Predictions of Memory

Filtering of the boating party.

2009-08-14T15:25:00.000-07:00

Renoir's "Luncheon of the Boating Party" offers a good example of compelling visual art. The woman near the center is a particular focal point, and one which evokes perhaps the strongest responses in many viewers. I've been playing with visual frequency filters for my research recently, and have found that her face produces a similar effect to that observed in the smile of the Mona Lisa by Margaret Livingstone in her work "Vision and Art: The Biology of Seeing".

If you look at versions of the woman filtered for high:

and low frequencies:

the high frequency images appeared to us to show a much happier face looking toward the other revelers at the luncheon, with the low frequency ones evoking a wistful or even sad individual looking straight out from the painting at the viewer.

Now, we're always building expectations. Even over the short period of time while we fixate our eyes on an object in the periphery. We expect to see a more detailed version of whatever we saw out of the corner of our eye just a few milliseconds before. In the case of this painting a viewer looking anywhere around the edges (i.e. at the other people partaking of the luncheon), the incongruous image of a sad person staring wistfully straight out of the image while surrounded by a jovial social scene would be enough to draw the eye. At the end of the saccade to the center of the painting, their brains would expect to encounter a detailed version of the sad woman previously seen in a low-frequency filtered way out of the corner of their eye.

These expectations would be violated however. Not hugely or strongly enough to really cause a double-take, but just a moment of uncertainty. The woman they are now looking at is smiling and appears to be far more engaged in the goings on than an observer would have previously thought.

The rest of the painting seems to be designed to take advantage of this, using long diagonal lines to draw the eye outward from her face in a spiral to the point where she appears to be a somber and detached figure in the middle of revelry, whose unexpected happiness helps to resolve this larger-scale schematic incongruity which led the viewer to focus on her.

CADIE

2009-04-01T16:06:00.000-07:00

I've been spending a lot of time today wistfully tracing google's CADIE narrative across all their different products, and am now just sitting at her blog, listening to the music and feeling oddly nostalgic for the present.

http://cadiesingularity.blogspot.com/

Literature Search

2008-11-19T10:12:00.000-08:00

Yesterday midterms ended. That's right, yesterday. They were spread out over about 1/3 of the semester, which had definite ups and downs.

In any case, after an evening of remembering what free time felt like, today has been the day that I launch into literature searches for my end-of-semester projects. Two of my four classes have only projects and no final exams, so they need to be good. Luckily, they both promise to be really interesting as well.

In my working memory and general frontal-lobe stuff class, I'm trying to work up a proposal regarding forward modeling and prediction...basically, how parts of our brains learn what to expect, and what happens when expectations are violated. I think unexpected reward is at the basis of why we feel good when looking at or hearing certain things which are considered "beautiful", and this will hopefully help me get a better understanding of how that occurs.

For my course in vision models the project is perhaps a bit more ambitious - I'm doing the preliminary work for an experiment to look at the relationship between eye movement patterns when viewing visual artwork, EEG mismatch negativity (MMN) responses, and the rules of visual design which say certain things create tension and draw the eye. If tension involves the violation of expectations, then a visual MMN response should be generated, which could signal for attention and cause a saccade to the point which violated expectations. There are some really interesting papers already on this topic, which I'll try to post later on today.

It's fantastic to have time to explore this stuff again, and really go where I want to go instead of worrying about finishing a problem set or what I need to know for a test!

What I've been working on recently

2008-11-12T05:49:00.000-08:00

As a quick update, these are the things that have really been capturing my interest recently:

- Perception of line ends
- The effect of the retina's resolution gradient (fovea->periphery) on perception of motion and form
- Mismatch Negativity
- Forward modeling and the circuitry of the hippocampus and cerebellum
- Broca's Area and ERAN/ELAN ERP's
- Dopamine systems and the conditions for unexpected reward
- The interaction of working memory with forward modeling and predictions
- The impact of elements of visual composition on saccades

Midterms have been keeping me busy, but I'm gearing up for end-of-semester projects, which all promise to be very exciting. I'll write more soon!

I still can't believe I'm actually here.

2008-09-18T13:45:00.000-07:00

The first couple weeks of classes have passed now. Two weeks of it being my job to sit around and learn fascinating things as fast as I can. It's amazing.

I'm starting to at least see the routine that things will settle into, and the good news is that it looks very survivable. I have a couple base rules that I've managed to hold on to so far:
- Work is not done in my apartment. There's a cafe downstairs, parks nearby, and I have a cube in a grad office at the department, which is where I'm parking myself every weekday morning at about 9:30 AM.
- Work with deadlines is not done on the weekend. I can laze around and organize my notes from the previous week, or read papers that interested me but weren't assigned. But the weekend is my time and I'm making sure everything that needs to be done for Monday is done by the time it rolls around.

I don't know how long those rules will last, but I know that I'm working an average of 12 pretty solid hours per weekday day right now, which means that there's still a moderate amount of wiggle room before I have to break into weekend activities. This is good, because I'm visiting New York every other weekend as part of a trading off weekends visitation plan with my girlfriend in Brooklyn.

