Intro to CogSci

Articles for the research paper

Thu, 2 Oct 2008 14:09:26 -0400

* Amnesia: [check www.pubmed.com and textbook]
* Autistic artists: Snyder_Perception_1997.pdf, Snyder_TBA.pdf
* Biological motion: Neri_Nature_1999.pdf, Hill_CurrBio_2001.pdf
* Blindsight (Weiskrantz): Gazzaniga_science_1992.pdf, Attending, Seeing and Knowing in Blindsight.pdf, Weiskrantz_PNAS_1995.pdf
* Chinese Room (Searle): Searle Chinese Room.pdf
* Dementia: [check www.pubmed.com and textbook]
* Echolocation: bat_echolocation.pdf , Moss_CurrOpinNeuroBio_2003.pdf
* ‘Evolving’ software: [check textbook]
* Face detectors (Desimone): Desimone_JNeuro_1984.pdf, Leopold_Nature_2001.pdf
* Inattentional blindness (Simons): O'Regan_1999_Nature.pdf, Simons_Perception_1999.pdf
* Intermodal transfer (Meltzoff): Meltzoff_Nature_1979.pdf
* Machine Intelligence (Turing): Machine Intelligence (Turing).pdf
* Mental rotation (Shepard): Shepard_Science_1971.pdf
* Morality as innate (Pinker): Brain and morality (Pinker).pdf
* Phantom limb (Ramachandran): Ramachandran_ProBioSci_1996.pdf
* Physicalism (the Churchlands): Physcialism (Churchlands).pdf
* Plasticity (Kohler): Kohler_inverting spectacles.pdf
* Qualia (Jackson): Qualia (Jackson).pdf
* Recovered sight (Fine): Fine_NatureNeuro_2003.pdf
* Rewired ferret: Sur_Science_1988.pdf
* Strong AI, the ‘Singularity’, living forever: [see www.kurzweilai.net to get started]
* Synaesthesia (Ramachandran): Ramachandran_ProcRSocLond_2001.pdf
* Wisdom of crowds: [contact me - erik.blaser@umb.edu]
* Flashbulb memories: [check www.pubmed.com and textbook]
* Child prodigies: [check www.pubmed.com and textbook]
* Asperger’s and humor: [check www.pubmed.com]
* Computers composing music: [start with a web search for Kurzweil]
* The ‘critical period’ for language learning: [check www.pubmed.com and textbook]
* God center in the brain: [check www.pubmed.com; one author is Ramachandran]
* Differences in the male and female brains: [check www.pubmed.com]
* Bilingualism and the brain: [check www.pubmed.com]
* Whorf hypothesis (Whorf): [check www.pubmed.com and textbook]
* Evolutionary psychology: [check www.pubmed.com and textbook]

Assignments, handouts, rosters

Mon, 12 Feb 2007 14:43:09 -0500

CogSci_Roster

Assignments:
HW1 (about concepts)
HW2 (AI position statement)
HW3a_determinism.pdf
HW3b_monism.pdf
HW3c_dualism.pdf
HW3d_freewill.pdf
HW4 Science definition.pdf
HW5 Research paper outline.pdf

Readings & Handouts (and copies of in-class exercises)
Hillis (Tinkertoy computer)
AOLiza (chatroom transcript)
Two Heads (New Yorker article about the Churchlands)
Analogy problem
Hungarian room
Robot with rat neurons
another article

What is Cognitive Science?

Wed, 24 Jan 2007 10:54:07 -0500

Cognitive Science is the interdisciplinary study of the brain and mind.

What is ‘mind’? This is best answered by what the mind does: vision, morals, language, problem solving, memory, movement, emotions, consciousness, etc.

Psychology, Philosophy, Computer Science, Linguistics and Neuroscience (among other disciplines) all contribute to this field.

Cognitive Science is not just about humans, but about the minds of animals, as well as other intelligent systems, such as computers.

Representations

Mon, 22 Jan 2007 09:45:47 -0500

Nowadays the mind is regarded as an ‘information processor’, with the best analogy being to a computer. Your computer takes in information, for instance, through the keyboard, represents that information as words and files, allows for manipulation of the information (spell-checking, changing fonts, changing formatting) and finally can produce an output, say, a printed copy of your term paper.

Analogously, the mind takes in information about the outside world (tables, trees, people, colors, sounds, tastes, etc.) through its sensory systems (the eyes, ears, nose, tongue, etc), creates internal, mental representations of these things, and then is able to perform computations (storage, comparison, on these representations. Ultimately, you exhibit some behavior, like movement, language, or an emotional response.

An interesting distinction is between digital and analog representations, like these two different ways to represent time.

