Artificial Intelligence vs Machine Learning vs Deep Learning: How are they all connected?
By Dr. Zain Masood, Chief Research Officer, Sighthound
One of the questions (an important one mind you) that I get asked a lot is the difference between Artificial Intelligence, Machine Learning and now Deep Learning. Are they the same? If not, how are they different? Please give me your thoughts and comments below and share it with your audience if you find it useful.
What is Artificial Intelligence?
Tesler’s Theorem: AI is whatever hasn’t been done yet.
The human mind is a mysterious and fascinating thing. It has the ability to learn over time and process information in a multitude of ways. What is even more astonishing is how the EXACT same information is construed by different people under different circumstances. It is fair to say that the human brain is still an enigma to us and we have a long way to go in comprehending all the different bits and pieces involved in shaping what goes inside this organ.
Looking at it from the perspective of science and technology, it is clear that the ultimate goal is to build systems that can think, act and behave like humans do. And in order to achieve this goal, one need not look further than designing constructs that help mimic how the human brain functions. Since the word intelligence is an adjective used to describe a human attribute, the term “artificial intelligence” was coined that relates to the ability for a machine to replicate how the human brain functions.
Artificial Intelligence can thus be thought of as a system to help mimic human intelligence by using algorithms; techniques, rules and/or methods to learn, reason and react to different situations and scenarios like a human would do. An important thing to understand here (a point that often is not considered) is that this encompasses taking information from all available sensory inputs that humans have at their disposal (seeing, hearing, smelling, tasting and touching) for the decision making process. Even though I am not a fan of the movie, Steven Spielberg’s “A.I. Artificial Intelligence” provides a good depiction of what the end goal (or close to it) looks like. It is for this reason that this is a vast area of research that encapsulates all sorts of learning and reactions that can be expected from humans and how to teach that to machines
Before we move on to the next section, I will leave you all with a clip from the movie “Imitation Game” where Alan Turing answers the question: Can machines think?
What is Machine Learning?
Artificial Intelligence can generally be divided into two categories:
Strong AI (or Artificial General Intelligence (AGI)) depicts a machine with general intelligence which can act almost similarly to humans. As of today, Strong AI machines are only limited to sci-fi movies and tv shows like the WestWorld or Star Trek: The Next Generation.
Weak AI (or Narrow AI) is part of AI that operates within a limited scope and can work with a specific task. These machines can dial in to focus and perform well on a single task. There is no mistaking that these are certainly intelligent machines that can store and process information much more efficiently than humans. However, they work under a far more constrained and limited environment than even the most basic human intelligence.
Machine learning is part of the Weak/Narrow AI paradigm as it is the study of algorithms and techniques designed to solve a specific problem. These machines have the advantage of improving through experience. What this means is that these machines utilize optimization of mathematical and statistical models for a specific task/goal based on sample data (or more commonly referred to as “training data”) to make predictions and decisions that are improved and reinforced over time without being explicitly programmed to do so.
We have seen a number of examples in the recent past where Machine Learning has experienced incredible breakthroughs that have helped shape the world we live in today. These range from Natural Language Processing (NLP) like autocomplete text to Optical Character Reading (OCR) like turning handwriting to text. A known household example of machine learning that nearly all of us use everyday is Google Search; it learns and personalizes the results and gets better over time.
What is Deep Learning?
Deep Learning is a subset of machine learning which is inspired by structure and functioning of the brain when it comes to solving a specific problem. It comprises a computer model referred to as a Deep Neural Network (DNN), a type of Artificial Neural Network (ANN), that mimics how the brain processes information (through neurons) to predict the outcome of a task. The earliest reference to such a system can be found as far back as the 1970s. However, it is only recently that advancements in hardware (GPUs) have allowed compute capabilities to really explore these methods and ensure that justice is done to the word “deep” in the term.
How is deep learning different from traditional machine learning?
One of the main differences between traditional (non-deep) machine learning and deep learning is the process of feature extraction. Feature extraction is a mechanism by which important/relevant cues are extracted from the provided data source, converted to machine readable form and then passed to the decision-making module. These cues or commonly referred to as “features”, help in achieving the machine learning task at hand. For example, a system for classifying “BMW” vs “Mercedes” would extract relevant features from the headlights, front grill, taillights, etc
Some of the popular feature extraction methods were Histogram of Gradients (HoG), Optical Flow (OF), Gabor, etc. Researchers would extract these features from the data and explore mix-n-match techniques that provided the best results.
The drawback or limitation of this traditional feature extraction technique was that it relied heavily on previous tried-and-tested approaches. Additionally, since the mix-n-match process was manual in nature, there was only a limited number of combinations that could be employed and tested.
With the advent of deeply learned DNNs however, it is the system that tweaks and adjusts these features in order to maximize the performance of a given task. This also helps deep networks to utilize significantly more data than traditional approaches which seem to saturate at some point.
On a final note, it is being said that deep learning might prove to be a pathway that finally connects and moves us from the Narrow/Weak AI domain to the Strong/AGI space. It does seem to have taken a good start in this past decade but it remains to be seen if this is indeed the vessel we all have been hoping for. Only time will tell.
Thank you for reading and please share any ideas, comments or thoughts below.