Quote:Original post by mnansgar Maybe this is a stupid question, but why didn't your dad look over this if he actually did research in the field?
>> It seems that mainly biology and chemistry projects win. I've heard this as well, and I've been told and believe that this is because at the high school level, even advanced students are typically not mathematically mature enough to understand modern research, muchless come up with novel research in the field. Also, it's easier to come up with good experiment ideas in for instance biology since so much is unknown. In math and computer science, you really have to know the field first before you propose a project. If you feel that you can understand the math and abreast of developments in a math field, then I have no doubt you'll have similar opportunities to win.
>> I was hoping that perhaps you could serve as a pseudo-mentor? I am entering >> this in another competition in December(PAPER could use help!!!! :) I'd be glad to assist you if you need some help. I'm just a student myself, going into Computation and Neural Systems for graduate studies (making mathematical models of the nervous system derived from psychological and neurobiological data). I currently take about half of my courses in physical brain science (biology (cell/neuro), psychology (cognitive/psychobio), and mathematics (cognitive neuroscience/logic)) and the other half in simulation (EE (digital electronics/signals and systems)), CS (AI)) for some very hectic semesters, but I have an excellent background in research and graduate coursework so could possibly be of some help.
For the paper though, you can improve the introduction and methods portions a lot, but unless you're willing to run more experiments, you can't do much with the results. Many ANN papers are just descriptive due to the nature of not having exact solutions for the involved differenetial equations, so you may be able to pass it off as that IF you have a good rationale for choosing your networks and some results. You also need to condense your writing a lot and fit everything in -- right now, the majority of it is fluff, so pack in the technical details! I know 20 pages seems like too few, but you'll have to condense everything into in some cases 3 page papers in the future when you publish, so get used to it! The general idea is that if someone can't replicate your experiment exactly from your description, then it isn't a good description. I should have no question in my mind about how you implemented this and why when I'm finished.
>> Tweaking the network would be near impossible(we're talking 100x100 ANN >> minimum). Tweaking weights in that is suicide. It's not the weights you'd tweak, but things like the learning alogrithm and the threshold values (or functions). Like I said, it's HIGHLY unlikely you'll get any learning at all by just smashing all of this together, and you'll need guides to figure out what went wrong. If you can't figure out what's wrong, then again you don't have results. I can't think of a quantitative error term off the top of my head since your system is rather arbitrary, but you'll have to find one or else you won't have a guide for what went wrong and why things aren't learning.
>> I used the word "novel" because.... everyone else who entered used it Well, I agree that it's a good term to use, but only if you can explain why.
My dad did look over it. He found a total of 3 grammatical mistakes, and pronounced the paper as bullshit(I kind of agree).
Since you are in the hotspot of the field, what kind of questions do you think I should be looking into? The hardest part is coming up with the research question.
Quote:Original post by Sagar_Indurkhya My dad did look over it. He found a total of 3 grammatical mistakes, and pronounced the paper as bullshit(I kind of agree).
Since you are in the hotspot of the field, what kind of questions do you think I should be looking into? The hardest part is coming up with the research question.
>> My dad did look over it. He found a total of 3 grammatical mistakes, and >> pronounced the paper as bullshit(I kind of agree). Maybe you should ask him for help with experimental ideas and ways to improve it? Surely he didn't expect you to come up with a feasible idea your first go, especially since you haven't really looked into the field's literature? Wouldn't he be a good point of contact for some questions you can look into?
It will probably take you a few weeks to do a thorough literature review alone (including understanding the topics of course), and that's if you have lots of time to spend on it. When you understand where the field is going and what avenue researchers want to go down, then choosing an appropriate research topic becomes easier. There are plenty of good questions out there; the trick as you acknowledged is in finding them, and they definitely aren't typically that obvious. I definitely don't think you have time in a month to do a literature review, come up with a new project, do it, then write a new paper. I'm somewhat dubious that your current setup will yield results quickly -- it'll probably take a lot of tweaking, and even then I'm not sure how effective it will be. I'm not sure what else I can say from that. I'm certainly no expert in neural networks though, so I may be wrong.
I won't have as much to add as mnansar, but I glanced over your paper and here are some of my initial thoughts on it:
- The format of the paper makes it difficult to read. If it was the format that was requried, then thats fine, but typically for computer science research papers I see single spacing and two columns. Here is a template that Stanford provided for one of their publications: ECJ_demo(tex file used for this on this page). If you have time, I recommend looking at LaTeX to write your report. It makes formatting of mathematical equations and referencing figures and bibliography entries much easier in my opinion.
- You have a lot of fluff in your writing. You should strive for succictness. The following quote in particular jumped out at me (from page 18):
Quote: We find three extremely important: The depth to which the Prediction Network will search the options tree, x, the number of possible action sets, n, and the number of nodes in a branch that will be searched to the next depth (to guard against the combinatorial explosion of the search tree), k. Thus, the runtime of the simulation will depend primarily on these two [three?] variables, and the runtime of the simulation may be noted as the big O of the function, Simulation(x,n,k).
The depth of the search tree, also referred to as the look-ahead depth, is x. n is the number of possible action sets, and k is the branch cutoff at each level.
