application to PhD programs
Dec. 14th, 2006 07:53 pm![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
guslIt seems I'm only applying to CMU this year (but at multiple departments).
Today, I decided that I'm not applying to Psychology, since it seems like the only faculty I could connect with are not interested in mentoring a thesis on methodology (model selection or model induction). Also, a lot of the work that they do is running experiments, whereas I am a theorist. I am really an AI person. Ion described me as being far on the Computer Science end among the ACT-R people. I don't disagree.
So this means that I'm only applying for MLD (Computational and Statistical Learning), HCII, and Philosophy (Logic, Methodology & Computation).
Please comment!
My goals and research interests
As an AI person, I have long been interested in the parallels between human and machine intelligence. For instance, learning (human and machine) can be seen as a general information process: learning is compression (philosophers of science will know that a unified theory provides a shorter way to describe the data). Furthermore, machines can be programmed to use the same representations as humans doing same task, as is evidenced by the existence of cognitive models (and unintentionally, many a program written by human programmers).
Machine Learning teaches us how to do induction effectively and accurately. Philosophy of science teaches us how to do induction "correctly". Active learning teaches us how to explore optimally (i.e. maximize useful information gain), and this knowledge is just as useful for the scientist choosing what experiment to perform next. In practice, however, much science uses intuitive methodologies and "tacit knowledge". I see it as my job to formalize these practices, thus putting them under scrutiny. Cognitive scientists have intuitions about what makes good cognitive models, and don't yet use machine learning to select among competing models, or to induce new models from data. I see it as my job to change this.
Despite the similarities, there are tasks that humans do better than machines and vice-versa. Human-machine collaboration holds the promise of augmenting human intelligence. However, most of this collaboration is based on the "calculator model", in which machines are used in a very restricted way, and there is no "transparency" in human and machine knowledge: human and machine only interface successfully at a few select places in the information-processing.
It is not easy to design machines to help us with general reasoning tasks because such tasks consist of manipulating concepts, which are too numerous and too slippery to "teach" to a computer. Reasoning systems in restricted domains have achieved some degree of success (after significant ontology engineering work), but scaling this to domain-independent reasoning has proven to be a very hard problem.
Areas of interest:
* scientific discovery, machine learning, induction methodologies: how can we combine theory and data?
* information theory (MDL learning)
* active learning
Domains of interest (i.e. data sets I would love to have):
* cogsci: behavior logs AS WELL AS human-made cognitive models
* demographics
* medical data
* natural language
* computational biology
Research experience
SimStudent Project (Sep 2006 - present)
I have been working as a programmer for the SimStudent project, which induces production rules from demonstrations. The idea is to learn cognitive models from behavior data. On the machine learning side, it lacks dynamic feature selection: we have hard-coded a set of operators through which the demonstrations can be learned.
On the cognitive side, it's not a good model of human performance, since it doesn't enforce cognitive constraints: i.e. it doesn't account for the fact that the learner is a human being.
Master's thesis (Apr-Sep 2005):
My master's thesis was about formalizing derivations in physics.
* created an equational theorem-prover that shows the set of models under which equations hold.
* created an ontology of classical mechanics, and formalized some derivations.
* used the above derivations to discover new proofs
Automatic dictionary builder (Jan 2004):
* designed and programmed an automatic dictionary builder for sentence-aligned parallel corpora, by doing Portuguese-English word alignment. Used several heuristics, including cognate-matching, correlations in word-length and word-frequency, and bootstrapping from a small hand-made dictionary.
Today, I decided that I'm not applying to Psychology, since it seems like the only faculty I could connect with are not interested in mentoring a thesis on methodology (model selection or model induction). Also, a lot of the work that they do is running experiments, whereas I am a theorist. I am really an AI person. Ion described me as being far on the Computer Science end among the ACT-R people. I don't disagree.
So this means that I'm only applying for MLD (Computational and Statistical Learning), HCII, and Philosophy (Logic, Methodology & Computation).
Please comment!
My statement of purpose
My goals and research interests
As an AI person, I have long been interested in the parallels between human and machine intelligence. For instance, learning (human and machine) can be seen as a general information process: learning is compression (philosophers of science will know that a unified theory provides a shorter way to describe the data). Furthermore, machines can be programmed to use the same representations as humans doing same task, as is evidenced by the existence of cognitive models (and unintentionally, many a program written by human programmers).
Machine Learning teaches us how to do induction effectively and accurately. Philosophy of science teaches us how to do induction "correctly". Active learning teaches us how to explore optimally (i.e. maximize useful information gain), and this knowledge is just as useful for the scientist choosing what experiment to perform next. In practice, however, much science uses intuitive methodologies and "tacit knowledge". I see it as my job to formalize these practices, thus putting them under scrutiny. Cognitive scientists have intuitions about what makes good cognitive models, and don't yet use machine learning to select among competing models, or to induce new models from data. I see it as my job to change this.
Despite the similarities, there are tasks that humans do better than machines and vice-versa. Human-machine collaboration holds the promise of augmenting human intelligence. However, most of this collaboration is based on the "calculator model", in which machines are used in a very restricted way, and there is no "transparency" in human and machine knowledge: human and machine only interface successfully at a few select places in the information-processing.
