/ Data Science, Machine Learning

Friendly Introduction to Machine Learning

What is machine learning?

Computer programming has traditionally been a tedious task. At the very start, women were employed to program computers because the male engineers considered the task so laborious. These days are now long past. It looks like both sexes will get an easy way out of it. All because of machine learning.

Machine learning means that instead of instructing computers in detail how to recognize a dog in the photo, the computer learns to do things on its own through the process of machine learning. This process involves application of learning software that goes through a big pile of examples to discover what constitutes a dog. As an outcome, you will get a software that recognizes animals, which can be further embedded in a product or service. This process can be continuous, meaning that even after the solution has been taken into use, the machine can still improve its performance by learning from new examples as they become available.

The Basics of Machine Learning

This story covers the basics of machine learning, or just ML as many tend to call it, its definition, what is being learned, how learning happens and what kind of things computers are ready to learn from.

Starting point: What does machine learning mean?

The traditional logic of computer programming has been that computers will do the things we want them to do only the way we have told them to do. Machine learning is removing the need to explicitly instruct the computer on how to match a certain set of inputs to an out. In other words, the task of naming a person from a photo is about examining facial features and producing the name of a person as output. With machine learning, the programmer doesn’t need to give detailed instructions on how to perform this task.

In the machine learning way of thinking, the computer learns the relationship after “studying” examples, pretty much as humans do. However, for the future, machine learning shows the possibility of outperforming humans and even learning without supervision, something in which humans are not that great in.

For practical purposes, it suffices to think about the product of machine learning as a black box within the system that allows the magic to happen with minimal intervention from software developers. Although the box is described as black, it is not fully opaque, meaning that the data scientist or other machine learning specialist can peer inside and tweak its mechanisms as needed.

What do machines learn?

Everything machine learning does comes down to two outcomes: regression and classification. They refer to a resulting capability of the system to respond either in real numbers (0…1; regression) or in categories (rabbit, lion, food; classification). See the illustration below:

Screen Shot 2016-11-01 at 10.30.12

While this should make things simple, it actually hides complexity related to the what can be achieved with this approach because applications can be built upon several such elements.

What do machines learn from?

Learning happens using training data. Usually quite a lot of data is needed to help the machine to learn its regression or classification task. What is a lot depends on the challenge, but can range from tens to tens of millions examples. The latter clearly exceeds what humans can typically encounter in a limited amount of time.

The learning material must naturally be in digital format. Luckily, the technological sensoring capacities already far surpass human senses in many domains and there is usually lot of data available. This helps computers to gain superpowers through learning when they have a different set of senses available than humans and considerable resources to process it. For instance, humans are really bad in detecting carbon monoxide before the level gets so bad it is intoxicating because of sensing, or sorting out 1000 spam emails in a fraction of a second because of processing.

Inputs for machine learning can be anything you can translate to digital format, including

  • Vision, images
  • Audition, sound
  • Smells
  • Vibrations
  • Cold, hot and pressure
  • Acceleration
  • Relative position of a body

And computers can utilize “super-human” senses:

  • X-ray vision (MRI, CT, etc.)
  • Infra and ultra sound
  • Radio frequency information (MHz, GHz, etc.)
  • Air quality sensors (temperature, humidity, O2, O3, CO, CO2, etc)
  • Radar/sonar

The only trick is that when we humans don’t sense these signals directly. Therefore we must first translate them to a domain we can read or hear to get an intuition is there something to be learned or not. Then we can try our own wits in learning about the possible regularities of the data.

How do machines learn?

Machine learning is dependent on the use different computer programs for learning, called learning algorithms. Sometimes people talk about neural networks, deep learning or other detailed techniques. The implementation, selection and calibration of these algorithms is the secret of machine learning. In this article, I don’t attempt to dive quite so far. Instead, let us consider the basic methods of learning.

Machines learn in three ways: through supervised, unsupervised and reinforcement learning. In supervised learning, the machine learns from examples that also include the desired outcomes, e.g. images are categorized into cats and dogs. In the case of unsupervised learning only images are needed and computers is tasked in grouping into what humans might call “4-legged animals”, “2-legged animals”, “legless animals” if legs happened to be a prominent factor in the training images.  Reinforced learning has the special feature of being able to work even without initial learning material as long as some feedback mechanism exists that can gather data while the system is running.

Getting complicated?

These are the building blocks for machine learning. Complex applications, such as personal assistant Siri residing in your iPhone can utilize several input approaches and different learning algorithms. The approaches that combine different algorithms are called ensemble methods. They have turned out to be the most effective, even if complicated to implement. For instance, the winner of the Netflix Prize in the highly regarded competition in 2009 was an ensemble solution.

A word about artificial intelligence, or AI. Is it one application of machine learning? Given that by 2016, computer have beaten humans in the game of Go and passed the Turing test, we can say that machine learning has definitely contributed towards building systems that display the powers of AI. What it will take to achieve singularity, no one knows, and few believe it is even possible.

Now that you’ve gotten the idea of what goes into a machine learning solution, you can lean back and wait for my next post that will outline different type of applications that machine learning can produce!

Confused?

SC5 specializes in state of the art digital services. We are also capable of helping our customers to benefit from machine learning in the process creating more value. Contact us if you want to discover more about the opportunities of machine learning for your business!

Text: Lassi A. Liikkanen, Data-Driven Design Specialist at SC5, @lassial

Illustrations: Eija Jokilahti, SC5

Thanks to Max Pagels, SC5 for commenting.