ML versus DL – understanding the difference

As a product manager, you’re going to need to build a lot of trust with your technical counterparts so that, together, you can build an amazing product that works as well as it can technically. If you’re reading this book, you’ve likely come across the phrase ML and DL. We will use the following sections titled ML and DL to go over some of the basics but keep in mind that we will be elaborating on these concepts further down in Chapter 3.

ML

In its basic form, ML is made up of two essential components: the models used and the training data it’s learning from. These data are historical data points that effectively teach machines a baseline foundation from which to learn, and every time you retrain the models, the models are theoretically improving. How the models are chosen, built, tuned, and maintained for optimized performance is the work of data scientists and ML engineers. Using this knowledge of performance toward the optimization of the product experience itself is the work of product managers. If you’re working in the field of AI product management, you’re working incredibly closely with your data science and ML teams.

We’d like to also make a distinction about the folks you’ll be working with as an AI product manager. Depending on your organization, you’re either working with data scientists and developers to deploy ML or you’re working with ML engineers who can both train and upkeep the models as well as deploy them into production. We highly suggest maintaining strong relationships with any and all of these impacted teams, along with DevOps.

All ML models can be grouped into the following four major learning categories:

• Supervised learning
• Unsupervised learning
• Semi-supervised learning
• Reinforcement learning

These are the four major areas of ML and each area is going to have its particular models and algorithms that are used in each specialization. The learning type has to do with whether or not you’re labeling the data and the method you’re using to reward the models you’ve used for good performance. These learning types are relevant whether your product is using a DL model or not, so they’re inclusive of all ML models. We will be covering the learning types in more depth in the following section titled Learning types in ML.

DL

DL is a subset of ML, but the terms are often used colloquially as almost separate expressions. The reason for this is DL is based on neural network algorithms and ML can be thought of as… the rest of the algorithms. In the preceding section covering ML, we looked at the process of taking data, using it to train our models, and using that trained model to predict new future data points. Every time you use the model, you see how off it was from the correct answer by getting some understanding of the rate of error so you can iterate back and forth until you have a model that works well enough. Every time, you are creating a model based on data that has certain patterns or features.

This process is the same in DL, but one of the key differences of DL is that patterns or features in your data are largely picked up by the DL algorithm through what’s referred to as feature learning or feature engineering through a hierarchical layered system. We will go into the various algorithms that are used in the following section because there are a few nuances between each, but as you continue developing your understanding of the types of ML out there, you’ll also start to group the various models that make up these major areas of AI (ML and DL). For marketing purposes, you will for the most part see terms such as ML, DL/neural networks, or just the general umbrella term of AI referenced where DL algorithms are used.

It’s important to know the difference between what these terms mean in practice and at the model level and how they’re communicated by non-technical stakeholders. As product managers, we are toeing the line between the two worlds: what engineering is building and what marketing is communicating. Anytime you’ve heard the term black box model, it’s referring to a neural network model, which is DL. The reason for this is DL engineers often can’t determine how their models are arriving at certain conclusions that are creating an opaque view of what the model is doing. This opacity is double-sided, both for the engineers and technologists themselves, as well as for the customers and users downstream who are experiencing the effects of these models without knowing how they make certain determinations. The DL neural networks are mimicking the structure of the way humans are able to think using a variety of layers of neural networks.

For product managers, DL poses a concern for explainability because there’s very little we can understand about how and why a model is arriving at conclusions, and, depending on the context of your product, the importance of explainability could vary. Another inherent challenge is these models essentially learn autonomously because they aren’t waiting for their engineer to choose the features that are most relevant in the data for them; the neural networks themselves do the feature selection. It learns with very little input from an engineer. Think of the models as the what and the following section of learning types as the how. A quick reminder that as we move on to cover the learning styles (whether a model is used in a supervised, unsupervised, semi-supervised, or reinforcement learning capacity), these learning styles apply to both DL and traditional ML models.

Let’s look at the different learning types in ML.