He gave me a lot of readings for you :)
My suggestion is that you read the long paper first, but focus particularly on the structure of MLN, rather than the proofy bits. So maybe skip 10-17 and read over the applications as thoroughly as you like.
Then please read the other 2 papers carefully :)
https://umdrive.memphis.edu/aolney/internal/cogsci_readings/pedro_readings.zip
Difficult readings. I could only get an idea of what it is all about, why it is useful, but not really how it works. I did find many interesting things, but I definitely lack the background knowledge to attempt deep understanding. These are some of those articles that tests your internet browser's capacity to keep opened 65,535 tabs of wikipedia articles! A few more weeks of the CogSci Seminar and I will start reading prof. Bollobas' articles!
ReplyDeleteOne of the interesting things that I read lead me to my (rhetorical?!) question: is there someone who can explain to me Approximate Inference in Markov Logic,together with a short example, in less than 20 minutes or in less than 20 unforgiving paragraphs?
I share Cristian's sentiments. I understand the utility, but not the application.
ReplyDeletePart of this seems like "chunking". Is that an inappropriate inference? As though the information from the larger, more explicit understanding is chunked into more efficient units?
Regarding giving weights to different variables... I may have missed this, but how are the weights assigned?
The results in the USP study were staggering? What happens when something like this clearly outdoes the competition? Will USP be the next bandwagon? Is it already?
Are the graphical models mentioned similar to the conceptual models Sowa preached about?
So... I finally am (relatively) understanding this stuff and why it's useful (with the help of also 65,000 tabs)! I couldn't write the programs or formulas, but it's starting to all make sense about why these different models matter. (Barely, but it's a start!) I do wish the MLN paper had been a little more relational between the different logic models and something that matters in "our world" than the pure logic proofing sort of things.
ReplyDeleteThat leads me to say, that I was more excited about the probabilistic theorem proving paper, in that there's a clear comparison between FOVE (which has taken me forever to figure out, and I still may not get it all, but...) and PTP. To me, it makes sense that PTP is more efficient than FOVE because we can infer with likelihood (i.e., the probability part) rather than using up all the computational resources on the different components of FO variable information. If it's still kinda unclear to people, take a look at Figure 2b in the PTP paper (it took me forever going through this stuff). You see that within just the first response, FOVE not only started with around 10 times the time to respond (.01s vs. 10), that grew exponentially within the first 100 objects (>100s. vs 1s). Just trying to help those of us who are "logically challenged." :)
I still am unclear on the weighting process, but I don't know how crucial it is for me to know that. Oh, Jeremy posted about that too.
Holy crap... those were tough readings!
ReplyDeleteThe only background I have in any of this is a beginners symbolic logic course I took almost two years ago. =(
I agree with Rick that what the articles lacked was a grounding in the real world. I understand that they are trying to make a model that combines logic and probabilities. It is an interesting project, but I am unsure what it has to do with cognitive science...
If the ultimate goal is to replicate how the mind does its work then I definitely have doubts that this is the right course. It has been shown (and was mentioned at a recent lecture) that human beings aren't even very good at simple logic so how could a much more complicated version be closer to correct? I would say just stick with pattern completion and association.
If the ultimate goal is to make some kind of AI that can outperform humans then maybe this is a good way to go. Unfortunately for me, I really can't contribute much to that discussion.
It seems for me that MLN attempts to solve the unsatisfied condition that the first-order logic that fails, and then it combines with the probability model. However, the algorithm MLN uses is more difficult, because it is involved in the profound knowledge of probability. However, it does not explain it from the basis, which makes it hard to follow the proving steps.
ReplyDeleteLike Benjamin said, I also have doubt about this models to simulate the human brain’s work. We can’t say it is on the right track or not in haste, but how can we evaluate its efficiency and accuracy?
Sometimes I feel shocked that these models have too strong reasons to lose the nature of the human reasons. They mostly depend on the diverse and powerful algorithms lack of the flexibility of human mind. Perhaps one day AI will outperform human minds, but it is not the real human, still a machine that is manipulated by human being.
When humans conveys their meaning, it is dependent on such various settings as body languages, contexts, facial expression, time and space and so on. Can meaning representation with algorithm reflect such complicated meaning?
