Modern Man is a wondrous being. Within a relatively paltry 50,000 years since Behavioral Modernity, we have managed to build a 500 acre temple complex, fling a hunk of metal 2.1 billion kilometers into space, and, of course, beat Dark Souls using only bananas. We rightly attribute this to what we term as our "intelligence". After all, that is what gave us our name, Homo Sapiens (wise man). But what is intelligence?
The word intelligence seems to have organically evolved to define certain mental traits of humans that other species rarely seem to exhibit, like creative thinking, problem solving, verbal and written communication, self awareness, ethical behavior, and a whole host of other things. With the advent of science and math, we were able to lay down certain rules (or algorithms) that, if followed by some agent, could lead to intelligent behavior, but it was the advent of computers that truly sparked the idea of Artificial Intelligence.
Fast forward 200 years, and we have gotten ourselves into a bit of a muck. The computers have lived up to their potential and beyond, no doubt. It is the humans that are having a hard time keeping up. With terms like Pose Estimation, Data Mining, Cognition, Computational Humor, Hidden Markov Models, Forward Chaining, and Loss Functions floating around, it is hard to retain one's bearing when wading in the pool of AI jargon. Of course, one way to understand this mess would be to check out all these terms individually and try to come to an understanding of all of them, but as with physical exploration, mental exploration could do with a map.
This is where PadhAI comes in. I recently enrolled in their course on Deep Learning so that I can adopt it in my research on Generative Design, and one of the preliminary assignments was to create an "AI Map" that tries to separate out common AI jargon into logical categories so that it could help us understand the architecture of the field of AI better. You can find my assignment attached below, along with a short write up to supplement it. Please do point out any errors that may have crept in so that I can correct it. Hope you find it useful!
AI MAP : Click here for the PDF version
One of the first things we must understand is why we have so much jargon at all. I guess there no need to go further into depth about why we need jargon. Jargon speeds up communication, consolidates ideas and research efforts, and can drive funding. But how does it end up being so "haphazard"? Why is it Machine Learning, but not Machine Vision? Are Computational Creativity, Generative Design, and Machine Art that different as to justify separate terms? Can we come up with a more uninformative term than GOFAI?(I don't think so). I personally wager that jargon originates from a mixture of "emergent" low level interactions and "selective" high level consolidations. Terms originate from the work of individuals and small teams. Most end up in publication material, from which they may end up being selected to be the representative term as more people pledge allegiance to it, or are set in stone by administrative bodies and conventions. Since there really is no central monolithic administrative body that looks over all proceedings in a given field, you see terms that don't have clear boundaries, have significant and ambiguous overlap with other terms, or ones that are simply re-branding of older terms to get in new research money. Also notice, these ambiguities mostly arise in areas and sub-areas of a field rather than very specific techniques and algorithms. Support Vector Machines means a very specific set of methods. Deep Learning on the other hand is an umbrella term for a lot of different types of Neural Nets based on one property, depth. The best response to jargon is not to be intimidated by them, but to rather embrace them. The AI Map is the first step in that direction.
As you can see in the AI Map, we were directed to classify terms between "Abilities", "Tasks", and "Techniques". To state it more clearly, what are the abilities we are interested in, what are the tasks that the system must perform to exhibit these abilities, and what techniques can be adopted to achieve these tasks. I decided to group sub-fields of AI under "spoke word" abilities to make the association easier. An Example : Perception involves sensing different phenomenon, be it obvious ones like light, sound, and taste, along with more abstract ones like gravity, proprioception, and drowsiness. Different sub fields like Computer Vision and Natural Language Processing feature in this division. Taking Computer Vision further, we see that it involves tasks like Object Detection, Classification, Tracking, Pose Estimation and Reconstruction. If we choose Classification, we can see it can be achieved using Decision Trees, Fuzzy Logic, or Support Vector Machines. This list is, of course, representative and not exhaustive. Doing so for a few more sub fields yields us actionable mappings of abilities-to-tasks-to-techniques.
While the overall make up of the AI Map is adequate, I felt the need to add one more section that shows groupings of techniques (Like how CNNs and RNNs are part of Deep Learning). This might clear up a few more doubts about what constitutes what.
Find the AI Map with the extra section below
AI MAP Extra: Click here for the PDF version
That's a lot of ground to cover, but as always, with a map and the will, it will be conquered eventually. Till the next post, toodles!