Speaking of work, I should acquaint everyone with the sources from which it flows. I'm taking 4 courses. Three 500 level and one 700 level. I don't know how it works in other departments, but in this one the 700 level courses are more focused on current topics (closer to a seminar), and the 500 are supposedly more focused on established information. In reality, we left the realm of "established true answer" about 10 minutes into the first lecture, but that's just because there is so little about the brain that has been truly established as definite.

So the classes are as follows:

Introduction to Cognitive and Neural Modeling
Vision Models
Speech and Hearing models
Planning and Timing models

I was sitting in on a computational neuroscience course as well (which is more based on the physiology/electrical activity of individual neurons), but when I found out I have to take it for credit next fall it unfortunately became something that gets put off until then.

The course readings are fascinating, and my biggest problem during the day so far has been getting caught up in my work and forgetting to eat. Luckily there's a 7-11 and a few restaurants close enough that I can reach them during the breaks in classes. There has certainly been some adjustment to the type of work being done. Over the past few years, people weren't interested in intentionally making something more challenging by withholding relevant data that they had in their possession. (Such as the answers to homework problems) Conversations happen in a very different and much more ambiguous way when one party has to avoid giving out information that the other party is ultimately seeking, and it's led to me misreading what the professor was saying on a couple occasions. Also, I'm very used to reading journal articles in my spare time, and having to put them down when I get to work. I have to keep telling myself that it really is okay to be reading them at my desk :)

Speaking of reading articles, I may have hit upon the perfect solution. I'm pretty good at reading on screens, but for mathematically based articles in particular I need to participate and scribble notes, restatements of equations, etc. So I got a tablet from fujitsu and stuck kubuntu linux on it (there was a bit of a saga where the touchscreen wasn't supported, so I had to write my first ever device driver, but that's mostly sorted now). Xournal is my first step. While reading a paper or in class it lets me write on, highlight, type on, and insert extra pages/images into pdf files. I knew it was working well when I caught myself doodling in the margins of a document during lectures. Combined with free access to a corporatequality scanner (one of those big document stations), I'm attempting to do a paperless grad school experience, and so far it's working out fairly well. I have two pages of unscanned paper on my desk waiting to be scanned and emailed to me tomorrow.

The second step is a little program called TomBoy (or if you use windows, zulupad is roughly equivalent). This is a personal wiki notetaking system, which lets me break out important terms, concepts, papers, people, systems, etc from my scribbled notes or during class. Each item has a title, and any time I type that title in another item, it is automatically hyperlinked (which is great for obscure terms and people's names). And it takes LaTeX input for mathematical formulas. I'm following a semantic tagging system I developed during my neuroanatomy course last autumn, which makes all the notes amenable to dumping into a SPARQL database, from which I can query things like, "all imaging studies published in the last three years which used human subjects and included a working memory test". And it will organize all the notes I've taken on items which match those criteria. Likewise, I could query for all the parts of the brain which help to connect the pons and the motor cortex, presented to me in order of connection from one to the other. It's fantastic, and it can just continue to grow throughout my time here and well beyond.

Anyway, that's where things stand at the moment. Boston is providing wonderful crisp autumn weather, and I'm going to see if I can't get my tablet onto the campus-wide wifi network and go do some readings under a tree by the river.

Exposition

2008-07-18T12:08:00.000-07:00

This is the part in the show where the narrator catches you up on everything that's been happening in the character's life between episodes.

- He put in graduate school applications, kept his job during the transition, and aced the neuroanatomy final.
- He didn't sleep much during this period, however.
- He was accepted into both Boston University and University of Rochester, with full funding via research assistantship, which was much better than he was expecting to do.
- There was a long period of somewhat obsessive research on the two schools and the academic traditions that he would be buying into at each, along with visits, interviews, and discussions with current and former students.
- While it came down to the last minute, he accepted Boston University.
- Somewhat burnt out, he then proceeded to spend the next few months enjoying his remaining time in DC and thinking minimally about Grad School, at least until it was time to move.
- The move to Boston went smoothly, thanks to help from friends and family. He found a fantastic apartment in a neighborhood that reminds him very much of DC.

And that gets us up to the start of the semester and the next 5 to 6 years of graduate school.

Busybusybusy

2007-12-13T10:30:00.000-08:00

Everything always seems to happen at once. In the next couple weeks I'm facing my neuroanatomy final, graudate school applications (the goal of this whole blog in the first place), and a transition at work -- my group in the company is being bought out, so we all need plan B jobs lined up soon, in case the transition doesn't go smoothly. In addition, I got caught up in the whole facebook/beacon fiasco, with the end result that googling my name is now far more interesting, and I actually have a conspiracy theory about me! You can read it yourself here.

