Speaking The Language: Artificial Intelligence (AI) vs. Machine Learning (ML) vs. Deep Learning (DL)

Every successful enterprise client engagement begins with a common vocabulary.

So much of the hype and coverage around this industry revolves around three terms that seem interchangeable at first glance: artificial intelligence (AI), machine learning (ML), and deep learning (DL).

Getting on the same page about these terms is a simple but important piece of a high-functioning engagement. Given how often they come up, everything flows more smoothly when all stakeholders are able to use these terms precisely. It gives clients a sense of empowerment and confidence in the exact value that we are providing them, and helps us to communicate more efficiently.

The easiest way to think of these terms is as a set of Russian dolls. The outermost layer is artificial intelligence. Machine learning is a subset of artificial intelligence, and deep learning is a specific machine learning technique. Let’s explore what this means in more detail.

What is AI?

In our context, artificial intelligence (AI) is the broadest term. It can be defined as machines exhibiting human behavior. Classic examples of AI come from the world of science fiction, such as C3PO in Star Wars: a robot that can function, speak, “think”, and move in a human fashion.

Expert discussions around AI[1] often revolve around the moral or existential implications of humanity creating something that acts human. These topics, while fascinating, are generally too vague to be directly applicable to a business bottom-line. We prefer to work in a specific subset of techniques that are proven to deliver the types of strategic results and competitive advantages that make this field so compelling:

What is ML?

Machine learning (ML) is a subset of the field of AI which deals specifically in algorithms which modify themselves using data to improve at a specific task.

Let’s break this definition down further:

• An algorithm is simply a procedure used to solve a problem. In the real world, an example would be a cooking recipe: a set of materials and steps in order to produce a delicious dish. In the world of computers, an algorithm is a piece of code which acts on the user’s behalf to carry out a task.
• The idea that ML algorithms modify themselves is what sets them apart from other types of algorithms. This is what puts the learning in machine learning. Whereas a classic algorithm changes completely at the will of the programmer, a machine learning algorithm can figure out its own solutions.
• Data is the heart and soul of machine learning. In every single scenario, you can get much greater improvements by acquiring more or better data than by messing around with different types of models. 80–90% of a machine learning expert’s time and focus should be centered around their dataset. It’s hard to overstate the importance of data to effective machine learning, and this is what makes it so salient for large enterprises who are sitting on treasure troves of it.
• As advanced as these techniques have gotten, machine learning algorithms are still constrained to a specific task or objective. This is extremely important to get right when developing a machine learning strategy — an algorithm that has been told to boost user engagement will act completely differently than one told to increase profits.

Once again, all together this time: ML is the use of algorithms which make changes to themselves in response to data in order to improve at a specific task.

The clear advantage of using machine learning techniques is that they are able to process exponentially more data and find patterns that a team of humans could never do in a reasonable amount of time. See here for some examples of how machine learning is being put to use in the enterprise today.

What is DL?

Last but not least, we arrive at deep learning (DL). Deep learning is a subset of machine learning techniques centered around the idea of simulating the neural networks of a human brain to make decisions. Briefly, this means configuring neurons into layers which collectively respond to input. The technical specifics here will not matter as much to the high-level strategist.

The reason why deep learning has gotten so much buzz lately is because a combination of better hardware and cutting-edge techniques has led to incredible state-of-the-art results across a wide range of tasks, including image recognition, speech recognition, language translation, time series analysis, and more.

However, this performance does come at a cost. Compare deep learning with “shallow learning”, the set of generally older techniques which were the focus of machine learning applications for many years. In general, deep learning solutions will be more resource-intensive (both human and computer), less stable, and less interpretable than their shallow counterparts. This is important to note since many companies have blindly reached for deep learning as a silver bullet, only to be met with the harsh reality that they lack the resources or expertise to utilize it effectively.

Speaking generally, the performance improvements of deep learning over shallow learning are only significant at the fringes (at the scale of Google, Amazon, or Facebook) where they have squeezed every tiny piece of juice out of other techniques. At the very least, I strongly recommend that any project’s first use of machine learning should be centered around shallow learning techniques; this is the same recommendation that the AirBnB (a massive technology company) engineering team made after experimenting with deep learning. Once all of the low-hanging fruit has been optimized — data, infrastructure, etc. — if, and only if, there is still a demonstrable need for better performance, deep learning should be considered.

Where should your focus be?

At Sensai Group, we believe in choosing the best tool for the job. From our point of view as practitioners, the highest-ROI area of this domain for enterprises is machine learning, and this is what we center our work around. This provides the grounding that can sometimes go missing in the clouds of AI, while maintaining the flexibility to choose the correct solution rather than the trendy one.

At the very least, I hope this document enables you to track conversations in this domain more easily, and cut through the noise when hearing about it in the news.

If you’re curious about the types of impact machine learning can have at your company, don’t hesitate to reach out to tiger@sensaigroup.com.

[1] An example of another AI sub-field would be robotics, the study of artificial movement through physical space