with a collective decision about what the model should output.
Deep learning systems can be extremely complex, but they can also be extremely powerful. These architectures underpin Facebook’s automatic photo tagging feature, Google Translate, and even algorithms that can beat world-class players at games like Go — which only a few years ago was a feat thought to be out of the reach of our current technology.
And what on Earth is NLP???
Natural language processing, abbreviated to NLP, is another subfield of machine learning that aims to design computer models capable of understanding human language and the intent behind it. NLP models may be designed and trained to simplify or summarise documents, to analyse the tone or sentiment of written text, to recognise and translate speech, or to predict the next word you want to type out on your phone keyboard. NLP is the collective name for the range of machine learning problems such as these, which may be supervised or unsupervised in nature.
This has been a very quick introduction to some high-level machine learning concepts. It may still feel like an incomprehensibly vast field of study — and that’s because it is! Like many other disciplines, people dedicate their lives to extremely narrow areas of research within it and have only found more and more depth to explore.
But hopefully, the next time you open up an article about a recent development in machine learning, you will be able to recognise some of the key themes this time. And then maybe you can pick one or two more topics or terms that you don’t quite understand and do some research to try and get a grip on them. You may end up with more questions than answers — but that’s not necessarily a bad thing. If you can do that, then at least you will know what you don’t know.
Socrates would be proud.
*Of course, there’s no record that Socrates himself actually said this.
**Or that he liked pumpkin spice lattes, for that matter.