Throughout my travels in Cyberspace; journeying across countless github repos, enduring several senseless Zoom meetings, delving into the depths of many package docs and interacting with Technical & Non-Technical folk alike, I have always wondered where the generally acceptable line between Machine Learning (ML) and Artificial Intelligence (AI) is to be drawn. It is a Friday Evening, I will be eating Ramen and will draw a somewhat arbitrary line in order to make what I believe is a meaningful (albeit nitpicky) distinction.
I assume here that there is a meaningful difference between the two terms, one could certainly argue that ML is AI, but I think that it might be more sensical to say that AI leverages or is one application of ML.
Definitions
To follow along with someone’s distinction between two things, you need a shared language to reason about what X & Y are and how they might differ or be similar;
- Machine Learning- a set of quasi-statistical, Mathematical and Computer Science-esque techniques for “learning” useful things from data
- Artificial Intelligence- a Man-made system that accomplishes some task rooted in the leveraging of knowledge
These are not Merriam-Webster’s words but these definitions follow the jist of the respective ideas as presented in our Cornucopia of knowledge, Wikipedia; ML and AI
AI in particular is suited to this definition as it encompasses Modern Systems like AlphaGo as well as Good Ol’ Fashioned AI like Eliza – also being flexible enough to include things that do not involve “learning from data” such as Static Knowledge Bases or the wizardry yet to be discovered that leads to AGI.
The Difference
Machine Learning techniques are brilliant at learning and AI Systems are those in which we apply knowledge – that knowledge can be transferred or learned but the AI system is not itself doing the learning (an aspect of it, say the Model/Controller/Head, is). Why then is this relevant?
For one, some of the technical folk get angry when any system that is capable of something we might call intelligent gets called AI, be it solving protein folding or becoming the world’s most diversely skilled artist. In this frame of thought, being able to use a neural network doesn’t make you intelligent – which is neither fun nor sensical. Systems like AlphaFold exhibit superhuman intelligence on tasks and are also artificial, your brain can do the sum() operation to combine your representations of these two concepts.
The non-technical folk also get into a ruffle when simple applications of ML expose themselves, achieving 87% accuracy on classifying the likelihood of crops going sour or reducing the dimensionality of a tabular dataset are both great applications of learning, but in vacuum are not themselves acts of intelligence. For methods leveraging Machine Learning to be intelligent, they must be implemented, and the implementation, the artificial one, is what gives birth to AI.
It is worth noting that the model of something like GPT-3 or Stable Diffusions is basically AI. (since they can exhibit intelligence in applications out of the box and great proficiency at those applications) But the models trained alone are not intelligent, it tasks usage in a downstream application for it to exhibit intelligence, up until then it has just reduced error on some data, effectively learning something useful.
All in all in is nitpicky to differentiate between ML and AI – but I think necessary to get all the great folks on the same page. ML is a toolkit of methods for learning and AI is the application of crafted knowledge – you just use backpropagation to sculpt that knowledge.