Graduate Education

[Elaro]

Hello. I searched for one, and couldn't find it, so could this be a general graduate education thread? Thank you.

I was envisioning this thread for people to ask questions about graduate education or tell stories about their own experiences. Just an opportunity to share data.

So I'm currently an undergrad in Computer Science at the Université de Montréal, and I have an interest in Computational Problem Solving & Design. As in, give a situation and a goal to an algorithm, and it spits out a set of instructions or a design for the solution. I heard from one of my TAs that this is part of Cognitive Science, but I'd like to hear from other people before I make any plans. I'm also interested in good schools for AI specializations, or ones that are just ok, but off the map, if that's possible. I'd also like to know whether to go straight into a Master's or work in the field for a few years. I'm especially interested Starglider's opinion, because he's the one with the AI development experience, but anyone with graduate experience (especially work and grad exp.) can help inform me.

[Starglider]

Cognitive Science is mostly psychology and neuroscience, usually with some language and information theory, supported by some stats. The artificial intelligence component varies from non-existent to substantial depending on the institution and how many of the teaching staff came from a CS background vs psych / med, but where it does exist it tends to be biased towards artificial neural networks, low-level brain simulation and biomorphic algorithms in general.

Classic logic based problem solving e.g. dynamic programming, parameter space search, Rete inference etc are almost exclusively the preserve of computer science. Essentially because cognitive science is focused on brains and 99% of what brains do does not look like formal/symbolic logic.

Curiously enough I have been getting a little sick of constantly implementing big data / statistical / trendy algorithms recently. I've been working on a little spare-time retrocomputing project involving reimplementing some narrative parsing and symbolic inference demos from the late 70s / early 80s, just for fun/practice, but I must stress that this stuff is deeply, tragically unfashionable and probably won't get you a job at a hot social search intelligence blah blah startup.

[Elaro]

My primary goal is, basically, "program an engineer" (or the part of the engineer that designs parts/machines). I was thinking of squeaking into the R&D lab of some engineering firm. Or working on idea processing, though that might require a math degree.

The problem with symbolic processing is that, at some point, you need to be able to evaluate what those symbols represent. Right?

[Starglider]

My primary goal is, basically, "program an engineer" (or the part of the engineer that designs parts/machines).

As you probably know these is ubiqitous within some domains (e.g. virtually all chip and circuit board design), somewhat common in others (e.g. parameter-optimising CFD simulation for aero etc) and absent for anything with fuzzy requirements and complex-structured / discontinous fitness / very-high-dimensional solution spaces. Computer science has developed numerous techniques for efficient directed search of well defined state spaces, such as chess games, PCB trace routings and logistics network patterns. Engineering as a whole definitely requires general intelligence which we don't have yet. You can choose to work on the frontier of design applications people are currently trying to automate, but that narrow band between 'old hat' and 'still impossible' might not include the stuff you're interested in. Of course the band is constantly moving.

I was thinking of squeaking into the R&D lab of some engineering firm.

I don't know for sure, but as of the 21st century I don't think engineering companies blow much money on bluesky compsci research.

Or working on idea processing, though that might require a math degree.

What do you think 'idea processing' is? I've only heard the term in a self-help-book 'improve the output of your brainstorming sessions' context.

The problem with symbolic processing is that, at some point, you need to be able to evaluate what those symbols represent. Right?

If you mean the so-called 'symbol grounding problem' this isn't normally relevant to automated design software, because the 'symbols' are the physical components and operations that are available to construct the product. Symbol grounding in AI is generally about the extreme complexity and conceptual gulf between everyday words and their referants, i.e. when first coined, toy lisp programs trying to make sense of news articles on a PDP-11 with 1 MB and 1 MIPS. The real world is incredibly complex and even human models of that are pretty complex, deep down researchers really knew that, but no one wanted to turn down grant money and that formal logic cribbed from the Maths department made the papers look terribly distinguished. In fact as noted above I am working on an illustration of this at the moment (when I get time). So 'symbol grounding' would be relevant if you were trying to make engineered products directly from natural language descriptions (without faking it and just searching Thingverse for something similar), but no one sane is trying to do this*.

* I tried to do this a while back for the moderately restricted domain of business software, but those pesky VCs wouldn't fund me. Thus ending up selling Red Queen's Racing Fuel in finance instead.