How Dogs Can Best Explain Machine Learning: An Executive Introduction
The Best Practice AI 3:10 approach to machine learning guides you from your initial goal to the required machine learning task
Machine learning is often a case of acronym soup. Terms like reinforcement learning or CNNs or back propagation or hyperparameter tuning make it feel impenetrable at times. We wanted to provide a more intuitive framework to introduce a non-technical audience to machine learning. And how to think about machine learning in the context of your business needs.
When we surveyed the 580 use cases and 800+ case studies in the BestPractice.AI library we wanted to build a consistent approach to tagging the sorts of artificial intelligence tools or technology that might be deployed. A significant number are what we chose to call ‘traditional artificial intelligence’, the heuristics or rule-based approach that drove the field prior to the current era of data and power hungry machine learning models.
While there are many different ways to categorise what machine learning (ML) does, at Best Practice AI we’ve distilled these into 10 primary tasks that machine learning can be applied to. Rather than be a definitive list, this roadmap is intended to guide thinking about the ways in which machine learning can be applied to solve practical problems, and help make sense of what is often just a set of buzzwords.
Before an appropriate AI solution can be implemented for a use case, it is important to identify what type of machine learning task you’ll need to implement. In our roadmap three initial paths lead to 10 machine learning tasks.
To begin we need to answer the question: “What is the ultimate goal?”
While it seems an obvious starting point, we need to ask what it is we want our machine learning model to actually do. Once you’ve identified your business use case, It’s important to recognise that a single business use case may have several different machine learning approaches which can be applied to it. Looking at the fundamental different machine learning tasks can help understand why one approach may be better suited than others, or how a combination of different techniques will be needed to help you arrive at your goal.
We’ve identified three ways to answer this initial question, each of which will send you down a different path:
Action Selection — choosing which action(s) the model should take to maximise a specified outcome
Grouping — deciding which inputs (data) fit together
Prediction — predicting an output that best corresponds to a given input (data)
Broadly speaking, these are the three main goals of machine learning. Do you want your model to act as an agent, choosing which action to take based on a feedback loop of data from its environment? Do you want your model to group your data inputs together based on some commonalities it discovers? Do you want your model to predict an output for your data inputs?
Being clear on what you’d like to achieve with machine learning is the first step. It can often be the hardest as it forces you to think through what sort of problem that you need to solve. But time spent on this early can stop you from barking up the wrong tree — and indeed the failure of a project.
The Ten Tasks
The ten tasks follow from however you’ve answered our initial question. The Action Selection and Grouping paths are relatively straightforward, whereas Prediction requires you to answer a few more questions.
The objective of action selection is having an agent choose what to do based on a reward to maximise or outcome to achieve, and then do it, receiving new information from its environment as a result. The field of machine learning concerned with this is called reinforcement learning (RL).
Reinforcement learning — determine the action(s) that will lead to maximizing a certain objective in the long run, based on a feedback loop. For instance, determine the feeding schedule that optimises a dog’s health indicators.
Grouping ML tasks do exactly that — they organise and arrange data according to some relationship between data points that the algorithm has detected. It’s important to note that you have no direct control over how the model chooses to group your data.
Clustering — grouping together data points with some identified similarities. These may or may not end up matching existing human concepts — for instance, clustering a collection of dog photos into groups of visually similar images may or may not result in the clusters corresponding to canine species.
Anomaly Detection — detecting outlying data points. For instance, a photo of a cat in a dataset of puppy photographs.
Dimensionality reduction — grouping data points closer together by moving them to a smaller dimensional space, simplifying the problem. For instance, mapping photos of animals unto a 2D coordinate system which would allow a quick visualisation of what animals are similar to each other.
Generally, these ML tasks predict the labels for new data that’s been inputted into the model. These labels are the model outputs, and we can classify the type of task based on what we want our label output options to be.
Generation — The model creates new data as its output. This data can be audio, images, text, etc. While the training dataset is provided with examples, the output is not expected to replicate exactly the input examples. It may also be non-deterministic, meaning that given the same input and algorithm, the output generated may be different each time the process is run. For example, new images of dogs playing.
Annotation — Makes separate predictions for different parts of the input data. For example, identifying what dog breeds are in each section of a dog park.
Binary classification — Classification makes a prediction about what label to give something, or to put it another way, what category of thing it is. Binary classification puts things in one of two categories, such as “yes” or “no”. For example, whether there is a dog in a photo or not.
