In any profession, it’s important that you know the acronyms. After all, how could someone be a professional in the world of computing without knowing what CPU stands for? Technology continues to innovate, and therefore new buzzwords, initialisms, and acronyms emerge each year. Though some can be dismissed as novelty and fads, others have been laying dormant and only recently ascended in the potential for changing the course of computing history. Artificial intelligence (AI), Machine learning (ML), and deep learning have recently resurfaced as hot topics in both jobs and innovation, and though some of these concepts have been with us since the dawn of computing, understanding the hierarchy and intricacies of each is important now more than ever.
The Foundation, Artificial Intelligence (AI)
Long before AI became a formal science it played whimsically into fiction shaped as of various robots, talking heads, and interactive mechanics. Humanity’s zealous imagination of artificial intelligence would continue even into science. People simply thought AI would be a substantially solved problem within a short amount of time.
The onset of AI is promising because it’s not difficult to calculate that which has structure. Machines were trained how to play tic-tac-toe, checkers, Othello, and more, while quickly advancing and beating humans. Games like this have a very simple concept of knowledge representation. Object, properties, and categories of board games can be easily simulated for a two-player zero-sum game. That’s why they were simple; they are games of perfect information. There are no augmented transition networks for parsing natural language or chance for perturbation of the input. Just a clean “good move or bad move” calculus where both humans and machines can do battle.
The common algorithm for such games, Minimax, is a beginning AI classic. These games lend themselves to a set of states which are possible (the positions that can happen on the board), and the computer is excellent at building a tree data structure of possible states from any given position. This is just as a person thinking one move ahead. You could then evaluate if that move is a good one, bad, or neutral. Most commonly you’ll eliminate the bad moves, and then begin thinking two moves ahead on your existing tracks. Humans do this slowly and run out of mental space, whereas computers have always excelled at this rote operation.
With enough number crunching this method can be 100% accurate on what move to make. As some of you may remember in 2007 Checkers was finally solved! But calculating 39 trillion positions isn’t feasible for every situation in every day. Before an algorithm is solved, it’s percentage of success is quite low. This simple algorithm, though exciting, is terribly limited by a person teaching (programming) what is a good outcome, and what is a bad one. It’s here that Machine Learning changes the game.
Programming without Programmers, Machine Learning (ML)
AI is actually a broad umbrella of concepts. Machine Learning is a subfield of Artificial Intelligence. The part that makes Machine Learning so interesting is that it allows us to train computers without explicit programming. At first, this sounds either magical or fraught with failure, but through Machine Learning we give computers the new ability; the facility of pattern recognition. To continue with the aforementioned example, a good chess player could look at a board and determine if the position is “good” without crunching numbers for follow up moves.
Certain patterns emerge which give the player a statistical advantage, and advanced players have seen these scenarios before. How could a computer know if a game is going well? Sure we could compute a large set of “good” and “bad” patterns for indexing and reference, but that’s going down the same strategy of crunching numbers and building our illustrious state tree. What we need is to build strategy. Machine Learning is building from statistical probability and patterns. As a liberating side-effect, we’ve not locked on the perfect knowledge of games on the board anymore. We can leave the realm of perfect knowledge and build strategies in places no human would ever dare to program!
Though the world of training machines is evolving drastically today, the concept can be described at a high level. A set of strategies is generated, and then in supervised ML, there is a grader (coded by a human) known as the “Teacher”. As each strategy is created, it evolves or iterates slightly. The best algorithm, as graded by the teacher, serves as the model for future mutations. It sounds like our Minimax algorithm from AI 101 but raised by order since it’s dedicated to the best strategy, instead of the best move. The end result is an algorithm that humans don’t necessarily understand, but is extremely useful! Humans made the mutator and the teacher, but not the trained result. So what else can we offload?
Deep Learning – Neural Networks
Just as Machine Learning is a subfield of Artificial Intelligence, Deep Learning is a subfield of Machine Learning. Deep learning speeds up the evolution process by letting the data results feed into the generation mechanism to make better decisions about other data. This creates several deep neural networks. This advances the pattern recognition into an almost comprehensive algorithm. Now the only thing a human programs is the Teacher of what things actually are. When HBO was creating the app for the popular TV Show “Silicon Valley”, they were able to use deep learning to create an algorithm that would properly identify hotdogs, by simply feeding the deep learning system a few thousand images of hotdogs. See their blog post for a comical and yet impressive story. Now to create algorithms, you simply need a large dataset to build your deep learning algorithms.
Which Technique is Best for You?