Or, for instance, the word “HAPPINESS” is a digital representation of the concept or the feeling of actual happiness. An analog representation might be this :)

There are pros and cons to each. An analog representation may have an isomorphism with its referent. This can be useful, for instance, if you’d like to remember if a tree is narrower on the top or on the bottom, you can scan your analog mental representation, which looks kinda like a tree, and figure out the answer. The digital representation of tree, namely “TREE” is useful because you can use it in language. You can say:
“Jim loves to climb TREES”, but it is awkward to say “Jim loves to climb ’s”

In other words, analog representations like images are useful because they bear some surface resemblance to their referent. However, if you are setting up a system like mathematics or language where you would like to have a syntax, that is, rules for changing and combining symbols, then it is more straightforward to use digital representations. Also, it is difficult to come up with analog representations for verbs and for abstract concepts: what is a picture of fly? bounce? justice? hunger? However, digital representations are trivially easy, because they don’t have to have any isomorphism, so we can just use the words; FLY, BOUNCE, JUSTICE, and HUNGER.

Note that for this type of representation to work, it has to be learned - so ‘Fa’ might not mean anything to you - it represents nothing useful - but ‘Tree’ does - because you speed English, but not Hungarian.

A representation is something that ‘stands’ for something else (the referent). A representation is symbolic. This has advantages: it is easier for many systems (the brain, computers) to deal with representations rather than the ‘real’ physical objects and concepts out ‘in the world’. Think about if you went on a vacation to Cancun. How would you tell your family what your vacation was like? One way would be to import all the sand and palm trees and Mayan ruins to Boston. Another way would be to fly your family out there on your dime. Of course, what you do instead is to take pictures or videos or keep a travel blog - or maybe all three. These things are representations. The picture of the sunset is not a real sunset, but an analog representation of one. The word ‘Tequila’ is a ‘digital’ representation of a delicious beverage.

Tri-level hypothesis

Sat, 20 Jan 2007 22:36:51 -0500

The Tri-level hypothesis, laid out by Marr (Chap_1) in 1980, breaks down the ‘Information processing’ (of the mind/brain and computers, and other devices) into three independent levels: Computational, Algorithmic, Implementation. These can be roughly understood as task, software and hardware levels, in computer speak. Each of these levels can be studied by itself, or changed, without changing the rest of the system.

For instance, there is no reason you could not take a computer, and change it’s Implementation level, for instance building it out of wood and plastic instead of metal. Daniel Hillis built a computer completely out of Tinker Toys that plays tic-tac-toe.

It remains an open question as to whether the brain is just a computer, but instead of being built out of metal and plastic, it is built out of flesh. This is one of the central questions of Artificial Intelligence. Wittgenstein reportedly, in answer to the question of whether it is possible to build a machine that can think, said yes, we have one in our heads. In other words, in his opinion there was no qualitative difference between brains and machines; and, given that, there is no reason to attribute intelligence, or ‘thinking’ to one and not the other.

To get a more concrete idea of what these three levels mean, let’s consider the following example.

Brain Calculator Abacus

Information
processor

Computational Addition Addition Addition
Level

Algorithmic manipulation of manipulation of binary manipulation of beads
Level arabic numerals code of electrical signals as decimal placeholders
and neural activity

Implementation
Level blood, neurons metal, plastic wood and metal
& electricity & electricity

The first, computational level, describes what the information processor is meant to do. Addition, Subtraction, ‘Find the red toy in the scene’, ‘Determine which object is larger’, whatever. At the algorithmic level, we describe the process by which - the actual computational steps - the device will make its computation. Algorithms act on the form - not the content - of the symbolic representations. The arabic numerals (‘1’, ‘2’, ‘3’, etc) are symbolic representations, so are the beads of the abacus and the bits inside the calculator (symbolic representations of what? Of quantity- quantity of anything.). At the Implementation level, we are interested in what the machine is made out of.

These three levels are independent, theoretically. That is, in principle, we could implement an algorithm in any material we choose, for instance (think about the tinker-toy computer, or the water valve computer). That is not to say that it is easy or natural or technically possible to build, for instance, a computer out of anything; the famous “I can’t believe it’s not butter” margarine-based PC of the late 1980’s comes to mind. Also, too, it may be awkward to use your abacus to play Doom.

The mind-body problem: Dualism vs. Monism

Thu, 18 Jan 2007 11:00:21 -0500

Very early in philosophy, going back at least to Plato, it has been controversial how to characterize the relationship between the mind and the body. Does the mind control the body or the body control the mind? Is the mind nothing more than the body? Might the body itself be a product of the mind? This is the so-called ‘mind-body’ problem. Nowadays it is discussed more specifically: What is the relationship between the mind (one’s thoughts, memories, perceptions) and the brain (the actual bunch of neurons in your head)? This problem still persists today, and may be the deepest philosophical question facing Cognitive Science.

