What is Machine Learning?
Machine learning (ML) is a component of Artificial Intelligence that is centred on the use of data and algorithms to mimic the way humans learn. Without ML, AI would not be able to do its job. If we break down ML it consists of three categories:
Supervised learning is essentially input = output. It is a very common method of machine learning to develop AI models. The reason why it is referred to as supervised learning is that the machine is given an exact specification on what the inputs are and an exact specification of what the outputs will be. The below example is an explanation of a use case to help you gain an understanding on how the whole process works. With ML it is likely to take the below process:
- The data input
This is the set of data (many people call them datasets) that the AI software will be working with once it has been processed or organised by ML. The data must have a clear structure and be organised (in many cases labelled) for ML to be effective and the data input in many cases (not all) can be the biggest challenge for the utilisation of AI.
An example of a data input would be a dataset of images for an image recognition.
A set of algorithms would then be written and utilised to process the data. What is important in this phase of machine learning is ensuring that the data is clean, structured and a very clear idea of the output or result.
So, to continue with this example, if the dataset of images was collected and the AI model needed to quickly identify if a picture was in black and white vs. in colour; the dataset would need to contain images labelled black and white or colour.
- The output
With the combination of the input plus the algorithms the output would be able to accurately describe whether a picture is in black and white or in colour. One important thing to note at this point is that the more data an ML model is fed, the more accurate it becomes.
Many people in our industry would call this training the dataset i.e. the model is being trained to be able to output something correctly if it is given a new piece of data. In this case, the new data would be a picture that the model has not “seen” before.
So, with this example in mind, the data being trained was comparable to a student being supervised by a teacher to ensure that they are correct in outputting. A lot of processes would build in a feedback loop where a supervisor is verifying whether or not an output is correct.
Unsupervised learning on the other hand is an ML model being presented with an input only and no guidance on what the output will be. The easiest way to explain this would be that the model is trying to find patterns or common features to output or present.
A great example of unsupervised learning is Netflix’s recommendation engine, the unsupervised model learns what to recommend to you overtime based on what you have watched in the past.
Reinforcement Learning is an element of machine learning where the model is presented with data and it needs to figure out how to traverse through that data by itself without any guidance. The “Reinforcement” of that learning is by building in signals of positive (rewarding) or negative (penalising) feedback based on the learning outcome of the model.
This means that if the AI produces results that are useful it gets rewarded, if the results are not so useful; it gets punished.
A great example of reinforcement learning is how social media platforms work. Social Media providers are always figuring out what content to show you so that you keep using their platform. Positive interactions such as liking, sharing, or saving would be regarded as reward signals for the AI model. Negative interactions such as hiding, reporting, or unfollowing would be regarded as punishment signals for the AI. The more you interact with a certain piece of content coming from a certain account; it is going to show similar pieces of content in the future.
If you have any questions about AI we would love to hear from you. Contact us by e-mailing email@example.com.