I agree that these were difficult readings primarily because of their density. Some walkthrough of the algorithm would be great. Because while I understand the general structure and the behavior of the model, I would appreciate a discussion about how it works particularly if there are any parameters that can be changed or added that may affect algorithm's behavior (how weightings are determined or used maybe?).
ReplyDeleteI disagree with the notion that the higher complexity of MLN makes it a poorer POTENTIAL (or at least closer) model of the human mind. One of the reasons MLN seem so attractive is that it sacrifices some "certainty" in the inference which allows for greater flexibility. This seems much closer to the human ability to handle ambiguity. Also when compared to FOL, MLN take significantly less computing resources mostly in decreased memory space. My understanding was that instead of keeping every proposition in the KB on hand in FOL that MLN can concentrate on the highest weighted ones. This seems to get much closer to humans in terms of capacity since humans (compared to computers) do not have a large, reliable enough memory to make use of it. MLN also strikes me as being much closer to the heuristic nature of human reasoning.
While MLN was impressive, I not ready to jump on the 'bandwagon' like others as it being THE model of human cognition because there seems as others point out to be more than just inference occurring in the mind, but I don't think that is there intention anyway (they struck me as mathematicians just seeing what they could do with Markov Logic :P). Another point that I thought was interesting was that MLN can handle both a stochastic and deterministic world since MLN which use just probabilities of 0 or 1 is pretty much FOL.
I was impressed by MLN doing a great job in finding the core semantic value for multiple syntactic structures. However, I am not persuaded by any reason how MLN achieve that goal. "To handle variable number of arguments, we follow Davidsonian semantics and further decompose a lambda form into the core form, which does not contain any lambda variable. (p.4)" This decomposiiton step seems contradictory (or just confusing) to its assumption of using lambda algorithm to encode semantic entities. Need some help here.
ReplyDeleteSince I'm going to be presenting this articles and leading the discussion tomorrow, I'm going to keep my comment/question short. I definitely feel like I got bogged down in the equations and logic statements in two of the papers. In the longer MLN paper I still feel like I could get a grasp on what was going on, but in the Unsupervised Semantic Parsing paper I was pretty lost. So my general question is, what was going on in that paper?
ReplyDelete-Blair
I know this is a very small point and only barely mentioned in the longest article, but it interested me, so here you go... I find it hard to believe that MLNs could be used very effectively for social network modeling. Perhaps that just because I’m being stubborn and don’t like to think that our actions, relationships, and choices in general can be summed up with something so rigid. They mention two people having the same habits or attributes, but I’m not convinced that that’s all there is to friendship. I see their point, though- obviously our interests must overlap with those of our friends to some degree. I was interested in the brief mention of where MLNs might go in the future as far as their use in learning.
ReplyDeleteAt first I agreed with Benjamin that this approach didn’t seem appropriate for how our minds work, but Chris had a point (and he clearly understands MLNs better than I do) that MLNs take up less memory space, and don’t need to keep every proposition, but only the ones with the highest weights. This makes it seem that MLNs could potentially be used in model of our cognitive processes. I still tend to I’d have to understand them better to really state an opinion.
So, from "attaching weights to first-order logic" I'm guessing that means that every semantic item should have a "weight" so not all items are analyzed equally right? So in this way you can know which items account for more of the logic and which items matter less. This is the elementary understanding that I'm getting.
ReplyDeleteA little disappointed that it's all logic and math and little else. I wish the papers talked more about the applications.
Well. These articles were... ahem, yes. I can unequivocally state that in fact words were committed to page, and that I read them. Err... they just didn't mean anything to me.
ReplyDeleteThat's not entirely true. I had a few glimmering moments of insight, particularly where problems we have encountered in our lab lined up with things mentioned in the articles. I finally found someone who addressed the antonym clustering problem - where opposite meaning words tend to be grouped together because of their similar properties and circumstances, with no clear negative "features" to tell them apart. Poon & Domingos (pg 5) say that they resolve it by looking for conjunction tags. Does anyone have a better idea of how this is implemented? Does it rely on "XYZ but ABC"? Or is it correlative conjunctions rather than coordinating conjunctions? And wouldn't it matter which conjunction was used? Using only the tags seems like it doesn't capture the purpose of the conjunction (e.g. 'and' is different than 'but'). I want to know - we had similar problems in the domain of biology, where contrasting happens quite often.