In the meantime, here's what I was planning on posting about a month ago, right before most of the busy times occured:

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

I've been great - neuroanatomy is hard, mostly because I was very used to math and physics and the sort of subject where I didn't need to memorize things if I was good enough at deriving from first principles. Unfortunately, it's hard to do that with the location of cranial nerve nuclei in the pons, so I had to turn to massive rote memorization, which I never liked very much. The course itself, however, is fascinating. We're currently studying the thalamus, which I'm absolutely in love with. Plus, being NIH and therefore a somewhat crazy place, they gave us human brains to take apart and we got to scrub in and observe a brain surgery :)

Other than that I'm agonizing over personal statements, and lining up a third recommendation since the guy I was originally hoping would do it has dropped off the map and can't be reached. I'll be far happier once my applications are all in, and I can go back to doing neural models. I've got a pretty fun system going on with izhekevich models for the neurons and a homebrewed neurotransmitter release/plasticity model (adapted from a few things I found online and a lecture Erwin Neher gave at NIH a few weeks back). It seems to produce a functioning monosynaptic reflex in my AIBO, which is pretty cool :)

Work's been good too - I've taken on hunting the storm worm, which is a hive-structured virtual parasitic organism that moved into an abandoned p2p file sharing network that was left over after edonkey was shut down by the RIAA. It's really advanced in some ways, and hard to study because if the other members of the hive recognize you as an outsider they spread an alarm signal and everyone nearby starts attacking your internet connection. We're attempting to adapt the math from computed tomography to essentially take a CT scan (deriving the internal structure without coming into contact with it - and therefore not triggering an alarm) of the p2p network, though, and it's looking really promising. Our first test is next week on a local network, if all goes well.

URBI is now working!

2007-10-04T03:58:00.000-07:00

[3263661:start] *** **********************************************************
[3263661:start] *** URBI Language specif 1.0 - Copyright 2006-2007 Gostai SAS
[3263661:start] *** URBI Kernel version 1.0 rev.1049
[3263661:start] ***
[3263661:start] *** URBI Engine for Aibo ERS2xx/ERS7 Robots version 1.0 rev. 240
[3263661:start] *** (C) 2005-2006 Gostai SAS
[3263661:start] ***
[3263661:start] *** URBI comes with ABSOLUTELY NO WARRANTY;
[3263661:start] *** This software can be used under certain conditions;
[3263661:start] *** see LICENSE file for details.
[3263661:start] ***
[3263661:start] *** See http://www.urbiforge.com for news and updates.
[3263661:start] *** **********************************************************
[3263661:ident] *** ID: U594360000
[3263788:power] *** Battery at 62 %
cpuload();
[3269224:notag] 0.045281
headPan = 30 time:3000,
headPan = 0 speed:1.4,
headPan,
[01710837:notag] 22.450094

It turns out my AIBO needed a firmware update. Once that was done, all the problems resolved themselves. The above text from a telnet session doesn't look like much, but it represents a connection, CPU load check, and turning of the head in a protocol that allows for realtime communication using java and c++ libraries.

The first step is to write a wrapper class which checks all the motor commands sent to the libraries and caps their speed, and a (probably separate) monitor to make sure that significant force isn't being exerted without motion occurring - because unlike muscles an AIBO's joints won't repair themselves after stress begins to wear them down, and I've found reports that it's possible to tell its joints to move so quickly that they break themselves.

Once these are in place I should be able to start playing with simple neural networks without having to worry about the output accidentally damaging the AIBO. I'm thinking of modifying a self organizing map which was designed to learn the pole-balancing problem (basically given input of the angle of a simulated pole or broomstick resting unstably on a "hand" surface, they learn which forces to apply in order to keep it balanced and not let it fall over) to try and keep the AIBO standing up. It's probably not very accurate with regard to how real animals manage it, but it'll be an interesting first step with somewhat similar output to an animal standing in one place (minor shifts and adjustments occurring, rather than locking everything in position), and once I have something like that going I can add some things like the simple knee jerk (patellar) reflex, which is astonishingly simple to model, being a bifurcating arc that excites a muscle and splits off to inhibit the appropriate antagonistic muscle when it is stretched, causing them to contract and providing a slight counterbalancing force to most motions - a natural damping factor which helps pull you back into place when you start to sway in one direction or another while standing around. This is also why your knee jerks when hit with a doctor's mallet - as I understand it the tendon at your kneecap is pushed back, which stretches the quadriceps slightly and particularly stretches the tendon, where the golgi tendon organs seem to effectively measure "strech". Inferring that the muscle has moved more than it actually has, this reflex arc kicks in and stimulates them to contract, which makes your leg jump forward, since all of the force which had stretched the muscles and tendons is now gone, so the reflex acts relatively unopposed. (Until the same reflex kicks in for the antagonistic muscles, which have now been stretched by the kick)

Since the AIBO doesn't have muscles, but instead direct sensors which provide position data and calculated force (torque) data on each of its joints, my first thought is to translate these into the same sort of measurements that animal nerves have on muscle contraction and stretching due to the intrafusal fibers of the muscle spindles and and golgi tendon organs, respectively. A relative change in position can be calculated from the absolute, which can be transformed to an equivalent "muscle stretch/compression", and the amount of force being applied could be transformed into a measure of how much tension is being generated on the tendons, or visa-versa.

I'm tempted to start off with the simple model of a neuron proposed in Izhikevich's "Neural excitability, spiking, and bursting", from the International Journal of Bifurcation and Chaos, 10:1171-1266. The easiest form of this model (for modeling purposes) seems to be the one put forward by him in 2003, in "Simple Model of Spiking Neurons", from IEEE Transactions on Neural Networks, 14:1569-1070. This model is a true spiking model, and depending on the parameters can be either an integrator or a resonator. Parameter adjustments have demonstrated that it can closely mimic the dynamics of a huge variety of neurons (everything from chattering spiny stellate cells in a cortex to mesencephalic V neurons in the brainstem or the neurons of the reticular thalamic nucleus). I like the consistency that comes in being able to model diverse aspects of the central nervous system by changing parameters, rather than by changing the underlying model.

