Sharon Rose Alterman

I found this week1s topics very interesting. What was most interesting to me was to see the parallels between the models that people are using to model computer reasoning and the early theories of cognitive psychology. In psychology, we do not call it a feature vector, but I have nearly the same diagram in my notes under the topic of sensory processing for both my cognitive psychology class two years ago as I put in my notes for Artificial Intelligence this week.

I am not exactly sure whether I should be happy at noticing this consistency, or not. First off, this sensory-processing theory in Cognitive Psychology has a lot of holes in it, and has been modified many times. Then again, we have been studying human reason much longer than we have been trying to simulate reasoning in computers. Secondly, I think it is very important to keep in mind that neurodes are not neurons. We are not trying to build artificial humans, we are trying to build intelligent agents. Human intelligence is simply the form of intelligence that we know the most about. Thus, I think that it is good that we have been able to function under the same principles that we think that humans may operate on.

On another note, I was very interested to listen to several philosophy students talking together after the talk on Wednesday night. They had gotten something very different out of it than I had. They were complaining about how one of them had 3asked a question about semantics2 and that the lecturer had just 3dodged it by pointing to the syntax and said the rest of it didn1t matter.2 My reaction to hearing this conversation was something along the lines of, of course he did; he1s in computer science and not philosophy. It was just interesting to see how people that are not studying the field react to what experts consider real advances in the field of artificial intelligence.

Rohit Apte

This week, we finished our robot assignment and it gave me immense satisfaction. we had actually programmed our robot to seek shade ­ a concept that before this semester, seemed to be something only people at MIT and Carnegie Mellon could do. Unfortunately, my euphoria was quickly shot down by Stu. The robot he demonstrated (Cassie) was an ongoing project that had taken over three decades of work to design! While Cassie was impressive in that she managed perform simple actions and basic communications, she was far from intelligent.

I am, however, looking forward to the upcoming robot labs. If my robot can show even the most primitive signs of intelligence (maybe the capacity to guide itself around a room without running into any obstacles) I would be satisfied

Durell Bouchard

Brianne Brown

Grace Chou

I was very impressed with the demonstration of the stimulus response program. The robots really seem to be able to understanding some fairly complex sentences. On the other hand, I think what it is able to do is very limited. Like a question being asked about the robot's ability to "learn" vocabularies, the robot can not synthesize knowledge on its own. Although it has a few hundreds words stored in its database, which makes it appears to be intelligent, the robot lack the ability to synthesize something based on what it perceived. I guess that's what computer scientists mean when they design a stimulus-response agent, that is really what you do, no more, no less. The agent is only capable of generate the right response according to a known stimulus and a given feature vector.

Dan Crown

Stuart Shapiro's lecture on Cassie gave us a good look at an actual project in Artificial Intelligence. Up to this point in class we have seen a number of completed projects (the robot that seeks light, etc) and much of the theory behind them and other robots like them. Outside of class I have seen a fair number of AI projects from a strictly outside-in perspective; in other words, I got to see the "finished" product only and didn't take the time to look at the code or theory behind it.

Mr. Shapiro's two day stay with us gave us a good look at the robot as a whole. On Wednesday night, we saw a "finished" product. We were able to get an idea of what Cassie can and cannot do. Class the following day gave us a look inside Cassie. We got to see the actual code behind some of the language processing that Cassie does, much of which was remarkably straightforward and easy to understand. These two sides of the same coin helped me to better visualize what it takes to produce something like Cassie.

The only aspect of the two days I thought was lacking was the fact that Mr. Shapiro didn't talk much about Cassie's learning sessions. I got the feeling that Cassie was simply told what to do. There was very little evidence presented to convince me that Cassie was doing anything more than following a very large set of rules. In my mind, this falls short of one of the major goals of AI - to produce a creature with the ability to learn something on its own.

Renee Findley

Common Language Understanding seems a rather daunting task.

I spent most of my AI ponderings this week considering Common Language Understanding and Cassie the robot. The ponderings weren't especially deep, but I did start considering something Prof. Shapiro spoke about.

He had programmed Cassie so that she could recognize in a statement that if someone typed in 'either a cauliflower is an animal, a mineral, or a vegetable' and later typed in 'cauliflower is a vegetable', Cassie would know that Cauliflower is a vegetable, and not a mineral, and not an animal. He went on to explain that a specific sort of logic sorting had to be designed for this, because although at first glance, and XOR gate would seem to do the job, this was not the case.

