AI and Machine Learning Explained Simply
An overview of artificial intelligence and machine learning concepts.
Posted Jul 08, 2019
Artificial intelligence (AI) is currently de rigueur and omnipresent. To have an understanding of artificial intelligence concepts is to gain a perspective on the technological lever whose force impacts not only modern everyday living, but also the future.
Machine learning is a subset of artificial intelligence, where computer systems are able to perform tasks without explicit hard-coding of instructions. The ubiquity of AI is owing to a large part to advances in the pattern-recognition capabilities of deep learning, a type of machine learning whose deep neural network architecture is somewhat inspired by the biological brain.
In order for machine learning to work, there must be data to train the algorithms—massive amounts of data. The ongoing global surge of AI in science and industry is due in part to the availability of big data, decentralized cloud computing, massive parallel processing power of GPUs (Graphics Processing Unit), deep learning, and improved algorithms. Computer algorithms are a set of instructions used to calculate and solve a problem.
The quality of AI deep learning depends not only on the algorithm but also on the data. It’s analogous to an internal-combustion car engine, where data acts as the motor oil that enables the proper functioning of the AI algorithm engine—the higher the quality of data, the better overall system performance. The quantity and quality of the data matters.
Supervised learning refers to the process of training AI deep learning algorithms with labeled data. Supervised learning is analogous to when a parent teaches a toddler what things are called. For example, a mom may point to herself and say to her child “mama”—in effect labeling herself. If the child learns to associate the word “mama” with the mother, then the child is rewarded with praise for the correct classification—for incorrect answers, no praise is rewarded. Now imagine training a deep learning algorithm not on just one concept, but on hundreds of thousands of labeled images.
Supervised learning is typically used for classification and regression. One example is email filtering for spam-detection, where data labeled as non-spam and spam email are used for training and testing the AI deep learning classification algorithm.
In reinforcement learning, AI agents make choices based on what it learned from sequential decisions from past experience and iterative attempts to gain the highest reward. Decisions are dependent. For example, when playing the game of chess, each move impacts the next move—decisions are made in sequence and are dependent.
Transfer learning is a machine learning approach where pre-trained models are used for similar purposes. Many deep learning models for natural language processing and computer vision use transfer learning as a shortcut to save time and cost in training neural networks from scratch.
For example, the infamous “Not Hotdog” app (a pivot from a “Shazam for Food” app) from season four, episode four of HBO’s award-winning television series Silicon Valley is not only comedic gold, but also an actual mobile app that runs on both iOS and Android with the purpose of identifying the presence of hotdogs in images.
“Not Hotdog” app developers used transfer learning to adapt Google’s image classifier Inception that was pre-trained with 14 million images from ImageNet, with “just a few thousand hotdog images to get drastically enhanced hotdog recognition,” according to a Medium article written in 2017 by Tim Anglade, who worked on the critically-acclaimed comedy series.
Unsupervised learning is when data is neither explicitly labeled nor has associated outcomes. A way to think of unsupervised learning is how a toddler learns that the reflection in the mirror is herself or himself through iterations of play, versus parent supervising, or guiding the learning process by labeling the image with the child’s name. Unsupervised learning is a form of self-discovery, as it were.
Semi-supervised machine learning learns from a small subset of labeled data along with a large data set of unlabeled data. For example, imagine a parent handing a toddler a chocolate-chip cookie, vanilla wafer, and macaroon and saying, “cookie” each time. Then the parent places a tray laden with more varieties of cookies mixed in with other finger foods such as raw broccoli, pretzels, and potato chips. If the child picks out the cookies from the tray and says “cookie,” she or he had learned from the small sample of training data—in this case, the chocolate-chip cookie, vanilla wafer, and macaroon—what the round, sweet, edible delights are called without having to be explicitly taught every type of cookie in existence.
Whether trained with supervised, unsupervised or semi-supervised, in the analogy, if the toddler selects the raw broccoli from a food tray and calls it a “cookie,” then as the familiar saying goes—"Houston, we have a problem.” Diagnosing why the training wasn’t successful requires examining both the training data and the algorithm.
- Was the training data properly labeled? For example, if the parent handed a piece of an orange as part of the small training subset and said “cookie,” then this mislabeling of training data will impact the training results and accuracy of the deep learning system.
- Was the training data set too small a subset? Just like how some children learn faster than others, some algorithms learn from smaller amounts of data more accurately than others.
- Was the data biased? For example, what if the three cookies from the training data set were all dyed the color green and were made without sugar? Then the answer broccoli would not seem so unexpected given the biased training data.
- Was the algorithm itself faulty? What if the child had reduced abilities in taste and smell due to nasal congestion from a cold, and therefore was unable to differentiate the taste of a cookie from the broccoli? In this case, the data and data labeling did not contribute to the unexpected outcome, but rather the processing analytical engine itself was operating sub-optimally.
Usually, in supervised learning, training data is manually labeled by subject-matter domain experts to prepare it to train the AI algorithm—a time-consuming, laborious, and therefore costly task. The workaround is to automate the training data by generating synthetic data or conduct unsupervised learning.
Whether it is performing a search query, shopping online, engaging in social media, streaming music, or even playing eSports—AI is augmenting daily life experiences. Companies across nearly every industry and sector, such as financial services, life sciences, and health care—nationwide and globally—are rapidly seeking ways to incorporate machine learning and capitalize on this technological wave.
Governments around the world place artificial intelligence as a strategic priority, and are supporting its growth through policy and investments. AI is a strategic endeavor at the forefront of forward-thinking CEOs, COOs, and senior executives leading the business units; it is no longer confined to the realm of CIOs, CTOs, and information technology department.
As fundamental and ubiquitous as the Internet and smartphones have become, AI machine learning is exponentially expanding on a global basis. AI machine learning is being woven into the fabric of modern society, and is shaping the future of humanity.
Copyright © 2019 Cami Rosso All rights reserved.