Here is an article that will stimulate your thinking on these issues. Here’s an assignment to provide further incentive.

Freedom of choice? Free-will vs. Determinism

Wed, 17 Jan 2007 18:42:57 -0500

Along with the mind-body problem, the only other question that vies for ‘deepest question of CogSci’ is whether humans, or conscious entities in general, have freewill, or whether our actions (including not just outward behavior but also thoughts and feelings) are completely determined, just as the paths of billiard balls on a pool table can be considered completely determined. This is the so-called Free-will vs. Determinism debate.

These two ‘deep’ questions are not completely unrelated. One’s stance with respect to one question might influence or constrain one’s answer to the other. For instance, it would be awkward to be a Materialist/Physicalist without also being a determinist. After all, everything we know about physical systems leads us to believe that they are deterministic. Even if we allow for true ‘randomness’ into physics, say under some interpretation of quantum mechanics, this is very, very different from saying that a mind has conscious control over what the brain and body do. Is saying that everything you do is random any more satisfying than saying it is completely pre-determined?

If you think that you have free-will, what is the force that guides the physical processes of the brain and body? Can this force be measured? If it interacts with the physical mustn’t it too be physical? But then isn’t it subject to the same laws of physics, which are deterministic, or at best at some levels random? It is certain that we have the feeling of free-will, but could it all just be an illusion? But then, what is the point of such an illusion?

How would you begin a ‘science’ of the mind? Structuralism & Gestalt

Tue, 16 Jan 2007 10:08:05 -0500

The late 19th and early 20th century were full of breakthroughs and optimism for the ‘hard’ sciences, especially physics and chemistry. In fact, at the time, there were scientists who all but thought that these fields were ‘done’, that is, that they had discovered all that there was to know about the world and the universe, from a physical and chemical point of view.

Medicine and Psychology were lagging behind, and many mysteries still abounded. It was around this time that a group of German scientists decided to approach the study of the mind using chemistry as their best analogy for how the mind worked (note that nowadays, we use Computer Science techniques and the computer as our best analogy for how the mind works, and before that electrical engineering and before that mechanical devices and before that, natural forms like plants, or clay). Chemistry had much success in attempting to explain all the matter in the world as mixtures of a finite set of elements. In fact, most of these elements were known, and were arrayed in a periodic table of the elements by a Russian chemist named Mendeleev. These German psychologists, led by Wilhelm Wundt in about 1879, founded the first experimental psychology laboratory. There approach was straightforward, just as matter is constructed by combining together particular elements, the matter of the mind - visual objects in particular - are constructed out of elementary, basic perceptions. There goal was nothing less than a periodic table of mental elements. This approach was called Voluntarism. Wundt’s position was that the mind, through the force of the will, combined the elements of perception into the objects that we see around us: tables, chairs, people, apples, trees. A specific shade of red + a type of sweetness and tartness + a roundish shape + a particular size and weight + plus a smooth, firm surface = APPLE.

Later around 1896, Titchener further formalized the study of the psychology of perception by introducing more rigor into introspection (only very well trained observers, and new perceptual elements were only allowed if there was sufficient consensus among observers). He also removed the necessary action of the will, supposing instead that mental elements are joined by more automatic, ‘mechanical’ processes. He further formalized the way in which an observer’s perceptions were recorded. This field was called Structuralism. Structuralism, before it was abandoned as a field of study, had accumulated over 44,000 sensation elements.

In parallel with this, Fechner was developing the field of Psychophysics. Of all the disciplines from this era that dealt with the ‘scientific’ exploration of mind and perception, this is the one that dominates to this day. Please see the lecture notes on Psychophysics.

The rise of Psychophysics

Mon, 15 Jan 2007 10:36:44 -0500

CLICK ME

Functionalism and 'emergent properties'

Sun, 14 Jan 2007 15:15:39 -0500

Functionalism tries to avoid the whole dualism/monism debate by saying that a mind is the result of the execution of certain processes or functions.

The mind is considered to be an ‘emergent property’ of the brain. (Emergence occurs when the global properties of a system arise from more than the local interaction of its parts. For instance, the substance ‘water’, and its features like liquidity and transparency, emerges from the interaction of H2O molecules, but is not a feature of any individual water molecule.) There is no reason, notice, under this view, why a mind might also not be an emergent property of an alien’s brain, or a computer’s hardware.

‘Digestion’, for instance, is an emergent property of the workings of the stomach. It would be embarrassing to be a dualist about digestion: to propose a physical world where the stomach lives and a ‘digestive’ world where digestion happens.

