Why do machines learn?
Since the onset of the fourth industrial revolution, we have been largely reliant on computers and computer algorithms to take care of things for us. This includes processing of the copious amounts of data we collect and making decisions and predictions based on this data. Although humans aren’t the best at decision-making and often make the wrong decisions, we are still able to compare outcomes, reason and make choices about most of our daily challenges. Thanks to our neural networks which have been developing pretty much since birth, and which we often take for granted, we can make sense of very complex or abstract scenarios without much advance training.
For machines to be able to make sane decisions and give us sound advice, they need a significant amount of programming, learning, training, testing, retesting, tweaking, re-tweaking and so on, until they can be relied on. In a way, they are like students who are very slow learners. Although it takes a while for them to start learning, once they begin to ‘get it’ right with a high degree of reliability, they are able to process massive amounts of unlabelled data in no time – a task that’s physically impossible for human brains.
As we know, advance machine learning is the process that fuels AI (artificial intelligence) – the new technology slated to displace humans in many fields. Although most of these fears are unfounded, there is a number of reasons why machine learning isn’t just the next tech hype. The computer algorithms that fuel machine learning mostly fall into one of three main categories: supervised, unsupervised and reinforcement learning; here’s what each one can do.
As the name implies, this type of machine learning requires a certain level of supervision of machine learning models, or bots. This is done by teaching the model, or loading it with knowledge, so that it can ‘’learn’’ the desirable behaviours it should perform. Within supervised learning, there are two sub-types of models: classification (the organisation of labelled data) and regression (the prediction of trends in labelled data to determine future outcomes). Both rely on knowing what kind of data they are dealing with in a controlled environment, with regular testing guiding the learning process.
Unlike supervised learning, this type of learning is different in that it isn’t observed by anyone and happens on its own terms. We let the model, or bot, discover information that may not be visible to the human eye or brain. Unsupervised learning relies on machine learning algorithms to make conclusions and predictions on sets of unlabelled data. As such, its biggest strength lies in the ability to find patterns and groupings from unknown types of data.
When comparing the two main types of machine learning above, unsupervised learning is more difficult and less controlled as the bots have little to no information about the data they are looking at or the expected outcomes; thus, the biggest difference between the two types is in the labelling of data.
There are sophisticated learning algorithms that are neither supervised, nor unsupervised. Deep learning, also known as reinforcement learning, relies on an agent – environment setup, in which a certain environment state is being passed on to the agent. In response to the environment’s state, the agent issues actions which affect the environment, thus creating a continuous feedback loop, where rewards guide the nature of the actions. In this way, the environment is reinforced by the agent at every interaction, as it continuously tries to ‘’please’’ or improve.
The biggest advances in AI today rely on unsupervised and reinforcement learning algorithms as they are able to uncover unpredictable patterns in datasets that could be interesting for humans to exploit. Data = power in a world that is based on it.
Copywriter: Ina Danova