I have not written anything directly related to my project there. You may wonder whether this guy is a machine learning student or a philosophy student. (; Anyway, I may change my high-security-with-copyrighted-material situation if everything goes this way. However, I try to write something about my project – wish it be fun and encouraging!
Let’s briefly discuss what I have done up to now:
As you know, I am working on learning in behavior-based systems. I have chosen Subsumption architecture as a base architecture due its success in designing a lot of behavior-based systems. I decomposed the learning process to two different situations: 1) structure learning, 2) behavior learning.
In the former case, I have supposed that the designer know how each behavior is working and s/he wants the learning mechanism places each behavior in its correct place. S/he guides this process by giving the agent a reinforcement signal that rewards or punishes its action. In the later case, the designer knows the correct structure of the architecture, but s/he is not aware of the way each behavior must act. For instance, s/he knows that there must be an obstacle avoidance behavior superior to any other behaviors, but s/he does not know what an appropriate action in each case is.
To learn a behavior-based system, one must solve these two problems. What I have done by now is trying to solve these two problems in a special case. I have got some partial results, but the problem is not solved completely.
A New Place: Control Lab
Now I’m in Control Lab! A Pentium 4 -2.8GHz with 512MB-RAM and 80GB of HD is set mine! From now, I may work on my thesis at the university instead of home. Let’s see if it actually happens!
Small-world network and Orkut
Have you ever heard of “Small World Network†or something so? The birth of Orkut encouraged me to see if I can find any related result that show what would happen in this network. As you may have noticed, your connection path length is a small number (mine is 2.8 currently connecting about 320K people), though you feel familiar with your network (it is clustered). D. Watts and S. Strogatz in their “Collective dynamics of small-world networks†(*) showed that this property is possible for small-world networks. This kind of network can be considered as a regular graph with some perturbation in its regularity. They hypothesized that social networks are of this kind and proved their claim by considering some different natural networks like actor networks, power-line networks (yes! It is not a social one!!) and neural network of a worm. I think Orkut is a fascinating resource for these kinds of research.
(*): D. J. Watts and S. H Strogatz, “Collective dynamics of small-world networks,†Nature, vol. 393, pp. 440-442, 1998.
MatLab’s bug: sresid
I am not sure, but it seems that sresid function of MatLab 5.3 is not correct. It is supposed to implement a residualization method of order reduction, but when I want to do so (not a balanced residualization), it do not change A or B or C matrices. Have you encountered this problem?
Meaning and causal relations
Usage is right
Usage wins
All language is folk language
All language is slang
…
So where does that leave us and our term “water” and our associated concept of water? We have 1) molecules of stuff somewhere out there in the world in our rivers and streams. These molecules, as we encounter them, cause physical events to occur, which cause still other events, etc. until some event(s) in this chain ultimately impinge in some way upon 2) some mysterious things happening in our heads; and finally we have 3) our observable linguistic behavior, which presumably is caused or influenced by 2). We have a long way to go before we understand 2) and the exact relationship between it and 1) and 3), but once we do understand these things, there will be nothing left to explain about language and meaning. It is sometimes said that meaning is merely mediated by causal connections between the outside world and our minds. I, however, would say that meaning just is those causal connections, plus some mysterious stuff happening entirely within the mind. Any talk of meaning beyond this has no explanatory or predictive power at all. There simply are no facts about the universe, either extrinsic, third-person “scientific” facts, or subjective phenomenal what-its-like-to-see-red-type facts, that are explained by assuming magic meaning rays connecting our thoughts to trees, cars, and the Milky Way galaxy.
…
from John Gregg, “Language and Meaningâ€Â
on Substance Dualism
“Two Cartesian Arguments for the Simplicity of the Soul”, American Philosophical Quarterly, Vol. 28, No. 3 (July, 1991), pp. 217-226
fMRI: an Introduction (Seminars’s report)
Yesterday I was attending at a seminar by Nancy Kanwisher from MIT. She is acognitive psychologist who is investigating visual activities of brain using fMRI. She is supposed to give 8 lectures about her works, but probably I won’t attend in all of them except those first two one as it is not that much relevant to my work and also I don’t have a necessary neuroscientistic background to understand all of her talk. Anyway, these kinds of stuff can be inspiring.
