Modern AI is an umbrella term encompassing several different forms of learning. The main buckets are machine learning and deep learning. But there’s overlap with broader data science as well. Let’s explore AI vs. machine learning vs. deep learning (vs. data science).
This is an excerpt of Springboard’s free guide to AI / machine learning jobs. Download the complete guide here.
What Is Artificial Intelligence?
Artificial intelligence is simply a system’s ability to correctly interpret data, to learn from it, and to use those learnings to achieve specific goals and complete tasks through adaptation.
In general terms, AI is great at automating the routine and repetitive. In other words, it’s great at optimizing. Here’s a familiar example: Amazon Prime used to be powered by people whose jobs revolved around getting your product from their warehouse to your doorstep. That process is a predictable algorithm that does not change from one day to the next. Because of that, the repetitive and boring job in the warehouse could be optimized and handed over to robots. Knowing this, Amazon built distribution centers to enable same-day delivery closer to our homes, and put robots inside of them.
What AI is not great at is creating and thinking outside the box. Things that fall outside of this core competency are: creativity, imagination, holistic viewing of something, arts, motor skills, rhythm, non-verbal communication and cues, musicality, feelings, visualization, and empathy. When the robot carrying the yellow bin filled with products you bought gets stuck, it doesn’t know how to solve that problem. This requires creativity and a holistic approach—a job meant for people. Workers in new Amazon distribution centers now help robots do their jobs. Contrary to the popular media narrative of robots stealing human jobs, Amazon hired 235,000 people to help its robots perform their duties.
What Is Machine Learning?
How are Amazon’s robots able to do what they do? They use machine learning. An algorithm that is coded by hand (like the Apollo 11 guidance system) is a zero-learning algorithm. When it receives inputs, it always responds in the exact same way, learning nothing.
Machine learning is a system of algorithms that receives inputs, produces outputs, then checks the outputs and adjusts the system’s original algorithms to produce even better outputs.
One type of a machine learning algorithm is anomaly detection, which looks for events that vary significantly from the majority of data. Anomaly detection is employed by Stripe’s payment processing service to detect fraud, which is (thankfully) an anomalous event.
An organization called Crisis Text Line uses machine learning to figure out which words, when typed in a text message, are the most likely to predict suicide. To isolate words, it employs a machine learning technique called entity extraction. Then it uses natural language processing and sentiment analysis to figure out that the word “ibuprofen” is 14 times more likely to predict suicide than the actual word “suicide,” and that the 😢 crying face emoji is 11 times more likely to predict that the person is in crisis.
The more complicated the problem we attempt to solve with machine learning, the more sophisticated the algorithms become. Consider computer vision and the effort to identify a red ball. The ball is a simple shape that’s easy to recognize and label correctly. However, if the ball is placed next to a mirror, it stops looking like a ball. If we dim the light and put a small plant in front of the ball, obscuring its features, recognizing the shape becomes very difficult.
The solution to this problem is a technique called representation learning, which just means that we break the red ball into its component features—things that represent a ball that is red. In essence, representation learning extracts high-level abstract features of the red ball—it can be curvature, relative size, and color. But just as the small plant, the mirror, and the dimmed light make recognizing and labeling the ball difficult, the real world can obscure component features and pieces of data that our machine learning algorithms are able to observe.
When extracting high-level abstract features is difficult, another type of machine learning has to be used: deep learning.
What Is Deep Learning?
Machine learning checks the outputs of its algorithms and adjusts the underlying algorithms to get better at solving problems. Deep learning links (or layers) machine learning algorithms in such a way that the outputs of one algorithm are received as inputs by another. Let’s go back to our red ball to talk about a deep learning algorithm, or a deep learning network.
The first layer of the network is tasked with only looking for, recognizing, and labeling dots. When it finds a dot and labels it correctly, it tells the next layer, “Hey, I found a dot! Here it is.” That second layer’s job is to look for dots that are close together. It receives the news from the first layer, and in turn says, “That’s awesome, I found two dots that are close together!” The third layer is responsible for looking for dots that are close together and look like a curve. It receives an input from layer two, which in turn got its news from layer one, and says, “Hey, I think I see a curve!”
These layers work together to first find dots, then dots that are close together, and finally dots that form a curve. On top of them, hundreds of layers are tasked with seeing dots that turn into lines, lines into curves, and curves into shapes that resemble our red ball. This is deep learning: machine learning algorithms that use a nested hierarchy of simple concepts to represent more abstract and complex concepts.
A company called Dialpad uses deep learning loaded with natural language processing and entity extraction to automatically transcribe calls. It then uses sentiment analysis—a deep learning technique—to discern whether the sentiment of the conversation is positive or negative, in real time. This gives people using Dialpad an opportunity to respond to negative sentiments with more empathy and data.
Ready or not, the future is one where people and AI work together to achieve common goals.
Data Science and Machine Learning
One of the reasons machine learning, deep learning, and data science overlap is that they all, in one way or another, deal with data. Massive amounts of data.
Let’s take a quick look at a company that’s making lawyers’ lives easier and making them infinitely better at their jobs. Everlaw developed technology that looks through case law during the discovery process to find documents that are relevant and important to the case. They help law firms, government agencies, and corporations sift through millions of documents in big lawsuits and investigations to find the proverbial needle in the haystack.
How are they able to do that? By creating and maintaining data pipelines for data analytics, storage, and reporting, and deriving insights from various data sources using statistical methods and machine learning models. In the case of Everlaw, data scientists help machine learning engineers design and build better ML algorithms, and use ML techniques to assist developers in implementing new AI features.
The role of a data scientist at Everlaw is a great example of what the job entails on a fundamental level. In essence, a data scientist looks for new data sources, creates pipelines for that data, designs dashboards that make sense of that data, and helps ML engineers build better algorithms.
Another example is what data scientists at Airbnb do. Their role is heavily focused on analytics and building data pipelines that help inform business decisions. They figure out what metrics are most important for the organization today, and analyze them in the right way.
On the other hand, machine learning engineers build and maintain scalable ML algorithms that are based on the core computer science concepts (like data structures, algorithms, profiling, and optimization). Machine learning engineers code more than data scientists, and data scientists make sense of the data that drives the business forward.
Mansha Mahtani, a data scientist at Instagram, said this when we asked for her take on the key differences between the roles:
“Given both professions are relatively new, there tends to be a little bit of fluidity on how you define what a machine learning engineer is and what a data scientist is. My experience has been that machine learning engineers tend to write production-level code. For example, if you were a machine learning engineer creating a product to give recommendations to the user, you’d be actually writing live code that would eventually reach your user. The data scientist would probably be a part of that process—maybe helping the machine learning engineer determine what are the features that go into that model—but usually data scientists tend to be a little bit more ad hoc to drive a business decision as opposed to writing production-level code.”
For more, download How to Build a Career in AI and Machine Learning here.
If you’re ready to transition to a role in this cutting-edge field, consider Springboard’s AI/Machine Learning Course. It’s a self-guided, intensive, mentor-led bootcamp with a job guarantee!