In the long paper, the authors mention applying this to computational biology, which I would find interesting since it's far more complex that departmental dynamics. Also, one of the subfields of computational biology is "computational neuroscience", which models brain function or structures in terms of its information processing properties. Is it ironic to try to come up with models of brain function with a method that some of my classmates claim is "un-brainlike"? Would this hinder this technique's ability to model these kinds of structures?
I think I'll briefly jump in again because Benjamin said something I don't agree with and Chris' response encouraged me to pursue this.
ReplyDeleteSo I would like to ask Benjamin how would he try to model the human mind in the following task: suppose someone has to solve the following physics problem:
"Suppose a car is able to accelerate at two meters per second. What acceleration can the car attain if it is towing another car of equal mass? Explain why."
Obviously, I don't mean to ask neither how the concepts and relations are extracted from natural language, nor for a very accurate neuronal interaction model (which is Science Fiction at the moment).
You say people aren't very good at even simple logic. This problem uses simple logic and most people solve it without difficulties (provided they have the basic conceptual physics knowledge).
So, Benjamin, how would you model this without logic? What about a more complicated problem, from, say, AP Physics, or Einstein's IQ Test?
I am not saying that the mind always works like when it is trying to solve logic problems. Actually, I do believe our mind tries to do logic using pattern matching, which is backwards and inefficient, just like trying to write something on a lawn just by using fallen leaves that you can scoop with a big shovel/rake. The position of each of the leaves does not mean much to the message, whereas the "whole picture", the sum of the parts, does!
So modelling the mind of the lawn writer, wouldn't you want to at least take into consideration the meaning of the lines and curves that he created with heaps of leaves?
I'm with Whitney. Holy confusing. Really looking forward to hearing some more about this from people who know what they are talking about tomorrow.
ReplyDeleteBasically, what it seems like to me is that this works... and works very well (e.g., USP results). Since I lack the background knowledge on most of this stuff, my main question is why doesn't everyone use this? Secondly, why does it do so well comparatively? I feel like these questions are really, really basic, and I apologize for those of you who aren't having trouble with this at all, but I would love the answer to be broken down to where I can understand it. :)
I agree with everyone else; who can read this stuff? I'm attracted to the idea of what I think that I understand about MLN. I guess one problem I had with logic was that the rules were so rigid; here we can see that exceptions are permitted, and I really appreciate that. I would like further explanation of how mcmc sampling is used for these networks. they have about a paragraph. As it samples these variables, what about them is being adjusted each time? The markov blanket? The number of neighbors? Who the neighbors are? Since this determines the weights I would like to better understand how. Sometimes, they don't use gibbs sampling (and if the future, plan to improve the 'form of mcmc'). Right now, what are the problems with it? To me it seems as if this is important since they named the whole thing markov logic networks. And since all the papers are a jumble of equations anyhow, I don't know why this was left out. Or maybe i just missed it...
ReplyDeleteTo say these readings were a bit over my head would be an understatement. I'm not even sure if I have a good idea about what the weights tell us (does it reflect the probability that a given logical statement is true?). There are a couple of things that were talked about that I would like to be able to understand a bit better, including how to deal with inconsistencies in the knowledge base and how the unsupervised parsing works.
ReplyDelete"So I would like to ask Benjamin how would he try to model the human mind in the following task: suppose someone has to solve the following physics problem:
ReplyDelete"Suppose a car is able to accelerate at two meters per second. What acceleration can the car attain if it is towing another car of equal mass? Explain why.""
Well Cristian, that is not a very ecological example... people are probably not born knowing how to solve this problem which means they need to learn the formula. Learning is pattern completion and association but not necessarily logic.
Obviously we can learn how to do logic with extensive training and with the aid of tools like writing etc, but that isn't what we are talking about now is it, Cristian? The question is whether or not the mind uses this process in its normal functioning.
Maybe a better example is, "I am pulling a wagon up a hill at a certain speed when someone throws a heavy object in it. How much do I need to compensate to get the same speed?" These are the kinds of physics problems that we face in the real world regardless of our education etc.
An embodied answer would probably work better than a logical one.
You kind of lost me with the last part...