First, she started introducing fMRI. fMRI which is stands for functional MRI, is a method to measure brain activity. In contrast to what I had thought before, it doesn’t record brain’s electrical activity (due to neuronal activity), but measures the amount it uses Oxygen and convert it to CO2. More brain activity implies more need for O2, and thus more CO2. Accumulation of CO2 changes magnetic properties of that part of the brain and thus it is detectable by MRI. That’s it! That kind of signal is called BOLD (Blood Oxygenation …).
Now, let’s report some of my notes during session:
-Low areas’ BOLD signaling is rather linear in stimuli number. It means that if you have five different stimuli, its activation level is rather the same as adding five BOLD signal of a one stimulus (considering time shift of BOLD response that can be up to 5 seconds). It is shown that this linearity is not the case in higher areas. I am not sure about the meaning of “higher†and “lower†areas. If it is somehow related to {high, low}-level activity, then it is plausible for me. This is my idea about its reason: Lower-lever ones do their job faster and would be free for other tasks sooner than higher-level areas. When there is an another task in the list, the lower-level part is probably free and can process it, so it needs more O2. But in opposition, higher-level wouldn’t get interrupted (due to increase behavioral coherence of the whole system) and so adding new tasks to the list do not need more O2 as it does not work on it. It can be useful for my behavior-based design by considering different modules with different time-scales which some of them is not interruptible (or any-time interruptible), though this is not a very new idea.
-In order to investigate brain, it is better to analyze only one part of it and not the whole. Is there any need to increase data in order to become sure of statistical significance of a higher-dimension hypothesis? It seems plausible but I cannot remember any theorem or clue for it (empirical means and probabilistic bounds!).
-How to make an average over many samples which are not exactly on the same part of the brain in different persons? (e.g. V1 part of the brain is not in the same place for every guy as the brain is not the same for everybody). Spherical coordination of brain seems a good solution to it. Isn’t there any better method? Spherical? Hmmm …
-Why can’t we use a regressioner like NN (or any other general function approximator) to fit measured data and make a simulated version of the brain (or at least, that part of the brain with a priori specified IO)? An important question: Do all of these stuff to get what? Beside that, we can analyze hidden layer of that NN and see what kind of features is clustered and processed together. For instance, if we see that part A1, A5, A17 are activated in the same time (e.g. when the person is given a face picture) and activation level of A3 is decreased below average activity, then it can be concluded that there is some kind of relation between those in face recognition. For example, A1, A5, A17 are active processes and A3 must be inhibited in face recognition. This can be done by observing neural weight of hidden neurons: w(A1) = 0.4, w(A2)=0.0012, … w(A5) = 0.8, … , A(3) = -0.9.
– All of these is somehow similar to system identification. I become curious to know whether neuroscientist have ever used that kind of modeling stuff or not?
– To design an experiment, we can change 1-stimulus or 2-task. They may be some internal clue that some tasks cannot be changes and sometimes it is reverse.
– Angular map between sight-part (in the focus region or far from it) and V1 of visual cortex.
Introduction
Hi!
This weblog is supposed to be a place for informal reporting of my academic activities, specially those related to my thesis. I will write some notes about papers that I will read, my raw ideas and thoughts and brainstorming sessions I may have.
emmm … it is a good place to point out that what kind of stuff I’m doing as my academic activities: I’m a control theory (engineering?! what an engineering course I had?) M.S. student who has been interested in machine intelligence for a long time (although I might not have known exactly what it is about), and most of my current activities is directed toward it. My thesis is titled Hierarchical Learning in Behavior-based Architecture which is a machine learning subject. Even thought I don’t restrict myself to write only about it, and I hopefully write about other stuff like classical control and also chaos control – which I’m working on as my seminar’s project. In addition to these, I may post some philosophical-oriented notes.
And at last, I must mention that the possible readers of this weblog are my thesis supervisor (Dr. Nili) and advisors (Dr. Lucas, Dr. Araabi) and a few of my friends as I do not wish everybody read this weblog or they obey strong version of intellectual Copyright law! lol!