For all that adaptability, it's a fairly elegant model as well. Modeling the change in a membrane potential v and a recovery variable u, in terms of four dimensionless parameters it runs as follows:

dv = I + v^2 - u
du = a(bv - u)

if v >= 1 then v <- c, u <- u + d

(The <- sign is an arrow, indicating that the left hand term will be set to the right hand term at the next time interval, a sort of state jump)

I haven't yet tried to implement a self-organizing map with this model as the basic building block, but hopefully over this weekend I can take a crack at it. I may also need to find some data on the firing patterns of the nerves which make up the loop of the patellar reflex, so I can make sure the parameters are right. I'll update and let you know how it goes!

More updates

2007-10-01T08:55:00.000-07:00

The past two months have been busy ones.

Linear Algebra went swimmingly. I love the mathematics of abstraction, and have started looking into nonlinear systems analysis and modern algebra. My professor also expressed interested in working with me later on once I've picked up more math, which is very exciting. More than anything it was wonderful to be in a room full of devoted, intelligent people who were really interested in and curious about the subject matter at hand.

Encouraged by my first class, I am now enrolled in a Human Neuroanatomy course at the FAES school at NIH. FAES may be my greatest discovery of the autumn. It apparently began as a program to conduct training for NIH employees, but it's now open to the public. It's fully accredited, cheaper than community college at just over $100 a credit, and offers amazing courses in the evenings and outside of normal work hours. Plus it's NIH, so you're being taught by some seriously knowledgeable folks, and your classmates are more often than not employed in the labs there doing cutting-edge research.

The class itself is a bit more wetware-oriented than I eventually want to focus, and it's taught from a more clinical than engineering perspective (I think I'm the only non-NIH person in the class). I'm still learning an amazing amount, though, and I'm hoping I'll end up with the right kind of baseline understanding of the jargon and quirks unique to this field that I'll be able to take in more advanced papers. Plus, in a surprise that unnerves me slightly, but mostly impresses the hell out of me, our professor announced that he had gotten permission to use some of the extra human brains for a lab section. I haven't done an exhaustive study or anything, but I'd imagine that the list of institutions that just have extra human brains laying around has got to be pretty short. I'm very lucky NIH is so close by.

In other news, my AIBO is now talking to my PS3 which is running yellow dog linux, and has RapidMind (http://www.rapidmind.net/) installed, and I'm wrestling with the URBI (http://www.urbiforge.com/) interface for remote control. It seems more bare-bones than tekkotsu, and has libraries for java so I can do some quicker prototyping before scaling things up in C/C++. There are some issue's I'm running into with the different versions of Open-R - 1.1.2 will run fine, but 1.1.5 r4 (in which URBI is written) doesn't seem to boot on my ERS-210. While those are being hashed out through trial, error, and forum posts I'm sketching out circuits to take care of simple motor reflexes and things like that. They should provide a good starting point for me to move up the "spinal cord" towards higher cortical functions.

Finally, I'm going to have to put some of this aside to finalize my applications and get them set out. They start coming due this month, so I need to get my personal statement into shape (something I generally hate doing), and wrangle my recommendations into a semblance of order.

Bayesian RSS readers?

2007-07-31T04:03:00.000-07:00

This is just a quick note wondering why there aren't more RSS readers with Bayesian filtering built in. It would be amazingly easy to let a user define a series of tags and then tag each story appropriately. You could then block stories with certain tags in order to tailor what was shown to you, and allow certain tags to override that and cause an otherwise blocked story to appear.

It seems quite simple really, but I've actually only found one attempt at anything like this, and it's a somewhat imperfect implementation designed for web-based use...it lives at http://www.nullwhore.com/sux0r/index.php.

I'm in the process of training it, but I think the fact that it performs some type of winner-take-all calculation and assigns only one tag to each story is going to hold it back.

Apparently I suck at regular posting.

2007-07-23T23:46:00.000-07:00

The good news is that I have a lot of stuff to report since last time.

I) I've set a firm date for the GREs, August 22nd. I'm pretty much caught up on multiplication/division, and now I'm just trying to get rid of those pesky careless mistake and speed up my answers. This is getting high priority because...

II) When visiting some schools, my top choices have so far informed me that they look for either near-perfect or perfect scores on the math section of the GRE. It seems perfect is only 94th percentile, and some of these programs want people to either be in the top 5% or have a good reason why they aren't. I'm closing in as time goes by, but it's going to be a race...

III) A paper of mine (and a co-workers) was accepted for publication! We're presenting early november...now we just need to finish running all the data and write the thing :)

IV) I've enrolled in a Linear Algebra class, to be followed this fall by Diff EQ. I never took those in undergrad, although I did take a class on the mathematics of computer graphics, which was an applied subset of Linear Algebra, and is making this class a lot easier. I have an amazing professor and a great bunch of classmates. Everyone's really interested in learning the math, and it's fantastic to be back in an environment like that, even if I have to get to work an hour earlier two days a week in order to fit the class schedule. The good news is that so far I've been able to get all my homework done on the Metro.