Computational modeling

Fri, 12 Jan 2007 10:46:18 -0500

To be able to claim - nowadays - that you, as a Cognitive Scientist, have fully understood some process or ability of the brain (motion detection, object recognition, memory, navigation, language comprehension, stereoscopic depth perception, whatever), requires that you be able to prove it. To prove it, means to be able to build it. The most convincing proof in this regard would be to build a robot or computer system that accomplishes the same feat as the brain/mind process in question. The other alternative is not to get bogged down in robot hardware considerations and instead do your building at a more abstract level, instantiating your proposed algorithm in a ‘computational model’.

Such a model should be motivated by results from experiments, and should be tested against the results of subsequent experiments. Once a model is built, one can look too for predictions that it makes, and then devise experiments on humans to test those predictions. The edge enhancement provided by models of retinal ganglion cell receptive fields and the motion detection of Reichardt detectors are great examples of computational models in human vision.

Here are two computational models for motion detection; the left is a ‘Half’ Reichardt, the right is a full ‘Reichardt’ motion detector. These models, similar in spirit to electrical engineering wiring diagrams that you might find for your toaster or thermostat, are sufficiently detailed to fully capture basic motion perception, while also being sufficiently abstract that one could instantiate them in various ways (with wires and photocells for the motion detector that turns on your porch lights, or with neurons in the case of a fly’s or human’s visual system).

Here’s a demo of Reichardt motion detection to help illustrate the action of the computational model pictured above.
http://neurovision.berkeley.edu/Demonstrations/matthew/reichardt.html

and here’s one for ‘Receptive Fields’:
http://www.sinauer.com/wolfe/chap2/ganglionF.htm

Here’s an even stronger version of the Hermann Grid. The original, classic version is on the right.

Artificial Intelligence: The Turing Test

Wed, 10 Jan 2007 11:28:22 -0500

We have spoken a lot about the fact that the same computation can be carried out by many different algorithms and indeed in very different hardware implementations. In our in-class example, for instance, we saw how much of mathematics (addition, subtraction, multiplication, etc) could be carried out by very different methods (sliding beads, pressing buttons, manipulating arabic numerals, exchanging electricity between neurons) and on very different hardware (abacus, slide rule, calculator, brain).

So, now, what of other computations? What about the ‘computations’ the underlie human navigation? We saw that the DARPA competition robots are getting closer and closer. What about the computations underlying great Chess play? Well, the algorithms may be different, but a computer can now fairly reliably beat the world’s best players. It seems in many arenas, we are able to carry out the same computation that the mind/brain can do, but implement it, for instance, on a computer. Can we continue chipping away in this fashion, making computers and robots that are able to perform more and more of the same computations as people?

And what about other, tougher computations that lie closer to the heart of what makes a human a human? What about human language comprehension and production? What about human problem-solving and ‘thinking’, broadly construed? Will these things ever be implemented on a computer? And, if they were, should the computer then be considered ‘intelligent’? Further still, would the machine be conscious, that is, self-aware? Would it have feelings, like pain, or hunger, or perceptions like ‘red’ or ‘sweet’? Could we make it aware? How could we ever know if it were? And, if we can’t know for a computer, or we would be unwilling to attribute intelligence to something that is just composed of a bunch of wires moving around little bits of electricity, then shouldn’t we apply that same skepticism to humans? After all, I can’t prove that you are ‘conscious’, and, our brains are - as far as science knows - also just a bunch of wires moving bits of electricity around. These issues are at the heart of Artificial Intelligence (AI) (AI though can be split into two movements. Proponents of ‘strong AI’ believe that a properly programmed computer can be self-aware and that such a computer is a mind; ‘weak AI’ adherents make no such claims, but allow that a properly programmed computer may be able to solve all the same computational problems as the human brain, but will not be conscious, self-aware, while doing so.)

Memory

Mon, 8 Jan 2007 14:13:32 -0500

Nearly every computation requires some memory, that is, a storage of information that can later be recalled, perhaps to compare to more recent information. In higher-level tasks, you use your memory all the time: you remember what your car looks like, you remember that you have a dentist’s appointment next wednesday, you juggle numbers in your head when calculating the tip at a restaurant, you remember an image from a music video, even though it was only flashed on the screen for a fraction of a second.

Nearly everything that you do would be impossible without such memory (and a whole host of other ‘memory’ systems that you are not even consciously aware of; for instance, the Reichardt motion detector we discussed before uses a ‘delay’ to store information about the output of one of its light detectors in order to compare with the current output of another light detector. Think too of the complex motor movements that you have to make while walking and driving that require that the past and current positions of limbs are remembered, but not necessarily consciously, especially after these skills are well-practiced).

Memory has been broken-down into various types - iconic/sensory memory, short-term/working memory, and long-term memory - based roughly on how long the memory lasts if you are not actively refreshing the information: long-term memories, like your name and what your car looks like, can last a lifetime, while the figures dancing in your head to calculate the restaurant tip will be gone before you finish you finish your breath mint.

