12/31/2005

Best Paper Titles 2005

The funniest/weirdest paper titles of 2005:

1) "Morphogenesis of the dentate gyrus: what we are learning from mouse mutants." I, for one, welcome our new mouse mutant overlords.

2) "Raise your hand if you think I am attractive: Second and fourth digit ratio as a predictor of self- and other-ratings of attractiveness." Check out these digits, baby.

3) "What does Batman think about SpongeBob? Children's understanding of the fantasy/fantasy distinction." Thanks, I've always wondered that. Also, what's the difference between "fantasy" and "fantasy"?

4) "Sickness absence and sickness attendance—What people with neck or back pain think." Wait... wait! I know! "OUCH!"

5) "Positioning terrorism in management and marketing: Research propositions." That's just evil.

6) “Ugh! That's disgusting!”: Identification of the characteristics of foods underlying rejections based on disgust." How does this one get past IRB? And why would anyone volunteer in the first place?

7) "Alpaca semen characteristics under free and directed mounts during a mating period." I feel sorry for these scientists. And for the alpacas, who must think the scientists are pretty weird.

8) "The Impact of Finding Mucus and Blood at the Tip of Embryo Transfer Catheter on Procedure Outcome." Let me guess.... not good?

Happy New Year !

12/30/2005

Eigenfaces

By using principal component analysis on a sufficiently large set of face pictures (5,000 to 1,000,000), you can break the pictures down into their simplest constituents (aka, their "principal components") which can then be recombined to create any given face within the set with high accuracy (often >95%). The principal components are considered "eigenvectors" of the original high-dimensional set of faces, or often just "eigenfaces" for short. As you can see, they often look pretty creepy.

This approach has proven fruitful in robotic face recognition, and it leads naturally to the question of whether the human brain is doing something similar. The fusiform face area (FFA) is an area of the temporal lobe that selectively processes face information - perhaps it is constantly doing some form of PCA on incoming face-like visual data, such that it determines the minimum number of neurons needed to represent all faces (the number of eigenfaces), and can then recombine these eigenvectors through coordinated firing to represent any particular face.

While it may seem like an outlandish hypothesis, remember that wavelet and Fourier-like transforms can be seen in the visual and the auditory systems. Also consider that such a "PCA module" in the brain might also be used more generally to distinguish between very similar objects: sure enough, the poorly-named "FFA" shows increased activation when bird experts look at birds, and when car experts look at cars. (It appears we are all face experts.)

Things get even more interesting when you look at autism, a disorder where kids show problems in the way they relate to people (to put it generally). Autistics show neural activation in response to faces in areas outside of the FFA, perhaps because their PCA system has become less localized to the FFA. Fitting with this interpretation, the regions activated in autistics differ from patient to patient. This "generalized PCA" theory would complement multiple theories of autism including those that state autistics are very good at systematizing ("extreme male brain theory"), and good at paying attention to detail but have difficulty integrating information ("underconnectivity theory"). Of course, this is pure conjecture and a "generalized PCA" system has not been shown to be at work in autism or aspergers. There are also clearly many other factors, such as differences in the long-range connectivity between brain regions and possible differences in the mirror neuron system.

12/29/2005

Synaesthesia Part II: Language Colors Vision

Richard Ivry's team of researchers at Berkeley have demonstrated a remarkable effect of the laterality of language processing in the human brain.

Because language is mostly processed in the left hemisphere, they hypothesized that light falling on the left side of each retina (which is relayed to the left side of the brain) would be more affected by linguistic cues than light falling on the right side of each retina.

To test this idea, they asked participants to find the square of a different color from a circular arrangement of otherwise identical squares. The square of a different color was either blue or a slightly different shade of green, and was located on the right or left side of the arrangement. If the square was on the right (and hence processed on the left side of the retina), the subjects took longer to identify the slightly-different green square than the blue one; there was no difference in reaction time between blue and green squares when presented on the left. (image © copyright Richard Ivry / PNAS)

Subjects were able to more quickly select the "one that doesn't belong" if its color has a different name than the others, but only if this was presented to the left side of each retina. There was no "color name" advantage when the objects were presented to the right side.

This research shows just one way that we are able to recruit processing power from other cognitive functions to help in a certain task. You would probably see similar influences from spatial processing on mathematical tasks, given that spatial processing is somewhat lateralized to the right hemisphere.

12/28/2005

A Mind of its Own: Wakamaru

This home robot can do all sorts of tasks, but owners say he is sometimes irritating. He interrupts conversations and ignores demands to be quiet. “One day I found him watching TV, which we never anticipated. Then sometimes he would start dancing..."

Via Digg.