V) I'm searching for books on neuro devo (which I honestly can't type without strains of 80's new wave running through my head), on the idea that everyone in this field is trying to model the really complex end result, which occurs after a highly plastic system interacts with the relatively unpredictable outside world for a good many years. It seems like it'd be a lot easier to work out the rules as the system is starting, and has not yet built up the levels of complexity we exhibit in our own adult lives.

VI) I have a scheme to build a really good robotics laboratory for under $2,000, thanks to Sony's financially questionable tendency to put out insanely high-end products at a loss to themselves. CMU recognized the fantastic advantages of the Aibo as a very cheap robot with color vision, stereo hearing, a rangefinder, touch sensitivity, a sense of down/inner ear, and a propriosensory equivalent. Their Tekkotsu framework allows you to connect to an Aibo wirelessly and control it with a computer program. The only problem is that neural-net based cognitive models are computationally intense, if highly parallel. Thanks to Sony's latest venture - the PS3 - I'm hoping to be able to cluster a number of defective PS3's (I'm leaning towards the ones that just have broken A/V output) pretty cheaply. At 6 cell processors, the theoretical performance starts to exceed one teraflop (1x10^12 floating point operations per second). It should be pretty cool.

Okay, that's about it...recommended reading is sparse because it's mostly been GRE prep books and my Linear Algebra textbook, but I did pick up a book called "Dynamical Systems in Neuroscience" which is wonderful: detailed and accessible for both neurologists and mathematicians, which is good because I lack formal training in either discipline :)

Wherein math starts becoming _really_ fun

2007-04-09T06:03:00.000-07:00

The boring part turned out to be not so boring.

Sure drilling math is tedious, but something amazing is happening. Math is taking less effort.

I had a sneaking suspicion that this would happen, like it has with reading and writing and thinking in other languages...I was at the point where I knew I liked the idea of doing math, but math of any complexity quickly became too much effort to be really fun. Mostly it was the results and their applications that kept me coming back and soldiering through. Now, I'm starting to see how if things keep up the way they have been, math could be the sort of thing I get the urge to do after a few drinks, or when I wake up on a saturday morning, and that just makes me really happy, because this stuff is beautiful.

I also feel like I'm solidifying the foundation that I lacked. It wasn't until mid-high school that math really turned around for me and became something I was a) good at and b) capable of enjoying, and since then I've done quite well, but I think I've always suffered from not a lack of grounding in the basic concepts. The hardest part of calculus problems were always the basic algebraic operations, and that sort of thing. If that changes because of all this drilling, then GRE prep may be one of the best things that's happened to me in quite a while.

The boring part.

2007-03-17T05:36:00.000-07:00

Okay, I've hit the boring part of trying to get into grad school. The non-learning part, or at least the non-learning-about-what-I-want part.

I have three things I'm trying to get accomplished before I apply:

1) Do well on GRE's. I have a feeling this is important. I don't think it'll be too much of a problem, but I should at least be studying.

2) Get published in peer-reviewed conference proceedings. I've presented at conferences before, but it was never original research, which is what we're trying to do now. It won't be in the realm of AI, but my current job is in a new field, so original research is still pretty easy, and my company is a fairly respected name.

3) Start something on the road to being patented. It won't be an earth-shaking patent. I do insist that it be something new, despite the fact that our current patent system seems to waive that requirement fairly often. I have two things submitted to my company's patent office, so hopefully they'll decide to pursue at least one of them.

Unfortunately, Most of this stuff is not nearly as interesting to me as what I was doing before, namely reading about neural networks and computational neuroscience.

Anyway, I just have to keep reminding myself that it's all going to lead to my being able to do something that I love, day in and day out.

In which game theory turns out to be far more interesting than I had thought...

2007-02-10T12:17:00.000-08:00

I know this is divergent from the usual subject matter, but I've found it very interesting this past week as a way to model the decision-making process from a very high level. I needed a way to represent the way that in ticketing systems (systems used to assign tasks and their priority in large companies), the priorities of an average assigned task always rise to the maximum allowed, making it eventually impossible to prioritize. For instance, we instituted a new one about 5 months ago in the department where I work. We're not huge, maybe 15-20 people, and another 10 people who can ask us to do things directly. We all know each other, and we're all fairly good at understanding systems. (Most of us are reverse engineers)

When it began, the average priority for a new ticket was 50 on a scale of 0-99. This is because it was the default setting. By now, the average priority of my tickets is in the high 80's, and my top four are all 99. It is to a certain extent obvious why this happens, but I wanted a better way of describing it, so I started reading Wikipedia and then found http://web.mit.edu/14.12/www/. (I love MIT's openness) Now I can talk about the way that in the zero-sum ticketing system game, players have access to complete but imperfect information. They are aware that every player is capable of altering their input in a way that will increase the amount of time (this system's resource in demand) in a manner that reduces the average utility of all other players. This is achieved by selecting a higher number than they think the other players will pick. A player does not, however, know which numbers other players have picked, and so in an attempt to increase their utility, they have to pick numbers higher than what they think everyone else will have picked, knowing that if the other players are rational, they will have picked high numbers in order to maximize the amount of time they get from the person to whom they are assigning work. With perfectly rational players, this leads instantly to the system's only stable state - where all players choose the maximum possible priority for all tickets. This ends up with an exactly even distribution of time, which turns out to be both a terrible way to prioritize, and mentally quite difficult as the number of tickets increases...in the end the priority of a given ticket in the system most often ends up being ignored, because once all players stabilize at the same input, the number ceases to contain any meaningful information, and its utility to the person on the receiving end drops to nothing. And there's probably a way to model that too. :)