As a programmer, a system of perfect information with a small amount of data, you can use basic AI strategies in a short amount of time. Applying elaborate and impressive controls, you can have characters in games respond to the player’s location, use compressed sensing single-pixel cameras for real-world tasks, or create an AI for chess.
If you’d like to advance your AI to take on imperfect information, like reading handwriting, you’ll need to implement Machine Learning to create a model of what handwriting should be. A digital version of Plato’s theory of Forms if you will. We will need to create a “Teacher”, which will train your model on what it means to be “good”, and you’ll need a set of variables for your model that can be trained. Armed with each of these, we simply need a large amount of data to train our algorithm. Shifting the algorithm is usually something like the gradient descent algorithm, where variables are shifted slightly, and each is tweaked as training is evaluated. The result is a fairly successful Machine Learning model we can use. In many cases, this is a successful method for your data, with high accuracy. Tensorflow, an open-source software library for AI Machine Learning, has an example of reading characters with ML at around 92% accuracy.
Once you’ve mastered ML, if you’re still unsatisfied with your results, or you need to train a model with even higher accuracy, it means it’s time for you to take your ML to the next level, and implement deep learning. By building a multilayer neural network for your model which is modified using Rectifier neural networks, you can evolve your model in ways to give yourself 99.2% accuracy.
Roles Relevant to Machine Learning, Artificial Intelligence, and Deep Learning
Though scientific AI has been with us for over half a century, the field is just starting to make new breakthroughs. How each level will advance is still unclear. Additionally, as we build deep learning neural networks, how they all work together? An algorithm that can identify road signs is still used by a program written by a human… right?
Now that we’re a long way away from playing checkers. As the field of AI has grown, the roles associated with each have also evolved. As Deep Learning continues to find application, opportunity has spread. If you’re interested in this opportunity here are some positions to consider:
All this data needs a person to store, retrieve, and integrate. That’s what a data architect does. You combine the business with the tech structures to make sure the machine learning lifecycle is fulfilled in the most efficient and scalable way. Glass door lists the salary average at about $113k a year.
Data Scientists (datalogy):
Using scientific methods, data scientists extract insight from information, checking for feasibility of solutions, and finding new solutions for existing problems. Like creating machines that can identify cancer better than doctors. Data Scientists help bring these complex theories to bear in the world. Glass door lists the salary average at about $121k a year.
Inevitably data needs to move, and processes need to scale. Moving thousands and even millions of data points through your ML models isn’t feasible without some kind of hybrid cloud. Cloud architects manage data like data architects, but in a distributed manner. Glass door lists the salary average at about $142k a year.
Creating a method or algorithm for the max output is a quest as old as time. It’s basic probability theory and analysis in a never ending loop. Our strategies have evolved farming, fishing, and culture. With the advent of electronics, delegating simple reactions to computers has resulted in AI. As our data continues to grow, Data Scientists have used this information and began training logic models with Machine Learning, making decent pattern matching analysis of a higher order. And finally, emulating the neural networks of the mind through deep learning.
Computer Science at UoPeople:
Our Computer Science degree programs offer hands-on data mining, machine learning, and AI knowledge and techniques. Example assignments include creating networks capable of learning simple languages or recognizing patterns, exploring algorithms used in board games, and more.
To start preparing for your programming courses and get started with machine learning, deep learning, and AI, check out our Computer Science degree programs and visit our “Prepare for University” section.
A Brief History of Artificial Intelligence (2017). Retrieved from https://www.livescience.com/49007-history-of-artificial-intelligence.html
Checkers ‘solved’ after years of number crunching (2007). Retrieved from
How HBO’s Silicon Valley built “Not Hotdog” with mobile TensorFlow, Keras & React Native (2017). Retrieved from https://medium.com/@timanglade/how-hbos-silicon-valley-built-not-hotdog-with-mobile-tensorflow-keras-react-native-ef03260747f3
A step-by-step guide to building a simple chess (2017). Retrieved from https://medium.freecodecamp.org/simple-chess-ai-step-by-step-1d55a9266977
The current state of the art in objects classification statistics (early 2018). Retrieved from https://rodrigob.github.io/are_we_there_yet/build/classification_datasets_results
How to become a great big data architect (2017). Retrieved from https://blog.panoply.io/big-data-architect-role
Salaires as of early 2018 Retrieved from https://www.glassdoor.com/Salaries/index.htm
This New AI Can Detect a Deadly Cancer Early With 86% Accuracy (2017). Retrieved from http://fortune.com/2017/10/30/ai-early-cancer-detection/