Iconic/Sensory - this is a very short-lived (milliseconds), and high-capacity store of information. Imagine an image flashed briefly on a screen - even though your memory seconds later might be poor, there was a very short period right after the exposure when much more information from the image was stored. In other words, a lot of information is initially coded, but it ‘fades’ or ‘decays’ (you forget it) quickly. If you glance quickly at this array of letters, then look away, you will feel the fading occurring.

Working memory (which is a short-term memory, but involves executive processes; in other words the information is being used or acted upon). This has limited capacity and limited duration. In perhaps the most famous experiment in all of cognitive psychology, George Miller estimated that it is possible to keep 7 +/- 2 (seven, plus or minus two) items (or ‘chunks’ of items; for example, KFC is a chunk of three letters, and will be remembered as a unit as opposed to, say, CFK) in your working memory. Information in this storage will decay quickly, in seconds, unless you continuously ‘rehearse’ the information, as you do when you repeat over and over again a person’s phone number until you can get to a piece of paper.

Long-term memory- is for all practical purposes of unlimited capacity (no one has ‘run out of memory’ that we know of), and comes in two forms: implicit, procedural memory (e.g., riding a bicycle) and explicit, declarative memory (e.g., Napoleon was born in 1769).

All that said, however, your memory is probably much worse (less information) and fallible (‘false memories’) than you think. What did you wear two days ago? Which way does Lincoln face on the penny? Is this one correct?:

Visual imagery

Sat, 6 Jan 2007 12:02:53 -0500

We often conjure up visual images, sometimes as an aid to memory, recognition, or problem solving. These images (visual representations) feel like actual pictures in our heads, so to speak, but how literally should we take this impression? When observers are asked to use these images to solve problems, for instance, will they behave as if they were manipulating actual pictures in their heads? If you believe this, you are a pictorialist. The problems shown to the right come from experiments that are taken as support for this point of view.

However, just because you have visual representations, and just because they may be activated while you are solving a problem like those shown here, doesn’t necessarily mean that you are using the visual representation to solve the problem. The alternative is that you are using a digital, as opposed to this visual analog, representation, and doing symbol manipulation, based on some syntax, just like a computer. When you do a math problem in your head, for example, you may picture the numbers, but what you are really doing has little to do with this visual representation, but instead you are manipulating the symbols for numbers, and basing your manipulations on the syntax for, say, addition. If you believe this, you are a descriptionalist.

Problem solving

Thu, 4 Jan 2007 12:21:45 -0500

Problem solving, broadly speaking, means getting from some starting state to some goal state. This could be crossing the road (getting from the starting state of this side of the road to the goal state of that side of the road), finding a partner (getting from single to hooked-up), solving a crossword puzzle (getting from empty boxes to properly filled-in words). To complicate matters, there are likely to be lots of rules/constraints along the way, like cars, competing suitors and those pesky crossword clues. To solve problems, you can use analogies from previous problems, you can try to manipulate visual analog or symbolic digital representations in your head, you can employ the help of other devices (computers, calculators, paper and pencil), and you can use heuristics (‘rules of thumb’, for instance, ‘when crossing a road it is a good idea to look both ways’, ‘when searching for a partner it is a good idea to shower first’, etc.). It is probably a good idea to set subgoals on your way to your ultimate goal (e.g. get to the median in the center of the road, make eye contact, etc.). You will apply concrete operations to help you progress from your start to goal state.

Let’s take a famous problem that people like to solve, and solve by computer too.
The problem is shown here on the right. Here is the goal: Get all the discs onto peg 3 in the same order. Here are the constraints:
1. You may move only one disc at a time.
2. You may only place a smaller disc on top of a larger one.

Here’s another one, the so-called missionary and cannibal problem:
Three missionaries and three cannibals are on one side of a river and need to cross to the other side. The only means of crossing is a boat, and the boat can only hold two people at a time. Devise a set of moves to reach the following goal: transport all six people across the river. But, bear in mind the following additional constraints: The number of cannibals can never exceed the number of missionaries in any location (for the obvious reason) and the boat has to have a rower in order to move.

Keep track of how you are solving the problem. What kinds of representations are you using? What is your starting state? goal state? subgoals? heuristics? operators? What does the problem space look like? How did you accomplish your means-end analysis (that is, how could you evaluate if you were getting closer to your goal or not)?

And here’s another:
http://tinyurl.com/3cgxzn

Intentionality and anthropomorphisms

Tue, 2 Jan 2007 11:09:04 -0500

We have discussed the likelihood of ‘Strong AI’ - whether it will be possible to create machines that are conscious. We have also discussed though that determining whether something is conscious or not seems to be impossible: there is no way even for you to prove to me that you are conscious and not just an automata and likewise there is no way for me to prove it to you. We just assume that other humans are conscious, feeling entities. That this is an ‘assumption’ is even clearer when one considers the case of animals, where the attribution of consciousness and awareness and feeling varies from culture to culture, person to person, and depends on the animal in question (chimpanzees vs. ants, e.g.).