So what I began wondering is what about in statements where something is typed in like 'either Rex is a dog, an animal, or a dinosaur' could Cassie know that the categories mentioned above overlap, or would it have to be entered in as though 'Rex is an animal, and he's either a dog or a dinosaur'? (In other words, a statement at the level of a pure XOR gate and not anything more advanced.) And if Cassie couldn't, I wonder how she could be programmed at an advanced enough level to compute overlap.

Scott Goldstein

Maria Hristova

I think last week was very eventful in CS 372. I found the Colloquium and the in-class lecture that Stu Shapiro did very helpful because it answered some of my questions about the field of AI in general. His explanations and answers were very clear and I think helpful for somebody at the introductory level. The discussion about computer vision and the connection between it and computer reasoning was very informative in the sense that it clarified some of the technicalities around it without going into too much details.

I think that it was very helpful that Stu Shapiro showed the code behind how Cassie interacts with the environment around her and the extent to which her communication skills can model human ones. Cassie's "vocabulary" looked very limited but the reasoning that she performed with it was far more impressive than the number of words that she could use. The actual LISP code behind her vocabulary made me think that the task of having Cassie interact with people or other robots freely, without any restrictions of word usage still has a long way to go before achieved. This is not to say that what Stu Shapiro is not significant but it just put the field of AI into perspective of how much is still there to be done.

I was left a little confused about Cassie's ability to learn from previous interactions that she has had. I don't think that it was very clear if she could learn and to what extent her ability to learn would influence her performance. I was not sure if the demonstrations of her abilities were a result of a learning process that she has been through or if the knowledge that she had was "given" to her instead of acquired. I found it very interesting to see how specific the research done with Cassie has been and to realize that putting the hardware version of her probably took many people's research in order to provide what the software simulation of her would do. Thinking that all the people working on the project should have coordinated everything that they do in order to have the final result the way it is makes me think that AI is years from achieving anything that would even closely resemble human-to-human kind of interaction.

Agata Jose-Ivanina

This week I was once again struck by the beauty and power of math. People keep talking about it but it is impossible to explain such things, one needs to witness them to realize what those words mean... I was fascinated by the fact that math can describe what we do, that it gives us the power to model ourselves. Perhaps, those models are not precise and they might never be able to explain who we are and why we are this way, but then they systemize our knowledge, even behavior. This makes math more poetic, not that I ever thought it wasn't, but also it might make us look more prosaic than we are...

Archana Joshee

Wednesday's colloqium with Professor Shapiro was excellent. First of all it cleared my ambigiuty in the connection of logic with natural language. "Regarding language as communication requires consideration of what is said (literally), what is intended, and the relationship between the two." Prof. Shapiro's example of the word 'now' explains how much of what is meant by a word depends on the context of its usage. A system has to know a lot about the world and its surrounding, to be able to use that knowledge (inferentially), and to be able to communicate it as intended.

In one of the web links provided in our class page, I found a link to a software available on the web called START (SynTactic Analysis using Reversible Transformations) The software system is designed to answer questions posed to it in a natural language. It does a very good job while answering questions about the geography, weather or any kind of facts. When I typed in, "Do you know who I am?", it spitted out these weird phrases and some names. I am just amazed by the complexity of our language and how much our brain has been processed in order to be able to understand sentences by inferring it to situations.

Kip Lewis

Stu Shapiro's talk Wednesday night was fascinating. Getting to see some of the research work done by someone who has devoted his career to Artificial Intelligence was a wonderful opportunity. It was mind-boggling to see how much organization and structure had to go into the current version of Cassie. Many of the aspects of conversation that we take for granted, such as keeping track of indexicals and dealing with tenses and time, seemed to be in no way trivial when trying to integrate them into Cassie's knowledge representation system. Cassie's level of complexity behind the scenes as compared to the low level of complexity of her conversations make me wonder how long it will be before she can converse at the level of human beings.

Stu explained briefly how Cassie goes about answering the question "Who are you looking at?". If I understood correctly, she goes through her entire database looking for objects that have the action "look" where Cassie is the agent, and then she reports all of these in chronological order. I'm curious to know if Prof. Shapiro thinks that humans "remember" in this way (i.e. by seaching for an event with a matching description).

