Machine Learning

Machine learning

Quick definition: Machine learning is a subset of artificial intelligence involving models that improve in performance when given more data.

Key takeaways:

The following information was provided during an interview with Alisha Marfatia, product manager in charge of operations and strategy for Adobe Sensei.

What is machine learning?
What are the main methods of machine learning?
What is the difference between AI and machine learning?
What is the difference between machine learning and deep learning?
What are the benefits of using machine learning systems?
What are the drawbacks to machine learning algorithms?
What is the best use for machine learning?
What scenarios would make using machine learning systems unnecessary?
How has machine learning become more democratized?
How will machine learning look in the future?

What is machine learning?

Machine learning is an application of artificial intelligence (AI), a system that allows computers to perform tasks better and faster than humans. Machine learning is a subset of AI involving models that improve performance when they’re given more data.

What are the main methods of machine learning?

There are three key areas of machine learning:

Supervised learning

Supervised learning is a type of data science that uses labeled data that’s been tagged with specific information about outcomes associated with that data.

A model is then trained to learn what features or variables are predicting the outcomes assigned to the labeled input data. The model can then use the information from the output data to assess its own performance and predict outcomes.

Supervised learning involves two main use cases: classification and linear regression. Classification predicts a class label.

For example, you might predict whether a customer will cut ties with a brand based on attributes such as purchase behavior.

Linear regression predicts a numerical label, like the expected revenue you think you’ll receive from a customer based on specific attributes. The outcome is numerical variable, as opposed to a condition.

Unsupervised learning

Unsupervised learning starts with a set of raw, unlabeled data. The main purpose of unsupervised learning is to find connections between the data set and any additional data points you give the model.

This method can help you find relationship-based groups within data, or clusters, which can be used to create customer segments.

Reinforced learning

Reinforcement learning starts by inputting a set of raw, unlabeled data into a model. The model then takes action.

Based on that action, the model receives feedback on whether it acted correctly or incorrectly and the outcomes of that action. The model then creates another action and keeps learning until model optimization is achieved.

A great real-world example of reinforcement learning is a recommendation algorithm on a movie streaming service like Netflix.

The service shows you a movie you may or not like and learns from your “like” or “dislike” rating whether it should keep recommending the same types of movies.

What is the difference between AI and machine learning?

The main difference between AI and machine learning relates to rules and specifications. Let’s say you design a program that takes American customers to a certain ad if they click a link.

But you’ve also made sure that Asian customers who click the same link are sent to a different ad in their native language. Because you explicitly defined those rules for your program, you are using artificial intelligence, but not machine learning.

On the other hand, let’s say your company has several data attributes about what your users have clicked on, where they came from, who they are, and how they’ve behaved.

You could feed this data into a model, which would then create specific roles for each user persona. The model would reflect your company’s historical data of your site users, along with information about their behavior.

This model can already give you a lot of information. But if you continue to input data, it will become more accurate, until it can predict whether or not a specific type of user with a certain background will click on a certain link.

This system is not rules-based like the AI system in the first example and is a prime example of machine learning.

What is the difference between machine learning and deep learning?

Just as machine learning is a subset of AI, deep learning is a subset of machine learning. It’s a series of machine learning techniques that are loosely modeled on how the neurons in our brain communicate.

Deep learning is a major breakthrough in the field of machine learning and many companies are investing in its capabilities.

Deep learning tends to require large amounts of data and uses neural networks. You usually use deep learning only when you have the right amount of data points available to make it useful, as well as the right amount of computing power.

What are the benefits of using machine learning systems?

The best part of machine learning is how it automates more tedious tasks and enhances your productivity. Additionally, companies use machine learning to optimize their products to make their customer’s job easier.

For example, Adobe creates machine learning-based features that allow you to spend less time on endless activities, like sifting through massive amounts of data to figure out who your best performing customers are.

Machine learning also has the benefit of uncovering insights that humans don’t have the brain power to even consider.

What are the drawbacks to machine learning algorithms?

There are a few drawbacks to using machine learning systems. One is the large amount of data required to create an effective machine learning model.

Size and quality of a data set are two of the biggest factors in determining how good a model is, and the more data you have, the more time it takes to label that data accurately for use in supervised learning methods.

You also need to think about the different types of data that you need to add to your data set to make the model a robust one, because you’re teaching the machine to make decisions the way that humans would.

Another drawback of machine learning concerns ethics, especially when it comes to deep learning. A lot of the ways these models make decisions stay in the dark, so you're not exactly sure what factors they are using.

Sometimes this results in models making predictions that seem biased. The model is only as good as the data that you feed into it, but you still might not know what kinds of relationships the model will notice and whether they’re morally fair.

What is the best use for machine learning?

There are many instances where implementing machine learning algorithms would be necessary for the optimization of an organization's resources.

One common case is when the big data is too much for a human to sift through, but it holds important information that could inform company decisions.

For instance, let’s say your marketing team is reviewing new data from different customer segments.

It’s difficult to sort through the different attributes of customers to identify high-performing segments, or what the preferences of Customer Segment A are as opposed to Customer Segment B.

With the amount of data available, it’s not likely your marketing team will have the brain power to process it all, let alone gather any useful insights. This is something machine learning can do quickly and likely better than humans.

Using machine learning automation could result in predictive insights your team might not have found on its own.

In general, AI today tends to be narrow, meaning it’s focused on a very specific problem and follows a set of rules.

This means that AI probably won’t be able to figure out who your best performing customers are, and at the same time, also make predictions about customer behaviors. Different models require different use cases.

Machine learning is better for solving these types of problems because you can use wider scopes and define the problem more broadly.

What scenarios would make using machine learning systems unnecessary?

There are some use cases where it’s better to just create a program with specific rules instead of taking the time to create the machine learning model.

Creating a machine learning model involves a lot of time collecting and organizing data and may also require a lot of computation.

How has machine learning become more democratized?

These days, machine learning is not only accessible to professional data scientists. More people can interact with machine learning as a built-in feature within products they use.

For instance, Adobe optimizes the relationship between our customers and our products, making machine learning more widely accessible.

Our engineers and data scientists train machine learning models based on our customers’ use cases. Then, we put the models into products to make them more useful and user-friendly.

For example, say you’re using artificial intelligence in an Adobe Creative Cloud application to identify an object in an image you’re editing. Machine learning models created by Adobe could outline the object in case you want to cut it from the image.

So, while you may not be creating machine learning models, you’re directly interacting with machine learning and using it to accelerate your workflows.

Another way more people are interacting with machine learning is that businesses themselves are now starting to train their own models and create their own model repositories.

Each business has its own use cases that are customized to their data. Because of this, businesses need the ability to train their own models, rather than having to use a one-size-fits-all model.

To help with this, businesses are turning to automated machine learning, or auto ML for short.

With auto ML, data scientists and machine learning engineers can automate some parts of the model training process to make it faster and to create programs and technologies that democratize machine learning for end users.

How will machine learning look in the future?

In the future, machine learning will play a bigger role in content generation. For instance, generative adversarial networks, or GANs, enable models to create content, as opposed to just making a prediction about content.

This raises more ethical concerns, as well as concerns around “authentic” and “fake” content. But for marketers, being able to generate several different versions of an ad for a campaign using a model instead of doing it manually could be a huge time saver.

People also view