I see several problems with this excerpt. Primarily, the second paragraph is redundant and should probably be cut out. Additionally, the first paragraph is unnecessarily verbose. I would rewrite it as something similar to this:
Quote: The runtime of the simulation depends on three variables: the depth of the search tree (x), the number of possible action sets (n), and the maximum branching factor (k).
Good luck with future work on this and other papers.
I don't have much to add in terms of critique, but the paper is definately not "bullshit". It may have limited applications in the field of artifical intelligence, but it has no doubt been an incredible learning experience for you, results or not, and at your level, this is really all that matters. I commend you for this.
One concern I have is that I am trying to test an analogy between the human brain, and this piece of software. However, ANN's don't do that great of a job representing biological neural networks. For one thing, the brain does not use error-backpropagation to train itself. I need to understand how the brain train's itself, if that is even the case. One deviation from this project I considered was to try and study biological neurons, and simulate biological neural nets in software. Of course, researchers don't have a full model of the neuron. I'm also thinking about instincts, and basic training. I was in the bio lab, and witnessed chicks hatching. Within the hour, they were walking around. How did they know to do that? Its not like someone taught them.
>> I need to understand how the brain trains itself, if that is even the case. You're correct in that we don't understand this topic yet. It's a big research area that's slowly making progress, biologically. The AI end is reaching somewhat of a standstill with ANNs. I think modeling holds promise.
>> One deviation from this project I considered was to try and study biological >> neurons, and simulate biological neural nets in software. I'd be very hesitant to suggest this to you as a project. The idea is nice (and it is in fact what I'm undertaking for graduate studies), but you're going to really have to really be an expert at both the neurobiology of what you're studying IN ADDITION TO mathematical modeling. The math involved is rather complicated for cell/neuron models, typically systems of differential equations, which can only be approximated numerically. Some software exists to do this (see XPPAuto and NEURON), but frankly you have to really know what you're doing to use them and interpret the results due to poor user interfaces (XPP is particularly notorious for this).
If you want to take the biological route, you'll need access to a university electroneurophysiology laboratory (if you want to study the electrical properties, which are what many neuronal models are based on). The good news is that there are lots of open-ended unknowns here, but the bad news is that good experiments take a long time to design and prepare, as in most scientific disciplines.
I think though that if you actually pull a synthesis of the modeling and biological experiments off successfully, then you'd have a great chance at winning, but these areas are difficult to merge. For instance, before you even begin to design your biological experiment, you need to understand the most recent math models, and also try to determine what factors are missing from them which might contribute most to their results. Personally, I think something that would be worth researching in this area would be the glial cells' effects on neuronal signaling (there is a lot in the literature on this, but not as much in the mathematical modeling literature, and really a good model for it hasn't been determined yet). This is still too general to study though, you'd have to whittle it down some more.
>> I'm also thinking about instincts, and basic training. I'd say no. Instincts (at least those with neural circuits in the spinal cord) are pretty well understood. I don't think there's that much left to model imo. It might be interesting to try and use mathematical cell models to simulate the instinct since the networks are pretty known, but I would bet that has been done (you'd have to check the literature).
I'd be very hesitant to suggest this to you as a project. The idea is nice (and it is in fact what I'm undertaking for graduate studies), but you're going to really have to really be an expert at both the neurobiology of what you're studying IN ADDITION TO mathematical modeling. The math involved is rather complicated for cell/neuron models, typically systems of differential equations, which can only be approximated numerically. Some software exists to do this (see XPPAuto and NEURON), but frankly you have to really know what you're doing to use them and interpret the results due to poor user interfaces (XPP is particularly notorious for this).
...
I've come across NEURON before(didn't understand how to use it). However, I believe that they do a lot of major work with it at Duke University. This is good because Duke is across the street from my High School(boarding, so we can go there any time). Also, Duke recognizes my schools name, so they are usually enthusiastic about kids from here working there. Do you think trying to work at a lab would be a good idea?
I have this notion that for a long time, AI work will be at stand still until biological work and hardware catches up, and then we will see major breakthrough's in AI.
Actually, working in a lab might be your best bet, since you'd have a professor there to guide your readings and research project. When you approach a professor, approach one who does work you're interested in, and also visibly be very excited about the work. You may even want to draft some more specific ideas for projects that are somewhat related to what he/she does to show that you've at least looked into the field a little. You won't be able to get anything for December's competition, but you would definitely get a good head start for next year.
Since biology experiments cost money for supplies and such, you may want to initially propose some specific ideas, ask the professor's opinion, then ask for paper readings along those lines so you can learn more about the subject. If your high school has a course partnership with Duke, ask the professor if you can enroll in an undergraduate research course at Duke next semester with him/her as the advisor (that is, if you are able to). Another opportunity may be to work as a research assistant in his/her lab on one of his/her projects, which would gain you some valuable lab experience (and which would look very good on university applications, whether you win a national competition or not).
If you are actually hoping to go into mathematical modeling of neurobiological phenomena, MAKE SURE THE PROFESSOR HAS A MATH BACKGROUND. If it's a biologist who got into math on the side with no major in it, then the professor isn't a good candidate. Something else good to look for is to check the professor's past papers to see if undergraduates are ever listed in the author list (although typically, you'd have to work with the professor for at least a year and in a significant way to get a publication).
You could check out these publications. The guy, Eugene Izhikevich, has a computationally efficient spiking neural model that can be used for simulating biological neurons. If nothing else, his results using the model that show polychronization are interesting.