What's particularly nifty is that I now have a way to test ideas on how to prevent this stabilization at maximum priority. Offering perfect information would perhaps slow the procession towards the stable state, since a given player would only have to beat all previous players, and might not need to immediately jump to the maximum, but the system still ends up in the same useless place.

It might help to punish players who escalate their input without meeting some set criteria which occurs less than p percent of the time, and from what I've been reading game theory offers a way to determine if there is a value for p which would shift the stable state, and where the different values of it would shift the state.

Another way would be to prevent access to complete information on the part of the players - essentially to prevent them from knowing how to maximize the time spent on their projects. This could be done by having them answer certain questions about a ticket which are then passed to an intermediary agent (either a person or a program) who attempts to assign priorities based on the current goals of the group, the deadlines for each piece, etc. Assuming that the abstraction was effective enough to keep players from discerning a winning strategy, then I think this might prevent stabilization at any particular point, which might actually be what we want, because that way priorities can be made to deal directly with the immediate situation, without having to work counter to natural systemic tendencies.

Anyway, hopefully by the end of the weekend or sometime next week I'll have a better idea how to formally represent all this and prove my ideas. Fun times! :)

Cripes, it's been forever.

2006-12-28T17:08:00.000-08:00

I apparently suck at blogging, but really for the most part I'm writing this for the ability to look back at it 3 years down the line when I'm pulling all-nighters and banging my head against math problems, and to say "Ohhh...so that's how I got here." So the intermittent posts will continue...with hopefully far less inter.

And speaking of banging my head against math problems, it turns out that's both the means and the end. I had noticed a disturbing trend in my head, wherein I had a strong tendency to just skip equations while reading. I really had to work to take them in - even if I was capable of working them out and understanding what they meant.

To remedy this, I decided that math and I would become best friends. Back when I was 5, Reading used to be a slow, plodding process that required directed attention - and now it's one of my favorite leisure activities. Eventually, I want math to be the same way. I have picked up a copy of "Calculus made easy: being a Very-Simplest Introduction to Those Beautiful Methods of Reckoning Which Are Generally Called by the Terrifying Names of the Differential calculus and the Integral calculus". It rocks.

Seriously, this is hands down the best math book I've ever seen, and while I realize that's not necessarily saying that much, it manages to make calculus not only easy, but appealing. I've started to find myself working out problems on the way home in the subway after a night out drinking...and I've had far less difficulty than I thought putting aside my 2-3 hours a day of reading papers and replacing it with a couple hours of mathematics.

Even though I've already learned everything in this book, usually twice over, I'm going through every example and every exercise, trying to drill this stuff in until it becomes fun. Wish me luck.

Who's down with LoC?

2006-11-05T06:38:00.001-08:00

Okay, I'm ashamed to say it. I've lived in DC pretty much my entire life, and I have just now found out about the awesomeness that is the Library of Congress. I mean, I know my way around THOMAS, but yesterday I went there to see their exhibit on cartoons/caricatures, and ended up registering as a researcher. That means, according to their website, that I now have 130 million volumes (counting books, videos, periodicals, etc) available to me, free of charge. And here I've been dragging myself from university to university in the distant suburbs, when one of the world's greatest libraries is right downtown! I'm going to be spending a lot of time there. Plus, the Jefferson building has one of the most beautiful interiors I've ever seen. It's the kind of place that elevates research to an art.

In other news, this whole literature search is going more slowly than I had anticipated. Work has been rough (crunch time) this month, and I've been exhausted when I get home, which is not conducive to teaching yourself new concepts and reading technical papers. Then when the weekend comes, there's always things to do and invitations to hang out with people...it's hard to say no, especially since I don't have a near-term deadline, and I'm not technically a student. When reading papers, I'm still fighting that basic tendency to skip the equations, and forcing myself to sit and work through them is slow going. The one upside I've noticed recently that neural nets and neurology are still holding my interest just as much as (if not more than) when I started in on this, which is a good sign that I might not get bored devoting a large chunk of my life to this.

I am making some progress, though. My BAM is working, and I've taught it some short tunes, represented in an extremely simplistic format. I've also come across Boston University's Cognitive and Neural Systems department, which has quickly jumped to the forefront of my interest. Putting their neural network studies into their own program seems to have given it the kind of focus and slant I want. I've been reading a lot of Grossberg's stuff as a consequence, and it's interesting. I'm currently on "ARTSTREAM: a neural network model of auditory scene analysis and source segregation", but his LAMINART model really interests me too.

