Our AI expert goes beyond the buzzword, in a thorough exploration of AI, ML, and DL to help guide decisions around the technology you use to innovate.
Artificial intelligence (AI) has made it into our daily lives more than we may realize. AI programs decide if our incoming mail is spam if we’re qualified for a bank loan, who we should match with on dating apps, and which song plays next in our auto-generated playlists.
Of course, AI isn’t only used for consumer-facing applications. Gartner found that 37% of organizations have already implemented some sort of AI, for backend purposes like chatbots, computer-assisted diagnostics, and predictive analytics. Accounting Organization/Big Four Organization Deloitte, proved the same, publishing additional benefits:
Despite the integration of AI into business and our daily lives, there’s still some confusion around the topic—especially when it comes to distinguishing between artificial intelligence, machine learning, deep learning, and understanding their various abilities.
It is important to know that artificial intelligence, machine learning, and deep learning all rely on statistical analysis to draw conclusions and take action. Part of the confusion around these sciences is that there is no agreed-upon definition for artificial intelligence, so there can be different definitions depending on the source. For clarity, here are three reliable definitions.
Artificial Intelligence is an umbrella term that encompasses machine learning and deep learning. It describes any code, technique, or algorithm which enables a machine to mimic, develop, or demonstrate human cognition or behavior. The associated capabilities are problem-solving, learning from examples, and making predictions. Originally, AI was created using rules-based programming. This means a computer was programmed to perform a specific task based on a set of parameters provided by a data scientist.
While basic AI can only perform tasks or draw conclusions it's programmed to draw, machine learning applications can parse data, learn from data, draw conclusions, and improve over time without following specifically programmed instructions. There are two main types of machine learning: supervised and unsupervised. In supervised learning, the data scientist trains the machine using labeled data and if the program returns an error or unsatisfactory result, the data scientist must adjust the algorithm. In unsupervised machine learning, the program processes unlabelled data to identify patterns and categorize information by common traits, thereby learning on its own.
Deep learning is a subcategory of machine learning that is able to process, learn from, and act on millions of data points. Deep learning applications may be supervised, semi-supervised, or unsupervised. Unsupervised applications can update their own algorithms as they learn to achieve better, more consistent results. Deep learning is commonly used for natural language processing, image and sound recognition, and categorization exercises.
To illustrate the different capabilities of machine learning vs. deep learning, a post on GeeksforGeeks.org gave the example of an AI-powered flashlight.
In the model equipped with machine learning capabilities, the flashlight will turn on whenever someone says the word “dark.” So, maybe someone says, “please get me out of the dark.” The word “dark” will trigger the light.
In a deep learning model, the light could be triggered even without the word “dark.” Someone could say “It’s hard to see with the light so dim.” By understanding the meaning of the speech and the required solution, the deep learning equipped flashlight could power on for any number of different sentences that don’t include a specific trigger word. This requires exponentially more computing power and the ability to learn from data.
Because deep learning programs are self-trained they can produce and act on infinitely more scenarios. But they are also much more complex to build, train, maintain, and fine-tune—for this reason, machine learning will be the preferred solution for many use-cases.
In the past, we measured AI capabilities against the human benchmark of performance.
Can the program solve a problem a human can solve?
Can the program come to the same conclusion as a person?
Now, we expect AI programs to outperform humans and come to better conclusions at faster rates.
When IBM’s Deep Blue supercomputer beat the chess world champion, Garry Kasparov, in 1997, we got a glimpse of a machine surpassing human ability. The algorithm was based on human-designated search procedures, and the AI won the match ultimately because it was programmed so well.
Fast forward 20 years, deep learning programs are now being used to soundly beat humans in multiple games of strategy. AlphaZero, a machine learning program designed by research company DeepMind, is a self-taught master of chess, shogi, and go. Instead of programming the AI to win, data scientists provide it with the rules of the game. The program starts out as a novice each time, but by playing millions of games against itself, AlphaZero trains itself to become an expert. Because the program was not trained on human strategies, it is able to come up with new and unorthodox paths to victory—paths human players could have never conceived. This is an example of unsupervised learning, where the AI can “think outside of the box”.
In today’s business world, machine learning and deep learning are used for everything from optimizing warehouse inventory to identifying at-risk medical patients.
Machine learning is used to make predictions and help companies make better decisions in real-time. For example, consider a machine learning-powered route optimization engine designed to help field service companies improve response time. By analyzing a mountain of data from various inputs including customer location, time of day, weather conditions, technician availability, and parts availability, the program can recommend the best dispatch location and route to achieve faster repair times and reduce fuel costs. As more labeled data enters the program, it learns to further optimize results based on trial and error.
This type of machine learning can also be used to improve inventory management, scrape consumer sentiment from the web, make customer behavior predictions, or dispatch targeted ads to customers based on predefined criteria.
Deep learning differs in that it can actually define or recommend the criteria. Deep learning programs can structure, categorize, learn from, and take action on millions of unstructured data inputs.
With greater computing capacity, deep learning AI is able to translate text, drive autonomous vehicles, analyze health data to predict disease, and so much more.
Most recently, in the wake of the COVID-19 crisis, deep learning is being used to enforce social distancing. In some of its warehouses, Amazon has rolled out Distance Assistant. By processing video images in real-time, the program measures the distance between workers and provides visual alerts when a safe distance is not kept.
Deep learning has other health applications too. For example, a Stanford University project used a deep learning program to analyze more than 200,000 medical records. By analyzing all health risk factors including medical history and genetic disposition, the program was able to flag patients who were more likely to have unexpected hospital readmissions, require long hospital stays, or develop a genetic condition. The program could then alert doctors to follow up with these at-risk patients.
In the same way that a deep learning AI trained on medical data can identify at-risk patients, an AI trained on customer attributes make it possible for brands and companies to predict their customer’s buying preferences. Insights like these help to improve messaging, optimize ad spend, and offer targeted promotions.
We have previously relied solely on our knowledge and experience to create standards of measurements, set parameters, and make predictions. Now, AI enables us to process massive amounts of data to make data-based rules, identify patterns to arrive at superior conclusions and more accurate predictions.
With the AI computing power available today, organizations can explore previously inconceivable ways to optimize processes, improve outcomes, and generate revenue. That journey begins by understanding the fundamental differences between these capabilities.