AI, Machine Learning and Deep Learning Share Genetic DNA, but Are No Clones – Unravelling the Labyrinth of AI Myths
Discover our AI blog series that debunks misconceptions about the brain power of AI. This is the first blog in the series, but you can also read “AI Does Not Learn by Itself” and “Reinforcement Learning Comes Closest to AI Self-Learning.”
AI was first coined in 1956 by John McCarthy to describe machines that can perform tasks characteristic of human level intelligence. Given its widespread use, AI is a vast topic that spans many different scientific, mathematical, biological, linguistic, technological and philosophical disciplines. Diving deeper, we can subdivide AI into “General AI” and “Narrow AI.”
General AI describes a human level intelligence capable of a general ability to solve multiple (if not all) tasks considered markers of intelligence, e.g. planning, understanding language, recognising objects and sounds, learning, self-awareness, abstract thought and problem solving etc. In other words, this is approximate to what you see in films with regards to AI: Skynet, Terminators, Hal 9000 from 2001 Space Odyssey, and Tony Stark’s Jarvis AI. We are a long way away from this scenario.
Narrow AI can produce results equivalent to human level intelligence in one or very few tasks, but it is often completely lacking in most other areas. Take for example Google’s headline-grabbing Alpha Go AI, which can outplay the Go world champion, Lee Sedol. Alpha Go AI is nevertheless Narrow AI because it can only play Go, albeit at superhuman levels of success.
Today’s AI is Narrow AI. In many ways, today’s AI is not entirely dissimilar to previous waves of automation, generating clever outputs, which are, however, unintelligent in a true sense.
Take the example of a dishwasher. You don’t have to create a robot with human level intelligence and dexterous limbs to get clean dishes. Instead, simpler mechanical processes achieve equivalent results. The dishwasher’s unintelligence doesn’t matter because it solves a finite problem better than human effort alone, generating huge time savings. Today’s AI is similar in this regard – it’s not truly intelligent, yet it is capable of outperforming specific human tasks in certain scenarios.
Delving deeper, Narrow AI can be further subdivided into its two most common techniques: Machine Learning and Deep Learning. Both techniques are rooted in well-established statistical, probabilistic and algorithmic techniques that take as input large amounts of data, typically to perform one of the following objectives:
- Identify patterns – e.g. to group together several documents that appear structurally and syntactically more alike than another group of documents, and so on
- Predict a value – e.g. to forecast a house price based on square footage
- Label a value – e.g. to classify a photograph as either “dog” or “not dog”
The difference between Machine Learning and Deep Learning is taxonomic. Deep Learning is a subset of techniques within Machine Learning, inspired by the human brain. These techniques are known as “neural networks.” Neural networks attempt to represent with code, a highly abstracted interpretation of how we currently understand brain structure and function. Concretely, this means a multi-layered network of interconnected nodes, known as neurons, that process and refine large volumes of information.
So far from being separate and competing disciplines, AI, Machine Learning and Deep Learning are in fact overlapping domains:
Hopefully, it is also clear that, at its heart, Narrow AI is a very complex calculator and not a ‘thinking’ machine. While still an evolving technology and indeed a concept, there are specific applications where this nascent technology (i.e. Narrow AI) can be successfully applied in a business context to achieve efficiency and productivity in manual, multifaceted and convoluted tasks by interrogating large volumes of good quality data.
To this end, depending on the technique used, the AI system must be taught the right things in the right way. For example, training an AI system to classify documents or clauses by type requires a corresponding number of examples clearly labelled to their type – e.g. “Confidentiality Agreement” in the case of document type and “Assignment” for clause type. Whether this effort falls on the user or the vendor depends on the nature of the classification task and, most importantly, on access to a commensurate quantity of quality data.
While there appears to be an explosion of AI technology, it’s imperative to note that its key driver is not necessarily new techniques, as is general perception. Fundamentally, the growing interest in AI is being driven by an increase in computing power, a reduction in cost of IT and an explosion of data. Together, these factors have supercharged experimentation with existing techniques to put theory into practice, which was previously too costly in terms of time, cost and data.
In conclusion, AI is a vast topic and today’s Narrow AI is but a single step toward the loftier goal of creating a General AI. Despite Narrow AI being unintelligent in a true sense, it can automate specific tasks to outperform equivalent human efforts and is teaching us where to look for the next breakthrough in understanding intelligence, and indeed how we might one day create true human-like AI.
Read more of our thought leadership
Learn more about our products