A conversation on the train

2006-10-08T21:15:00.000-07:00

This past Saturday I got to use the absolute nerdiest opening to a conversation that has ever crossed my lips. I was on the train, which had been stuck the station before so it was packed sardine-like with all the people who had built up waiting on the platform. Standing next to me were a group of people who I overheard talking, and then I suddenly turned and asked, "Excuse me, but I couldn't help overhearing. Were you talking about modeling the thalamo-cortical loop?"

It seriously made my day. Particularly because I had a paper about that in my hands at the moment. I mean really, what are the odds?

What was particularly encouraging is that I generally held my own in the ensuing conversation, despite the fact that they were studying it actively, and I'm just trying to pick it up on my own enough to study it. We started talking about neuron and cortical column models, and before they got off a couple stops later, one guy suggested I look into something called the max model, which apparently compares the maximum input to the neuron's threshold level, rather than summing them. The obvious question in my mind is how such a system could handle inhibition, other than resorting to summation or perhaps taking the input of the greatest magnitude, which would allow a high-magnitude negative input to prevent firing. Anyway, I'm going to look into it.

Recently I've been reading a lot from UT Austin, trying to size up their program there. Some of my favorites have been:

Yoonsuck Choe's Analogical Cascade: A Theory on the Role of the Thalamo-Cortical Loop in Brain Function, from 2002

Srinivasa V. Chakravarthy and Joydeep Gosh's Studies on a network of complex neurons, a paper on the role of chaos in a more complex variant on a hopfield model

And I've also been enjoying Eytan Ruppin's Neural Modeling of Psychiatric Disorders, from 1995. It's probably a bit out of date, but it provides a nice overview of attempts to model various disorders such as alzheimers and schizophrenia, which is a very interesting area to me, because it should provide a good test in the future to see if we're close to the way human brains work. (Damage them in the same way, and see if the same effects occur)

Practical playing around - and some other papers I've found interesting.

2006-10-01T06:40:00.000-07:00

I always lean best by doing, and at this point I've been reading and reading for weeks, so I decided to make some of my own neural memories, just for kicks. I went back to Kosko's 1987 Adaptive bidirectional associative memories paper, and have just about gotten one working in python.

An ABAM is essentially a neural network with full feedback, where the weights (which can be thought of as long-term memory) are manipulated to store an association between two pieces of information as a stable state. That is, when information A (which in the network can be thought of as a short-term memory of A) exists on one side of the network, then it produces B, which in turn produces A. The great bit is that these memories can stabilize on known states if the input starts out as being somewhat close to one or the other. The cleverly named bidirectional model can recall an association from either direction, so I could start it out with something like B, which would then produce something like A, which produces something a bit closer to B, which produces something a bit closer to A, until the stable state is reached and no more change occurs in the two short-term memory sections. Hopefully my explanation made enough sense to convince you that these things are pretty awesome.

I unfortunately had issues getting numpy set up on my Zaurus, so I had to write the matrix math functions myself, which slowed things down. (If you don't know what a zaurus is, it's my little Japanese palmtop. If you at all like to tinker with computers, I highly suggest you check it out) I do a lot of my reading on the train and bus during my rather long daily commute, and so it's been particularly handy to be able to program readily while I'm in transit as well.

Anyway, I figure an ABAM is a nice starting point, and from there I can begin to play more with input filtering and try linking them together in different ways. I think the first thing I'm going to do is try to get one that can learn up simple melodies. Pretty much all my practical experience up to this point has been with feed-forward networks, because they're faster and make it easier to get immediately applicable results, and the business world likes that.

Anyway, while the programming has been fun, my literature search continues. I'm hopefully meeting with a professor at Johns Hopkins sometime over the coming week to talk about their program, which has me very excited! On top of that, here are a couple of the niftier papers I've read recently:

Bidirectional Associative Memories, by Yeou-Fang Wang, Jose B. Cruz, Jr., and James H. Mulligan, Jr.
(this one is a good overview of BAMs, as far as I can tell)

Biological Plausibility of Artificial Neural Networks: Learning by Non-Hebbian Synapses, by Daniel L. Alkon, Kim T. Blackwell, Thomas P. Vogl, and Susan A. Werness.
(I'm finally starting to find more biologically-based papers, and even though this one is old it was a fascinating one, where they studied the central nervous system of snails and rabits in an attempt to craft a more accurate model of synaptic learning than the standard Hebbian model - which has been in use since we knew next to nothing about how nerves worked.)

The Man Who Mistook His Wife for a Hat

2006-09-25T11:44:00.000-07:00

There's something fascinating about scary things. Maybe it isn't this way for everyone, but for me the things that disturb me exert a huge attraction. That's why - when I saw "The Man Who Mistook His Wife For a Hat" (By Oliver Sacks) on the shelf of a bookstore this past weekend - I knew I would be buying it, and reading it, and possibly not sleeping because of it.

It is a book of real-life horror stories, from a neuropsychological perspective. Stories of people who have lost all sense of who they are, and any ability to form new memories, so they must re-invent themselves from scratch every 5 seconds or so in a frantic race to establish themselves before it all disappears and they have to start again.

Stories of people who loose most of their memories of life, but from the present backwards so that they think they are 20 when in fact they are 60. Unable to create new memories to replace these, every time they open their diary they find that it contains records of the next 40 years of the "future" that they had already lived. And every morning when they awake, they look in the mirror to discover that they have aged 40 years overnight.