It is not easy to design machines to help us with general reasoning tasks because such tasks consist of manipulating concepts, which are too numerous and too slippery to "teach" to a computer. Reasoning systems in restricted domains have achieved some degree of success (after significant ontology engineering work), but scaling this to domain-independent reasoning has proven to be a very hard problem.
Areas of interest:
* scientific discovery, machine learning, induction methodologies: how can we combine theory and data?
* information theory (MDL learning)
* active learning
Domains of interest (i.e. data sets I would love to have):
* cogsci: behavior logs AS WELL AS human-made cognitive models
* demographics
* medical data
* natural language
* computational biology
Research experience
SimStudent Project (Sep 2006 - present)
I have been working as a programmer for the SimStudent project, which induces production rules from demonstrations. The idea is to learn cognitive models from behavior data. On the machine learning side, it lacks dynamic feature selection: we have hard-coded a set of operators through which the demonstrations can be learned.
On the cognitive side, it's not a good model of human performance, since it doesn't enforce cognitive constraints: i.e. it doesn't account for the fact that the learner is a human being.
Master's thesis (Apr-Sep 2005):
My master's thesis was about formalizing derivations in physics.
* created an equational theorem-prover that shows the set of models under which equations hold.
* created an ontology of classical mechanics, and formalized some derivations.
* used the above derivations to discover new proofs
Automatic dictionary builder (Jan 2004):
* designed and programmed an automatic dictionary builder for sentence-aligned parallel corpora, by doing Portuguese-English word alignment. Used several heuristics, including cognate-matching, correlations in word-length and word-frequency, and bootstrapping from a small hand-made dictionary.
![[identity profile]](https://www.dreamwidth.org/img/silk/identity/openid.png){kind=small}
(no subject)
Date: 2006-12-15 04:42 am (UTC)I didn't know CMU allowed this. Berkeley only allows you to apply to one department in a given year and I was of the impression that this was a common practice.
Anyway, to me your statement of purpose makes you seem a bit unfocused. I think it would help to state your goals more explicitly and say what you intend to do as a graduate student to advance those. I don't know if this is the place for it, but somehow you should definitely say how you fit into department X: why they are a good match for your interests, skills, etc..
(no subject)
Date: 2006-12-15 05:01 am (UTC)http://www.cmu.edu/prospective/graduate.shtml gives the impression that graduate admissions are decentralized. I never read any rules against applying to several departments.
I don't know if this is the place for it, but somehow you should definitely say how you fit into department X: why they are a good match for your interests, skills, etc..
Yeah... but I guess a perfectly good reason is that the department is inter-disciplinary.
My goal is to be an AI researcher. I can't be more specific than that: I can only elaborate on my motivations (which are largely philosophical), what kinds of problems/questions I find interesting/fun. I have a hard time doing this without diverting into philosophical ideas.
Then again, "AI" is more specific than "computer scientist", which is all that they expect you to know when you're starting a PhD in SCS..
As far as seeming more focused: I guess I could talk about the specific problem that I'm working on now, but who the hell knows if I'll still be working on it a year from now? OTOH, it seems very likely that I will be working on similar ideas for a long time to come... I have always been interested in epistemology, formalization and AI.
Argument-mapping is only its current incarnation. Over break, I'm going to do read up on causal induction.
(no subject)
Date: 2006-12-15 06:07 am (UTC)Then again, "AI" is more specific than "computer scientist", which is all that they expect you to know when you're starting a PhD in SCS..
That's not quite true. In my statement of purpose I said something like I definitely wanted to do theoretical CS, but hadn't decided yet between algorithms and complexity theory. And from what I can tell, my SoP was a bit less focused than most people's. Saying "CS" isn't enough for SCS, and "AI" might barely be enough.
But I think it's not true that you can't be more specific than just "AI", since there are all different kinds of AI, and presumably you're not interested in all of them. For example, there's machine learning vs. other kinds of AI. And there's learning theory vs. practice. And there's planning, sensor networks (from an AI perspective), distributed agents, etc. Any PhD program is going to require you to specialize at some point, to do a thesis if nothing else, so you can't really hope to work on all of AI. You might as well take this time to figure out what you're really interested in and focus on that.
You're not applying to CSD? We're the coolest of the SCS departments :-).
(no subject)
Date: 2006-12-15 06:22 am (UTC)I'm into symbolic AI: reasoning at the conceptual level. I don't know that much about planning or all these formalisms for probabilistic reasoning... it seems like it would take lots of work to get up to speed. But it might be worth it, who knows?
AFAIAC, CSD does cool stuff in two areas, namely logic and algorithms.
Logic here (i.e. the PL people) is too theoretical for my taste: it doesn't seem to concern itself with how real mathematics is done. I believe that mathematical intelligence should be about reverse-engineering human mathematicians. But to be fair, I don't quite understand what their goals are.
Algorithms make fun puzzles and useful applications, but I'm mostly interested in automating you guys away. As you can see, I really am an AI person.
(no subject)
Date: 2006-12-15 04:10 pm (UTC)This is definitely not the case. Most statements of purpose of people who they accept are going to be much narrower than this. Writing "Some specific problems I am interested in working on are X, Y and Z," isn't a firm commitment to doing X, Y, or Z. It shows them that you've thought about what you want to do as a graduate student, how you fit into the program, etc.. And that you're willing to focus most of your energies one or two narrow problems.
Great
Date: 2006-12-15 07:37 am (UTC)(no subject)
Date: 2006-12-15 07:50 am (UTC)