Multi-task classification — Like binary classification, except multi-task classification predicts which of several different types of things a data point belongs to. For example, what breed a dog in a photo belongs to.
Regression — Predicts a numerical outcome for a data point. For example, the age of a dog based on a photograph.
Data transformation / translation — changing data from one format to another, for instance from text to speech. For example, automatically generating movie subtitles for Isle of Dogs.
It’s also important to understand that these ML tasks are not mutually exclusive — not only will there be more than one way to get to a workable solution, but it’s likely that multiple approaches will be needed to get to an optimal solution. Testing, refining and combining approaches is a key part of the data science role. However, being able to frame your problems in terms of machine learning tasks at a broad level will help you understand how your in-house or external team approaches implementing machine learning for your business use cases.
Learning Types
You may have heard of “supervised”, “unsupervised”, “semi-supervised”, self-supervised” or “reinforcement” learning. What they refer to is how a task is accomplished by your ML model. We’ll briefly discuss the relationships between data input and the solution output here, and what type of learning they refer to.
Supervised — In supervised learning the ML model is provided with a set of labelled data to train on. The model then makes a prediction about what the label should be for new data it’s given. What do we mean by labelled? If you have an image dataset of dogs and cats, each image needs to be labelled (correctly) as a “dog” or a “cat”. This labelling is manual work done by humans and can be very time intensive.
Self-supervised — This is similar to supervised learning, except the labels do not need to be provided by humans, as they can be found in the original dataset itself. This could be the case if we wanted to predict stock price values for Dog Food Inc. and had a dataset of historical values. We could create a new data point for each day by using the price values up to that day as an input and the price on that day as the label.
Semi-supervised — In semi-supervised learning we have a combination of labelled data (as we do in supervised learning) and unlabelled data (as we do in unsupervised learning). Semi-supervised algorithms can combine these two types of data in order to make better predictions for new data points than could be achieved using the labelled data alone. The unlabelled data could prevent your model from overfitting or allow you to capture the fundamental structure about your data points.
Unsupervised — Unsupervised learning does not require any labels at all. Instead, it discovers patterns and commonalities in the unlabeled data set, and in the process can produce its own labels. Unsupervised learning may discover similarities among the data a human wouldn’t necessarily grasp, but there’s also no control over the output.
Reinforcement — Reinforcement learning is different than the other learning types here, as it doesn’t require datasets, but instead an environment that the model (called an ‘agent’ in this context) can interact with and get information from. Because it differs fundamentally in this way from other types of machine learning, it is also one of our 10 ML tasks.
The Acronym Soup — Data, Models, and Algorithms
In order to accomplish machine learning tasks, computer scientists and data analysts use a combination of machine learning models and algorithms. These models are constructed in different ways, and many different factors can influence which “architecture” is chosen. These include the generally accepted best practices at the time, weighing various trade-offs such as model stability, consistency, data and computer power requirements, as well as the machine learning practitioners’ personal preferences.
Some currently much discussed and implemented model architectures include artificial neural networks (ANNs) and their variations: convoluted neural networks (CNNs), recursive neural networks (RNNs), generative adversarial networks (GANs); decision trees, which can be combined in an ensemble as random forests; and support vector machines (SVMs).
In addition to the mathematical bones of problem-solving, you’ll need data — and in certain cases, the tools for capturing that data — in order to build [flesh out] your machine learning tool. We’ll take a closer look at data and model considerations in future posts.
Applying the 3:10 Model in the Real World
To help illustrate how these machine learning tasks apply to the real world, we’ve thrown you a bone and tagged the use cases in the BestPractice.AI library with AI type. We’ve also indicated where ‘traditional AI’ is likely to be used right now in addition to the machine learning techniques covered here. As machine learning researchers continue to discover new applications and ways of implementing the science this is by no means definitive — consider it a useful starting point for identifying the usual suspects. Hopefully our roadmap will also guide and encourage you to think about the nature of your project, its data considerations, and how machine learning may — or may not — be the best tool for you to use.
In the following posts of our roadmap series, we’ll take a more in-depth look at how specific use cases map to these machine learning tasks. In the meantime, check out the BestPractice.AI library for examples of the types of problems being solved with these tasks in your industry.
And please, do get in touch with any feedback on either our approach or the data library at BestPractice.AI.
Sam Work
