What is Machine Learning? We often explain it with a chart that looks like this. Here is what is missing from this table.
Machine learning is positioned such that all machine learning is artificial intelligence, while not all machine learning is deep learning. It’s great for giving an abstract idea of how these terms are all related, but doesn’t explain well what each category is.
That was the biggest question I asked myself when I started my research on artificial intelligence: what do these categories look like in the real world? The first step is to find out what this classification system means.
How is artificial intelligence classified
When we classify things as a society, we tend to define groups well, because often classification is a way of organizing and sorting a collection of things. We sort our fruits from our vegetables, our shirts from our jackets, and very honestly everything that surrounds us.
Like everything else, artificial intelligence is also categorized. Specifically, we tend to classify artificial intelligence in three main ways. Here are some of the most commonly used classifications:
- Artificial intelligence
- machine learning
- deep learning
All three are mainly used to identify the type of intelligence and learning capabilities of a machine (this is a generalization – these classifications also include other smaller aspects)
Intellectual capacity (compared to humans)
- Narrow artificial intelligence — Less “intelligent” than a human being
- General artificial intelligence — An “intelligence” equal to that of a human being
- Super Artificial Intelligence – “Intelligence” significantly greater than a human being
- Responsive — Responds to real-time stimuli
- Limited memory — Uses previously learned information
- Theory of Mind – Understands thoughts and emotions
- Self-awareness — Is self-aware
All of these types of classifications can be used to describe different types of artificial intelligence, but as you may have noticed by now, the first category is the most relevant when it comes to AI (why this All the machines we’ve ever made fit into ‘narrow AI’ for the second category and ‘responsive’ and ‘limited memory’ for the third, so these classifications are less descriptive and useful).
Artificial Intelligence – The “Umbrella”
Artificial intelligence is a label for any machine capable of displaying “intelligent” or human-like behavior. While you might imagine this referring to a machine’s ability to have human thoughts or emotions, AI can be much less complex.
Take, for example, a website search tool. This search tool, implementing exclusively AI, would be extremely primitive, only being able to return results based on exact rules created when programming the tool.
Here’s what it might look like:
Search query: “How to create a table”
- Result 1: “How to build a table”
- Result 2: “My learning experience of building a table”
Undesirable information: This search query retrieved what the user was most likely looking for; instructions for building a table. You may have noticed, however, that this query returns unwanted information about someone’s personal table building experience.
Missing information: Besides including unnecessary information, another problem arises if someone publishes an article with the title “How to build and assemble a table”. This article, because of the “and assemble”, is not recognized as a search result for the user’s query, which means that he does not see this article. To solve this problem, the programmer would have to add more rules to the engine, constantly struggling to maintain the balance between too many false positives (unwanted information returned) and false negatives (relevant information hidden).
This prototype struggles to provide all the good results while hiding the wrong ones (while still being reasonable to create). Keep this prototype, and let’s come back and improve it with machine learning.
Machine learning — (mostly) dependent learning
Machine learning is when artificial intelligence can refine its future decisions using specific algorithms. With the help of a human to label and provide feedback on its performance, machine learning can be used to automate various tasks.
Most of the time, a machine learning algorithm will take a set of inputs and can output as classification or regression. In a classification model, the algorithm would use data from an input to classify it. Examples of classification could include determining if a light is on in an image or determining if it is going to rain on a given day.
However, using classification is not always the best solution. For example, instead of being able to return if the algorithm thinks it’s going to rain, wouldn’t it be more useful if it could give a percentage chance of rain? This is where regression can be more useful. A regression algorithm, such as linear regression, could be used to take, say, any day’s air humidity and turn it into an output of the percentage chance of it raining based on of this data.
Adding Machine Learning to our research tool can help put an end to the problems mentioned above by retrieving unwanted information and missing important information. Here are some examples of how we might go about implementing this:
One way for a search to return relevant information is to use a tag system. If each article in a database is associated with the relevant tags (eg “How to create a table” – construction, DIY, instruction), then if the user enters these tags, the tool can return the information to a higher level of relevance. .
Unfortunately, it would take potentially hundreds of tags for a detailed search to be possible, and it would be a lot to ask every user of such a tool to remember them. This is where machine learning can be implemented. Machine learning could be used to take an input string and extract tags from it.
If all of this is possible, why can’t the machine learning algorithm just do a search and decide for each article if it’s relevant?
Well, often in AI, the easier the task for a human, the harder it is for the machine, and that’s partly true in this context. There is no simple algorithm that could be used to decide if a given search is “relevant”. It would still be possible to try, but with such complex input and output, the results wouldn’t be stellar. This is where deep learning is most useful.
Deep learning – (mostly) independent learning
One of the main disadvantages of machine learning is that it often still requires a lot of human work. Data often needs to be categorized or labeled before being fed into an algorithm and must also have an “answer key”. Deep learning can often be used to remove the need to label data so that the focus can be on feeding the machine more and more data.
As stated above, unlike machine learning algorithms, deep learning continues to improve as it receives more and more data.
Deep learning is able to achieve this level of intelligence often through the use of neural networks. Neural networks, more formally artificial neural networks, were created to mimic the way the human brain makes decisions. Most standard neural networks employ the use of interconnected neurons to translate a given input into an output.
Here are some of the most relevant neural networks today:
- Convolutional Neural Networks (CNN) — Most commonly used with images. These neural networks specialize in identifying characteristics of objects, which leads to object identification and classification.
- Recurrent Neural Networks (RNN) — Recurrent Neural Networks use previous information in addition to new data to determine future data. These are specialist areas such as natural language processing (NLP) and text translation.
- Generative adversarial networks (GANs) — These neural networks are made up of two parts, the generator and the discriminator. They work in constant opposition to each other, which leads them to improve each other. The generator’s job is to generate something (like fake video, audio, etc.) which is mixed with real data for the discriminator to separate.