The point is that there is something more going on there than just an intellectual determination: this computer passed the Turing Test therefore it is conscious. I think people would have a harder time taking a sledgehammer to a Furby than to a computer that passed the Turing Test. We will attribute consciousness to robots not when they pass an intellectual test, but when they pass an emotional test; we will attribute conscious awareness when our robot can behave like a kitten or a chimp or a person. At that point, we will not be able to help but anthropomorphize, and attribute intentionality to this new creature. This is the insight of Caleb Chung. His robot Pleo’s only purpose is to engage us emotionally, as if it were a pet. For sure in terms of ‘thinking’ and ‘Weak AI’ and processing power, this robot is much, much dumber than, say, Stanley (Blue team) or Sandstorm (Red team) from the Darpa race. Only one will be able to engage our affections.

It turns out it is remarkably easy to get a person to anthropomorphize and to attribute intentionality, just by making things move properly. In the case of Biological motion, shown here at the right, just the proper movement of a few moving dots is sufficient to tell you that this pattern is actually a walking person, a man, of about average height, early 30’s average weight. It is not just that this is what you see, but you can’t help but see this. It is an automatic process. It turns you you have specialized brain areas dedicated to detecting Biological motion.

Similarly, Heider and Simmel (1944) showed how easy it is to compel people to attribute intentionality (that actions or behaviors have a purpose or reason; in this context these reasons are typically characterized as mental states: ‘The cat ran away because it was scared’, for instance). Even simple geometric shapes are viewed as intentional - automatically - if they move properly.

A robot that one will engage as truly a ‘pet’ is easy, inevitable in fact, if one exploits Biological motion and these principles of intentionality attribution.

Robotic life forms: Pleo contest DUE 3/3

Mon, 1 Jan 2007 07:04:32 -0500

Here is the information on the Pleo, and your next HW assignment. I would like you to write the short, 250 word essay to enter the contest (due Sat. 3/3!). Maybe one of you will win the contest and win a Pleo to serve as the mascot of the burgeoning Cognitive Science program at UMass Boston. We’ll even name it after you. :)

Cognitive neuroscience, a brief history

Sat, 30 Dec 2006 12:54:17 -0500

Neuroscience is the study of the nervous system. In cognitive neuroscience, the emphasis is not on algorithms or computational models, but instead ‘wetware’: what are the brain areas, circuits, and anatomy underlying cognitive functions. As an analogy, imagine that some aliens drop off a fantastically complicated robot. Psychophysicists and cognitive psychologists would study the software, how the computer behaves in various experimental situations, its abilities and function; Cognitive neuroscientists would take it apart, like electrical engineers, and study its circuitry. Cognitive neuroscience focuses primarily on the how and the where; the what is typically left for the psychophysicists and cognitive psychologists.

This approach has some appeal as being less abstract than computational approaches, but there is plenty of assumption and simplification that must go on in this field as well. After all, there are approximately 100 billion neurons in the brain, each with as many as 10,000 connections to other neurons. Such complexity far outstrips any current attempts at understanding.

A bit of history
Phineas Gage (1850’s), by virtue of having a metal post shot through his head and frontal lobe, and thereafter becoming something of a jerk, gave strong evidence that even loosely-defined characteristics - like one’s personality - are apparently stored in the tissue of the brain.

There has been dispute over whether different areas of the brain are responsible for different functions. Though the early movements in both directions had fatal flaws (phrenology of Gall in the early 1800’s and the rat-memory experiments of Lashley in the 1950’s, respectively), ‘localization’ has largely won out over ‘mass action’ theories of brain function.

Before microscopes were well enough developed, it had been controversial whether the brain was made up of individual neurons, as had been discovered in other parts of the nervous system, or whether the brain was one single organ, acting as a unit. It wasn’t until the work of Ramon y Cajal (in the late 1800’s) that there was anatomical evidence for individual neurons (like those pictured to the right; to tie back to topics we’ve talked about before, I also included a picture of how an individual interested in computational modeling would model a neuron. This little circuit could be build out of wires, or simulated on a computer. The approach to Artificial Intelligence that uses algorithms that are like neurons is called connectionism).

Further, compelling evidence for the selective role of particular neurons in brain function was provided by the direct electrical stimulation experiments of Penfield in the 1940’s and 50’s. He stimulated various brain regions of patients who were undergoing brain surgery, but were awake and conscious. His discoveries about the relative sensitivity of different parts of the bodies has been made into a ‘homunculus’ sculpture, shown here:

It is now relatively well established that one can conceptualize the brain as having various areas, each specialized for a particular function (motor movement of the left index finger, sound frequency analysis, face detection, biological motion detection, long-term memory storage, long-term planning, etc. etc). However, these areas are not isolated; communication takes many forms: feed-forward (bottom-up), feedback (top-down), horizontal, within-area, and between-area.