I found it fascinating that Stu thinks that humans essentially super-computers, which obviously leads him to conclude that the AI problem is solvable. This is in sharp contrast to my philosophy class, in which it is taken as a given that the mind is separate from the body. Almost all of our discussion is based on this assumption that the soul exists and works on a different plane than the body. I've found it difficult to move from one class to the next (both are on the same day) and to try and switch back and forth between such opposing world views. At least it is making for a very interesting semester.

Creence Lin

I know that the interactive online lesson on Percetrons isn't really part of the readings, class discussion, and lab, but I thought the serendip experience really helped me to understand more of what the web of threshold logic units and weight vectors really stood for. At first it didn't really occur to me that a number like 1 is supposed to represent an elephant. One thing that made me think was when the lesson said that, "There is increasing evidence that things are not quite so simple, that genes, while not determining behavior either, also influence it... Two networks, given exactly the same training experiences, will, in general, come up with different solutions (and hence exhibit different behavior) if their initial starting characteristics (the starting synaptic weights) are different." What if the initial starting characteristics of the two networks are the same and given the same experience? Do they yield the same output? What about for people? Does a certain set of genes and a certain set of experiences (inputs) at specific times yield a unique output? If so, where does free will fit in? If not, where does

Oops! I forgot to finish up and send out my week 3 response. This week made me think about the idea of creating computers that are isomorphic to humans. "Genetic" and "evolution" are words that I normally associate only with things that are alive like humans, but then I find them in phrases like "genetic" programming and "machine" evolution. The thing is though, even if the embodied Cassie eventually has the exact same output as a human, I don't think I will ever treat her like a human being. If, for example, I hit Cassie really hard and she acted like she was hurt, I wouldn't feel bad for her. If I had done damage I might feel badly for Stu or Deepak or other people that created her, but I wouldn't feel bad for Cassie. On the other hand, if I hurt a person, I would feel badly for him/ her. Yes, it's possible that this person might be some sort of Terminator or Robo-cop. But I guess I just assume that the people I interact with are wired like me and have experiences similar to me when they have pain, as opposed to being embodied Cassies dressed as humans and I just can't tell. Of course, how do we even know that we are going about modeling rational agents in the right way? Is my brain really a bunch of meural nets? I don't know if my brain is currently organized in a way that is similar to Cassie's when it comes to semantics. Probably not everyone is wired the exact same way, but we are all similar in the sense that most of us have two eyes, a certain number of bones, and our digestive system works pretty much the same way. So it would make sense that our brains would be similar too. When I listened to the Net Talk thing, I am pretty sure that when I was learning to talk and read that I did not talk like that. Nor have I ever heard a child talk like that, although I do think that babies do babble. Perhaps if we fast forwarded my life while I was learning to talk and read, I would sound like the computer.

Martin Lukac

I didn't get a chance to ask Stu to about the belief revision system. The concept is really cool and and im really interested in the code for it. It seems like the code for the acutal revision part could be very simple, depending on the structures that are used for the initial beliefs. I could imagine that all that would have to be done is just add the new belief to the new structure and flag it or give it a higher priority so its used in interpretation first. What starts to confuse me when i think about it, is what the actual structure of the 'belief' would look like and how its used. I supposed that when the robot see's a certain situation that it recognizes, it could just cross reference the action it was given with what it 'knows' it can do in the situation. The 'knows' being what was programed in that it could do in the certain situation. But, that still doesn't show how the 'belief' is structured and stored.

Over all the code that Stu showed us seemed pretty simple, but taking into account how specific you have to get when writing rules and that you have to break down stuff that nobody knows about and turn it into rules, i can understand why this project has taken the amount of time it has. It seemed to me that they basically were creating their own definition of knowledge and finding the best way to implement it. Also, from the clarity and apparent simpleness of the code, its easy to tell that they did the best possible job they could have (or maybe its just that i've never seen lisp before!).

A comment on the NetTalk tape. Besides making a great addition to a haunted house or a Mark Lord play, I thought it was pretty neat. It got me thinking that if they were able to do something like that useing the nonalgorithmic approach of a neural net, then would it be possible to create a robot or system that combines a neural net and algorithms. has that already been done?

Reshma Menghani

I thought that an interesting point was brought up in the colloquim in respect to Cassie. Cassie is able to perform certain tasks such as following, sleeping, looking , etc. However Cassie doesn't really know what the concept of any of these tasks are, yet she is able to perform them. The question of how intelligent such a project is comes into question.

Another intersting point was brought up in lecture. The average human has 10^10 number of neurons. There is no such program in AI that has this many number. Yet if one were to construct neurodes of this quantity would we be able to mimic the human brain?

Todd Miller

I'll take a break from deep thoughts about neural networks this week and wonder what about Shapiro's talk impressed people so much. Most of it seemed to be rather straightforward adaptions of Lisp language features -- part of the reason the language is the one traditionally chosen for AI work -- and not particularly insightful. For instance, the `new' logical connectives are (or seemd to be) straightforward boolean functions; it's the same issue as recognizing that the English 'or' is almost always the logical XOR. That this doesn't work for multiple or's is hardly suprising to me -- a XOR b XOR c is /not/ how I would've translated it, but instead to XOR(a,b,c). It's a coincidence that A AND B AND C == AND(A,B,C) and A OR B OR C = OR (A,B,C); that it doesn't work for other gates/functions shouldn't suprise anyone. But enough of that rant.

The interesting realization was that coreferential entities can be distinct, thought I'm certain that this isn't original. It seems rather obvious that an intelligent system has to be able to talk about propositions; even first-order logic can do that. The understanding quoted sentences problem would seem to have a rather straightforward solution -- call the analyzer recusively with the 'memory set' of that time. (e.g. rewind 'now' until it's 'then'; probably a good idea to chop off the (new) future, as well). This 'alternate universe' would be maintained between sentences so questions could be asked about it. This also applies to hypothetical scenarios, except 'now' in the hypothetical is ill-defined w.r.t. the 'current' 'now'. (e.g. If you were next to Stu, how many red robots could you see?) The multiply-now facilitly should extend cleanly to handl the multiple 'now's Shapiro talked about (e.g. I'm living in PA now vs I'm typing now vs I'm 21 years old now, etc.)

I noted that distinguishing between perceptually identical things is (obviously!) not a problem for the vision subsystem to deal with; it could, in fact, be AI-complete in the general case. While dealing with explicit feature vectors simplifies the coding of things, it's a bad idea for the general case; it removes the ability of the learning and reasoning system to feed back into the vision system. (Remember that video? We see what we believe.) Object recognition in humans usually (e.g. with low noise) takes place subconciously, but it is controlled by the concious both when necessary (e.g. is that person /really/ far away someone I know) and to learn (e.g. pick up object, name it, remember it). (New patterns can be learned and become subconcious to the extent where SWAT team members can tell if you're a terrorist or a hostage in less than the time it takes to bring their gun all the way around.) This seems like an ability necessary to a complete AI.

Maria Pace

I found Professor Shapiro's talk on "Building a Cognitive Agent" interesting. I think his research on building robots that can communicate based on natural language recognition is certainly a step in the right direction in the field of A.I. Though Cassie stuck to relatively simple conversations, I was impressed with the idea that she keeps track of time, and seems to understand references to "I" and "you". This is a good simulation of the some of basic characteristics of a conscious human. What also struck me was that she admitted when she didn't know the answer-- many humans aren't even up to that level.

However, one obvious aspect of consciousness was missing from Shapiro's robot-- namely the ability to learn. Perhaps AI researchers should program a robot that integrates Cassie's advance language skills with a model of learning such as the neural net.

Heather Palmeter

Well, this week was certainly interesting. Seeing something put into practice is always an interesting addition to any class discussion. While the lecture on Wednesday was a great overview, I found that it was the more in-depth discussion during Thursday's class that left me with a much greater appreciation for the work that was being done in the field.

One of the things that struck me was the relative shortness of the code ­ not that I have any benchmarks against which to compare it. For some reason, I just though the code would be well, that there would be more of it. I'm not certain if that's a common misconception which I fell for or just one that I had but, either way, it was a surprise.

Another thing I began to realize during the talk was the high degree of specialization within the AI field. It became quite apparent when someone tried desperately to get to get an answer to a question about computer learning with little regard for the fact that it was an area of study far removed from the topic of the lecture. That in itself isn't such a startling revelation until you think of everything required to unite any two aspects of AI, or more somewhere later down the line. It wouldn't surprise me at all to find that the search for a way to unite all of the work in AI had become a specialization in it's own right.

Megan Rutter

Recently I wrote a response titled "Can Machines Think?" I have since formed new opinions about machine intelligence, learning, and thinking.

I believe that machines can learn. Since our most recent class lectures and Stu Shapiro's demonstration, I have redirected my thoughts about machine intelligence, and I came to the satisfying conclusion that, while it is hard for me to say that machines can think (maybe I just need to reexamine my definition of "think"), I am perfectly happy stating that machines can learn.

I'll admit that there is a difference between humans and robots, but I don't think it's so much a matter of intelligence. From Stu's lecture, it appears that everything can be boiled down to rules. All one has to do is program these rules and their exceptions into a machine. Undoubtedly the machine will make mistakes, but no human has gone through life completely error-free. Humans make mistakes, they're pointed out and they're fixed. It's the same with machines. If a machine makes a mistake, it is reprogrammed. The present rules are adjusted, and a new rule is programmed. We, as humans, are constantly being told that we are wrong, so we adjust our belief system accordingly. A robot is no different. It will make mistakes, but they are mistakes that can always be corrected. It may be hard to accept. But if this isn't learning, what is? Maybe we just need to adjust our definition of learning. I can see why it is hard to say a robot is thinking. That word has many more humanistic qualities to it, at least to me it does. Maybe the material we've been studying in class is a lot more convincing of machine learning. After all, why would it be called the "Learning Rule" if it weren't learning??? I've said this before, but I think it is just a matter of time before machine intelligence is accepted as intelligence, and robot learning is accepted as a form of learning. Over time our definitions will bend to include the "artificial" forms of intelligence.

Brian Simms

It was cool to meet someone who works in AI on a daily basis. It lends a little more of a concrete feel to the algorithms and ideas we cover in class. I was particularly spooked by the program he ran on the projector where he told his construct to "come here" and the little robot onscreen moved over to his position and said "I came to you, Stu. I am near you." It reminded me a lot of 2001: A Space Odyssey. I could hear Hal's voice emanating from the computer.

I'm a little baffled by the apparent lack of concrete language learning. I would have thought that better algorithms could be developed to make a language generator that sounds normal. Children learn through repetition and continual training, I don't see how that's any different than a computer. Is this a matter of improving the algorithms or is there something innately different in the way a computer learns language? It doesn't seem likely that a computer is incapable of learning to speak normally...

Matthew Spigleman

The two major agents discussed this past week were NetTalk and Cassie. NetTalk used a distinctly bottom up approach while the research embodied in Cassie focuses many of the subtleties involved in language comprehension.

NetTalk is a curious project, its developmental approach to learning appears to be mimicking human speech development, yet it is not taught as a human would be. Babies learn simple words first and then progress to harder ones. The NetTalk agent appears to have been taught from the get go with a fairly complex set of sentences. This contradiction is hard for me to understand as it does not make sense why one would try to teach a human like agent in a non-human manner.

For this reason the approach that Stu Shapiro takes with Cassie appeals to me. My impression is that the majority of the programming done in regards to Cassie's understanding of language deals with the many quirky and difficult aspects of the English language. This approach was very interesting to hear about because I was previously unaware of many of these language features. While Cassie may not be able to carry on a conversation about the weather or politics she is able to dissect the meaning of many complicated sentence structures. In the long run I suppose that this approach has great potential, as it is probably comparatively easy to develop basic conversational skills once the more fundamental issues of language understanding are in place.

Andreas Voellmy

the discussion of neural nets was very stimulating. i'd like to know the extent of the mathematical analysis of neural nets. i am curious if there are rigorous answers as to what different kinds of neural nets are capable of.

while neural nets as non-algorithmic programming tools are very exciting,one limitation they have is that they are still deterministic. after learning, the particular network always responds in the same way and with the same result to a given input. there is no fuzziness of logic built into it. it is possible that indeterminism and fuzzy logic are integral parts of the human consciousness and that this may be one of the major limitations of neural nets.

Nicholas Yee

To me, the most fascinating aspect of using neural nets to find a solution to a problem is that you end up with a black box function. In algorithmic programming, once you solve a problem, you have the code which demonstrates your understanding of the solution. You can share this code with someone else. But with neural nets, you end up with a black box which you can demonstrate is anywhere from 90% to 99.99% accurate, but you have no idea how it goes about doing it. Sure you can sit down and tease apart the weights, but then you can only do this for simple nets which have several layers. This would be impossible to do in a large neural net designed to accomplish complex tasks.

And you end up with a kind of paradox, either you can put a lot of effort into finding an algorithmic solution (which we know doesn't work well in AI-complete problems), or you can let a neural net take care of it, but then you don't know what the solution is.

Back to CS 372 Course Materials | Back to Written Work Main Page