There is a woman who went to sleep and dreamed of her childhood in Ireland - of a celebration with music and dancing. Only when she woke up, the music didn't stop. It just kept going.

There is also a man who through a stroke not only lost his vision, but simultaneously loot all memory of ever having had vision, or hearing about it, or ever knowing that anyone else had some weird sense where they could "see" things.

And a story of a woman who was raised from birth without being allowed to use her hands, until she became convinced that they were artificial and she did not really have hands. She lived like that for 60 years, until the day she discovered that the "useless lumps of clay" on the end of her arms were in fact part of her body.

Stories of diseases that give people more energy, and make them think faster, act faster, and overall feel so "dangerously well" that they loose themselves in the bursts of impulsive behavior, running everywhere, flirting shamelessly and offering themselves to strangers, often having to struggle with every ounce of strength just to remember that they had an existence separate from the mad, dizzying impulse that their life has become.

Stories like this may be part of why I want to understand the mind so badly...because understanding diminishes fear, and these things creep me out.

"You have to begin to lose your memory, if only in bits and pieces, to realize that memory is what makes our lives. Life without memory is no life at all...Our memory is our coherence, our reason, our feeling, even our action. Without it, we are nothing...I can only wait for the final amnesia, the one that can erase an entire life, as it did my mother's..."
- Luis Bunuel

The Amygdala

2006-09-09T19:03:00.000-07:00

Okay, so I'm back. I wish I had a good excuse, but I've just been busy reading about fantastic things that I probably should have posted here. I'm new to all this, so you'll just have to bear with me. :)

Anyway, aside from the continual reading, I've been delving into a real-valued neural net designed to recognize randomness recently. With preprocessing to account for character frequency, it seems to be doing great so far at recognizing short strings like filenames. Anyway, this all originally stemmed from a conversation with a friend of mine, which also led to our looking into the amygdala and its connections with the brainstem and forebrain. He maintains that it can be modeled as a chaotic system, which means that I'm now frantically reading up on chaos theory and how to model chaos in neural networks. It is some really exciting stuff!

As for general reading, I'm working my way through Hassoun's Fundamentals of Artificial Neural Networks, reading bits of the IJCNN proceedings for 1992 (I'm particularly liking A Neural Network System Model For Active Perception And Invariant Recognition of Grey-Level Images from volume IV, and working my way through a ton of papers off of citeseer, which I think may be my new favorite web site.

Oh, and for anyone who stumbles across this and is new to neural nets, wikipedia isn't a bad place to start. I'd check out their entries on Neural Networks and on the Memory-Prediction Framework, and then go from there.

It is not okay for people to break.

2006-07-31T20:55:00.000-07:00

This past saturday a good friend of mine died of unknown causes. His girlfriend found him lying in bed, dead several hours. I have numbly avoided facing this so far - instead I've been throwing myself in a dazed and fevered way into research on biological neural nets.

It's comforting, in a very empty way, to reduce people down to machines. That makes it okay for them to break.

Literature Search Update

2006-07-30T18:56:00.000-07:00

So going into undergrad, I (surprise) had no idea what it was I wanted to do with my life. My alma mater, therefore, was not the most prestegious of computer science schools. This was driven home to me over the weekend when I found out they had only one year of the proceedings of one neural network conference in their stacks. That year? 1992.

And since I can't check books out of the DC Consortium system as an alumnus, it looks like I'm going to be spending a lot of time reading in the libraries of GMU and UMD. Furthermore, when I say a lot of time, I mean a _lot_ of time. Assuming the 1992 IJCNN proceedings are representative, then I have an estimated 20,000 pages of very dense literature to search in order to cover the last five years of neural net and major general AI conferences.

A lot of that will get tossed out after a brief glance at the table of contents, but even if I throw out 9 out of every 10 published papers as not what I'm looking for, I will be reading a more over the coming months than I may ever have before.

And I can't wait.

Initialization

2006-07-29T11:20:00.000-07:00

So I spent two hours talking with the department head of computer science at my undergrad school, and the end result is that I know I'm going to go get a PhD.

I don't know where just yet, but I have a plan for finding out where would fit me best. It involves as its first step 3 months of literature search to determine which school has professors who publish in the areas of AI and neural networks that interest me the most. Even the literature search really excites me, which is probably weird, but a friend of mine told me recently that getting a doctorate is really a pretty weird thing to do anyway.

After helping with that plan, he gave me a lot of suggestions on how to get into the places I want to go, how to make contacts, lists of people he knows, the way some other students have made it into schools that they normally wouldn't be admitted to by working as lab assistants and other things...basically he laid it all out in a way that makes it seem possible - even for a guy who left undergrad never expecting to go back into academia and has been working in the corporate world for the past 4 years. And that possibility made me realize that I really want to do this. The research, the late nights working with other people as obsessive as I am, the 75-80% pay cut I'd be taking, all of it.

Coincidentally, for the first time ever I feel as though I have something truly worth blogging. This is a new thing, something I've never done before. It'll be long and hard and probably entirely different from what I expect. So I started this blog, and now I'm off to my old college library to start pulling all the proceedings of AAAI, IJCAI, IJCNN, ICANN, and ESANN I can find.