Cognitive neuroscience, methods

Thu, 28 Dec 2006 10:36:27 -0500

Things to keep in mind when evaluating the pros and cons of a particular method are cost, invasiveness (are we just talking about taking pictures of your brain, or do we have to get in there with a scalpel?), whether the technique is passive or active (are we just passively observing the brain, or are we actively messing with it?), is the technique functional or anatomical (are we watching the brain do something, or are we just looking at how it is wired up, its anatomy?), and spatial and temporal resolution (how small of an area can this technique deal with? 1cm by 1 cm, meaning millions of neurons, or a single neuron? How small of a time slice can this technique deal with? If a brain area changes what it is doing for just 1/10th of a second, can we catch it?). For instance, Brodmann made purely anatomical (cytoarchitectonics, he called it) distinctions between brain areas; the map he made is shown here to the right. Others prefer to focus on functional distinctions, with the brain divided up based on which part do what. If you do a map even of, say, the visual areas of the monkey you will find that at least 30 different brain areas have access to the ‘retinotopic’ representation of the visual scene as recorded by the monkey’s retina.

In lesion studies an electrode or a scalpel is used to selectively destroy a specific brain area of an animal. It is then observed how this affects behavior, function, or brain organization. Damage due to stroke, head trauma, or other injury or genetic deficit allows for similar work in humans.

In Single-cell recording an electrode is inserted into or adjacent to a neuron. The electrical activity of this single cell is measured. Such measurements are often combined with some stimulation of the organism (say, shining different patterns of light onto the retina) in order to determine functional relationships. In Multiple-unit recording, a larger electrode is used to measure the activity of a group of neurons.

An electroencephalogram (EEG) provides an even broader view of brain action. Electrodes placed on the scalp measure the gross electrical activity of entire brain regions. An EEG recording in response to the presentation of a stimulus is an event-related potential (ERP).

Recent years have seen the introduction of more sophisticated methods.
In Computer Axial Tomography (CAT), X-rays passed through the brain from different perspectives are used to construct 2-D and 3-D images.

With Positron Emission Tomography (PET), radioactively tagged glucose molecules are ingested or injected to measure which brain areas are most active. Since brain areas that are more active require more blood and oxygen, these molecules will cluster there.

In magnetic resonance imaging (MRI) soft tissue structure is measured by the alignment of protons within a powerful magnet. Functional magnetic resonance imaging (fMRI) is a version that shows changes in brain activity over time.

Some very recent, but admitedly crude sounding techniques are direct current (DC) and Transcranial Magnetic Stimulation (TMS). In DC, current is applied to the skull, and therefore the underlying cortex, in order to enervate or scramble signals there. With TMS, a large electromagnet is placed on the skull, transmitting large magnetic field that can effectively (and thankfully, only temporarily) ‘knock-out’ certain brain regions, simulating a brain lesion.

