I know you’re excited about machine learning, but you may wonder is it the time yet, can I use machine learning? Is machine learning suitable for your application?
Did you know that without a single hour invested in development, you can approximate the usefulness of machine learning techniques in your product or digital service by answering two simple questions. Read and learn!
So, what’s your problem?
To exploit the capability of machine learning, you must grasp the nature of a task as a computer might see it. This is what you might call the core problem of learning. This refers to defining what exactly we would like the computer to learn in order for it to complete the task we have assigned to it. These goals are not always evident at the practical, holistic level of a finished product (say, Tesla’s autopilot function).
Once the core problem of learning is defined well, then it is possible to say whether the machine learning (ML) computers of today can solve it with adequate accuracy and in a decent amount of time. This is what matters most for actually making the ML application work.
Can a machine help solving that question?
Say you are now convinced that your users and customers would benefit from a ML-boosted service. Next, you must consider the business and technology perspectives. The first question to ask is if ML works at least as well as you want, what value would it add to your product? Be honest: traditional business logic applies here, too.
If machine learning doesn’t create value, then it is probably a waste of resources. Marketing people might fancy your new implementation, but the business folks will not necessarily fund your next machine learning experiment unless you have a business case for why machine learning would improve the customer experience or revenue.
OK, suppose you’ve passed the viability check.
I expect you also have at least a hypothesis of the core learning challenge. This might be, for example, “can a computer learn to predict when the user would need to take an umbrella or a waterproof jacket when they leave the house in the morning?”.
Next, you’ll need to figure out whether you can expect the machine to learn the job you wish to get done. We’ve come to matters of data. Data broadly refers to any information you can feed the computer: weather data, shopping data for umbrellas and waterproof gear, social media updates and so forth.
Here are two simple diagnostic questions to assess the data situation that can get you pretty far:
- Do we have enough examples for the machine to learn from?
- Are there patterns in the teaching material that can be recognized by a human expert?
The first question looks easy but is difficult to answer in advance. You need both good quality and a sufficient quantity of data. Some signals are noisy (that is, have poor quality). You might need more examples in these cases than in others. Some features are very prominent and easy to learn, such as in the case of detecting nighttime or daytime from digital photos. This can be solved with as few as 30 training images!
Typical applications require thousands of instances of input data and possibly the desired answers. The more complicated the task, the more material will be needed. AlphaGo learned to master the game after analyzing a staggering 30 million games and then playing some against itself! This is likely excessive in most situations, but it gives you an idea of the scale of what computers can and may need to swallow — from 30 to 30 million. This is one of the reasons why increasing computing capacity and certain hardware really makes computers smarter by speeding up the learning of truly big data.
The second question is, are there patterns in the teaching material that can be recognized by a human expert? If a human can do the task, then there’s a fair chance that there are regularities in the data that might be picked up by the computer as well.
If you have a positive answer to least one of the questions, you can go forward with some confidence that machine learning might indeed be of use to you. What you’d do after this discovery is another story, which you can work out, for instance, with SC5.
Thanks to Max Pagels, SC5 for the original ideas underlying this article!
Human-centered machine learning by Google Design
Tips for deploying machine intelligence by Max Pagels
Text: Lassi A Liikkanen, SC5
Featured illustration: Laura Rantonen, SC5
This text is derived from an article originally appearing in Smashing Magazine: