Apology for lack of posts

Sorry about the lack of posts, something came up that required my full attention. I should be doing my first Research Blogging post soon and I've got a post on the adaptive significance of humour and a bit of commentary on the methods of evolutionary psychology in the works.

The Science Network #greatwebsite

The Science Network is one of my favourite resources on the Internet. It's one of those gems on the Internet that doesn't get nearly as much publicity as it deserves. Some of the biggest names in science have lectures and interviews on there. All in high quality video that are downloadable so you can watch it in your own time, usually offering a PDF transcript of the video. If you're interested in the recent debates provoked by Sam Harris' book The Moral Landscape the Science Network covered a conference about it here. In this an event, Harris was joined by some big named philosophers and scientists such as Simon Blackburn, Patricia Churchland, Peter Singer and Steven Pinker. There is also coverage on cognitive science conferences and interviews with scientists on their upcoming books and this just scratches the surface. Everyone should take a look

The use and abuse of the normal distribution

"The normal probabilty curve has been so generally abuse in psychological and educational measurement that one has reason to be fearful of criticism from the very start in even mentioning it. The only valid justification for bringing in the probability curve in this connection is that its presence can be experimentally tested. The wrtier has found experimentally that the normal probability curve was not applicable for certain stimuli. In most of the experiments the distributions are reasonably close to normal." Louis L Thurstone in Psychophysical Analysis 1927

I wonder what mistakes we make now, will we still make in 84 years time. I hope not this one!

 

Friday Links #science

Botanical Gardens

Julia Galef on the Social Psychology of Burning Man
http://measureofdoubt.com/2011/09/08/the-social-psychology-of-burning-man/

Jeffrey Shalitt at Recursivity on the value of robot companions
http://recursed.blogspot.com/2011/09/robots-as-companions.html

Jerry Coyne links to a video of Richard Dawkins talking about his new children's book: The Magic of Reality: How We Know What's Really True.
http://whyevolutionistrue.wordpress.com/2011/09/08/dawkinss-new-childrens-book/

I'm South African and was 16 when it happended, but remember watching CNN and seeing the second plane hit shortly after I had got back home from school. The BPS Research Digest has a nice round up of links to do with 9/11
http://bps-research-digest.blogspot.com/2011/09/911-links-round-up.html

Leo Ingwe's piece on Butterflies and Wheels about being a skeptic in Africa
http://www.butterfliesandwheels.org/2011/being-a-skeptic-in-africa/


Vaughan Bell on the lesser known pioneer of the cognitive revolution
http://mindhacks.com/2011/09/04/the-spark-of-the-cognitive-revolution/

Essence Seeking and Aboutness: Creating meaning in the world.

Story time post: gross speculation about stuff I vaguely know

I've been reading Lois Leon Thurstone's original article on his method of psychophysical analysis for Uni and it got me thinking a lot about the relationship between reality and our perception of reality and how this relates to meaning. Psychophysics is the study of how your subjective perception of change relates to actual changes in the outside world, for example a circle projected on a screen could undergo a minute change in size and that change could be below the threshold of what is perceptible to us, this hows that there isn't a one to one correspondence between the object and perception of that object. One of the points Thurstone makes is about not needing to restrict psychophysical analysis to things with magnitude.

It is not necessary to limit psychophysical analysis to stimuli which have intensity or magnitude as their principal attribute. For example a series of handwriting specimens may be arrange in a continuum on the basis of general excellence

It could easily be used it to gain a effective measure on legibility of handwriting, from unreadable to most legible. That is perhaps a trivial example because inability to read would be a good criteria to rate handwriting in terms of legibility thus making it a rather simple discrimination, but what about rating handwriting in terms of elegance. This it would seem would be much harder and there would be greater variation in ratings due to people having different tastes. If asked to explain what elegance is, it would be a more difficult  concept  to define than legibility. Yet you know it when you see it (well some do) it's somewhere in you already, how did it get there, and how come it is so hard to make what it is explicit.

To make this more concrete think of the difficult of programming a computer to recognise elegance. Is there a list of simple axioms that you could give it to look out for when seeking elegance in handwriting.  The free dictionary gives the following definitions of elegance:

a. Refinement, grace, and beauty in movement, appearance, or manners.

b. Tasteful opulence in form, decoration, or presentation.

Could we use these to somehow come up with a list of features that we could program a computer to look for, and judging by the amount of features satisfied rate handwriting in terms of a percentage of those features satisfied. I think this is highly unlikely, one problem is that those features would need an explanation too, and how would the robot recognise them in the first place. It also seems very alien to how humans learn about elegance, no one gives us a list or verbally instructs us. Another way that a computer may be "taught" to recognise elegance in handwriting would be by presenting a neural network with many examples of elegance in handwriting and examples of inelegant handwriting and hoping that it will pick out the features of the former while excluding the features of the latter. This would create a kind of statistical prototype which later instances of handwriting could be judged with in terms of distance from the prototype. Problems also occur with this solution, firstly we haven't been able to programme a computer with a random neural network to learn to recognise elegance in handwriting yet. It seems to require some sort of pre-existing structure that either provides a framework for concepts to be structured around or something that biases the attention of the learning.Without this neural networks require excessive training, and what this entails is usually decided by a human, and therefore does not give us insight into how the human learnt it.  Another problem with the neural network approach is the difficulty of manipulating prototypes as holistic concepts. By this I mean the ability to think of the concept "dog" and the concept "collar" and combine these two as if you were manipulating symbols to get "dog wearing a collar", or being able to chunk the concept of "dog" into a higher more abstract concept "dog parlour". This was one of the strong points of symbolic AI. What this I think points to is some sort of hybrid model of the two and likely other approaches that will still be discovered. Below I sketch out a rough sketch of what I think about right now when I think of a hybrid between the two. It is definitely wrong but I'm trying to be a honest as I can at the risk of showing my ignorance about all of this.


The ability to label prototypes, that is to in someway summarize the spreaded activation of a neural network when the concept of elegance is activated, by making an analogy between a word/symbol and that activation so that now that symbol is analogous to that activation.  This then allows that symbol to be available for further manipulation and can be feedi back into the system that created it  and provide a building block for higher level concepts. This means that there could be many levels of concepts and multiple feedback loops between levels. Connections between things disparate as lines on a piece of paper and fuzzy abstract concepts  can be achieved through higher level connections even if the lower levels don't have anything to do with each other. The neurons that fire when you look at handwriting don't have to directly spread activation to those that fire when elegance is detected they can be connected by indirect multilevel networks that operate at a higher level. Therefore elegance and handwriting do not always trigger each other. Things can be elegant and not handwriting and handwriting can be inelegant.

Something like these large multilevel symbolic networks represent what I think is the best way of explaining aboutness/semantics/meaning. How can a mental state be about something, perhaps by having a sufficiently rich and multifaceted multilevel model of it in the brain. Perhaps this is why a concept such as elegance which is clearly an extremely complex multidimensional concept, can be be so easily used by us.

Rural Village : Okhombe, South Africa

These photos were taken when we were doing research in the area.

Local houses - More tradtional ones on the right

Fires had recently burnt much of the grass leaving the cattle with not much to eat

Soil erosion is a big problem here - Teams of locals build walls to try limit it.

View of the valley

From up quite high in the mountains

Free Stanford Course in Artificial Intelligence: @aiclass

I've signed up for the free Stanford AI Course: Introduction to Artificial Intelligence taught by Sebastian Thrun and Peter Norvig.  I really enjoy being able to access courses like this, and have long been a fan of things like MIT OpenCourseWare and the Khan Academy. Those resources are very good if you have enough discipline to get through the subject by yourself, but I find I can't keep this up for very long and can only handle it in bursts. Knowing that there is no test or time limit doesn't help. With the Stanford course on the other hand there is a set course that run parallel with the actual Stanford course, and video lectures are meant to be watched by a certain date providing incentive to push through those times when you don't fell like doing it.

They offer two options for taking the course online

1. A basic track where you watch all the lectures and do basic quizzes. Not the full course
2. An advance track which is the full course and aspires to be Stanford difficult. Homework assignments are given and exams are taken.*

I would have liked to have taken the advance track but you  need to be available on the days the exams take place and I'm going to be overseas during the final exam and won't be near a computer. So I'm taking the basic track, I will still  learn the basic concepts of AI but without having to do the assignments and exams. I don't know if that's a good thing or not, different parts of me have different opinions about it.

I will definitely be blogging some of my thoughts about the course and the general state of AI.

*You do not receive credit for completing the course, you receive a statement of accomplishment.