Human rationality, the lack thereof

Tue, 26 Dec 2006 21:10:08 -0500

To be rational means to make decisions based on logic and a fair assessment of available data and outcome probabilities; rational choices and decisions will be optimal in this sense. To be sure, given this definition, humans are far from rational. We get by because for the most part because our lapses in rationality do not greatly affect our chances of survival in the natural environment in which we evolved. However, especially when it comes to making decisions nowadays, we could probably use a dash more rationality. In everyday life our lapses, and those of others, may not even be apparent, but there are several ways to expose them.
In class I’ll discuss the dollar auction, the Wason Selection Task, Linda the bank teller, the power of positive thinking, a new test and a new drug for depression, and pushing people in front of trains.
In addition to certain fallacies and lapses of logic, we have several biases that cloud what might otherwise be rational decision making. The following is a list, from Dr. Roger Frantz, of common biases. Each of these exposes a situation where logical, ‘rational’ decision making breaks down. These are from Dr. Frantz’s book X-Efficiency. Theory, Evidence and Applications. I’ve left some out, and made some notes or edits in brackets [...].
Bias # 1: Acquiescence to Majority Opinion. Studies have shown that even when asked to judge the length of a line, individuals knowingly answer incorrectly in order to agree with the (overwhelming) majority. Individuals also violate their own personal judgments when faced with a presumed representative of community standards. [Asch]
....
Bias # 4: Confirmation Bias. Once [you have made a particular decision, you are more likely to] listen only to those [individuals or sources of information] who agree with [you]...
Bias # 5: Halo Effect. We tend to experience clusters of properties and evaluate something or someone on the basis of the cluster as a whole. The quiet CPA with a wife and three children who attends church regularly with his entire family could never be a child molester. The Halo Effect distorts our ability to correctly evaluate information and hence to make appropriate decisions.
Bias # 6: Ignoring the Odds. We ignore the odds even when they suggest that we are incorrect. This bias is also known as the representative heuristic. (A heuristic is a shortcut device for making decisions.) For example, is Michele a librarian or a salesperson? She is shy and withdrawn, a meek soul, always helpful, has a need for order, and a passion for detail. You probably believe that Michele is a librarian because her personality more adequately fits the description of a librarian. Odds are, you're wrong. Don't worry: approximately 67% of 1000 business executives surveyed made the same error. They, and perhaps you as well forgot to figure the odds: there are approximately 200,000 librarians in the U.S. but 75 times as many salespersons - approximately 14 million. Despite certain personality traits, Michele is more likely to be selling something than she is to be stacking books at the local library.
Bias # 7: The Self-Serving Bias. We tend to attribute desirable outcomes in our life to our own intelligence, effort, or diligence. On the other hand, we blame the external environment for undesirable outcomes. We also overestimate the chance of positive things happening to us (winning the lottery) but underestimate undesirable outcomes happening to us (being in a car accident). Distorting our abilities and the reasons for success and failure cannot help but reduce our decision-making abilities.
....
Bias # 11: Data are Good. We rely on (inadequate or irrelevant) data just because they are data. Psychologists Amos Kahneman and Daniel Tversky (1974) refer to this as an anchoring heuristic. For example, in a study of anchoring, a group of real estate agents were shown a house, were given a talk about the neighborhood and home values, and were told that the asking price was $65,900. The agents’ evaluation of the house placed its value at $66,755, on average. Another group of agents shown the same house and given the same talk was told that the asking price was $83,900. They evaluated the same house at $73,000, on average, or approximately 10% above the average figure given by the other group of agents. We tend to use any data given to us and to "anchor" our judgments with them.

The wisdom of crowds

Sun, 24 Dec 2006 23:33:04 -0500

In the late 1800’s, Francis Galton did an experiment where he asked a crowd of people at a county fair to guess the weight of a recently slaughtered ox. He was sure how the experiment would turn out: all of these random people, most of whom were likely to have no idea about oxes and carcass weights, would yield a horrendously mistaken weight. It turns out he was completely wrong. When averaged together, the crowd estimated the weight of the ox at 1197 pounds. Its actual weight was 1198 pounds. This estimate was more accurate than that of any particular person in the crowd, even those of the couple cattle ranchers on hand.
This touched off a spark in research on how well groups could make judgments. Before this, the idea that a random group of people could be anything but a mob, that it could be ‘smart’ and even outperform a particular expert in a given domain, was crazy. Further research showed this to be anything but crazy, and in fact the basis of many stable systems.
Here are some examples of the wisdom of crowds at work. None of these systems could work without it:
Democracy . The brain . Free markets . Who Wants to Be a Millionaire? . Race Track betting odds . Natural Selection
You’ll note that second one. That’s why the wisdom of crowds is relevant to Cognitive Science. We have 100,000,000,000 or so neurons. That’s a lot. For any particular task or judgment (say, recognizing a face, or listening for your ring-tone, or deciding whether to duck or jump) you don’t rely on a single neuron. Instead, hundreds of thousands, at least, neurons all ‘vote’ - by virtue of their activation level (‘firing rate’ of electrical impulses) - on what they think you should see, hear or do. Their collective action - the wisdom of this crowd of millions of little dumb single neurons - is responsible for the smartest decision maker in the known universe. The same thing that holds true for neurons that make up a particular brain holds true for a group of such brains (i.e. a crowd of people). For tougher decisions - those beyond basic perceptual ones - it is almost always better to have a group judgment (a ‘committee’ if run properly) than a particular person (“I’m the decider”).
There are a couple important things to keep in mind when designing such group judgments:
1. There should be a ‘true’ answer out there: How much does this ox weigh? What should the exchange rate be for Hungarian Forints? How much should bread cost? What price should this company’s stock be? etc.
2. It is also important that though people may share information, that this be done judiciously; it is critical that ultimately people’s decisions are independent and not coerced or biased by those of others. This is not about reaching unanimity, but instead capitalizing on an averaging over everyone’s diverse knowledge. Secret ballots are good for this. The Social Psychology experiments of Asch (around the 1950’s), mentioned in the lecture notes on human rationality make this point clear. In those experiments, people were biased to misjudge the length of a line just because other people (friends of the experimenter) were basically coerced by ‘peer pressure’ into making obviously bad judgments.
3. There must be some straightforward way to ‘average’ together the judgments of the group. This is easy if judgments are quantitative (yes/no, pounds, dollars, etc.).