Neurogenesis in Kids, Adults, and Silicon

In the Jan '06 issue of Developmental Science, Durston et al. show a "focusing" effect of activation in frontal and precentral gyri, as well as posterior cingulate brain regions in a longitudinal study of response inhibition such that smaller surface area and yet greater signal strength is associated with increasing age. The authors point out that this increasingly focused pattern of neural activity is reflective not only of structural change (neurogenesis, pruning, myelination) but also of a complex interaction between brain structure and experience.

In fact, the brain is remarkably plastic even into adulthood, such that taxi drivers grow larger posterior hippocampi the longer they've been on the job, musical proficiency in pianists and violinists is predictive of motor cortex thickness (Gaser & Schlaug, 2003), and just a few hours of phonetic discrimination training results in significant changes in the distribution and laterality of fMRI activation in perisylvian regions (Golestani & Zatorre, 2004). Neurogenesis has recently been caught in the act via a cutting-edge imaging technique known as two-photon microendoscopy. Even cannabinoids, a class of chemicals which includes the psychoactive ingredients of marijuana, have been shown to cause neurogenesis.

We can observe the same pattern of increasingly "focused" neural acitivity in artificial neural networks; as the networks are trained for specific tasks, hidden units show less diffuse activation. At the same time, some units become increasingly specialized for representing certain aspects of the task, and will thus fire more strongly than they did during the earlier diffuse activations.

This brings up several questions. What developmental mechanism determines whether task-relevant brain regions undergo pruning or neurogenesis? What is different about the situation of taxi-drivers, in which they show posterior hippocampal neurogenesis, and the situation of people trained in phonetic discrimination, in which functional reorganization and not neurogenesis occurs? And are there artificial neural network models which shed light on these questions?

There are supervised learning models of neurogenesis, such as cascade correlation among others, but it's not clear that these algorithms are biologically plausible. Recent work from Johns Hopkins suggests that neurogenesis can be a very counter-intuitive process: GABA has been seen to act as an excitatory neurotransmitter in new neurons while simultaneously serving as an inhibitory transmitter for more mature neurons! Given the confusing nature of what little we know about neurogenesis, it seems that biologically accurate models will be some ways off.

References

1) Gaser, C., & Schlaug, G. (2003). Brain structures differ between musicians and non-musicians. Journal of Neuroscience, 23 (27), 9240–9245.

2) Golestani, N., & Zatorre, R.J. (2004). Learning new sounds of speech: reallocation of neural substrates. Neuroimage, 21 (2), 494–506.

12/22/2005

Tuned and Pruned: Synaesthesia

According to this paper and others, there is evidence that we were all synaesthetes as babies: capable of hearing color and seeing sound. This synaesthesia is thought to be the product of massively interconnected limbic neurons that are relatively unspecialized, and not yet yoked to the slowly-developing cortex.

During the first year of birth, several recursive stages of neural specialization, pruning (neural death), and then neurogenesis serve to turn this baby brain mush into something more child- and eventually adult-like. But in the meantime, there are several curious U-shaped patterns of performance on tasks that require integration between modalities (such as vision and hearing) such that infants are temporarily better than their older counterparts!

For example, 3 to 4 month olds will preferentially look at the one of two adults who is reciting a passage that matches what is heard, presumably through some early integration of auditory and visual processing. Yet they will not show this behavior again until 7 to 8 months, and in the meantime (5 to 6 months) they will simply look randomly between the two adults. Other more traditional synaesthetic tasks include spontaneous habituation to a light that is roughly the same intensity as a sound to which they had previously been habituated, preferential looking towards one of two objects after tactile habituation to only one of them, and imitation tasks.

What is surprising about all of this is that subjectively, the experiences of seeing and hearing seem very different: vision involves the sensations of color and brightness, and audition involves the sensations of loudness and pitch. At a deeper level, however, audition and vision are very similar, both in terms of the type of information processed and the type of information processing that ultimately results in the sensations of vision and sound. Both the visual and the auditory systems rely on mechanisms that perform Fourier-like transformations of sensory data. These mechanisms are composed of components that are “tuned” for sensitivity to specific frequencies of incoming sensory data. Further, the mechanisms themselves are tuned for a specific balance in the time vs. frequency resolution tradeoff (in the case of the auditory system) and the spatial-frequency vs. location tradeoff (in the case of the visual system).

Visual information comes in the form of quanta varying in frequency and the varying location of the photoreceptors which transduce those quanta into neural activity. The brain can subsequently derive perceptions of brightness, hue and saturation from the magnitude, dominant frequency, and spectral frequency distribution of quanta, respectively, at various locations on the retina, perhaps through a biological analogue of Fourier or wavelet transforms.

Similarly, auditory data comes in the form of vibrations varying in frequency, which is ultimately transduced into neural activity at varying locations along the basilar membrane of the inner ear. From this information, the brain is able to construct the distribution and magnitude of these frequencies, which can be subsequently transformed into perceptions of loudness, pitch, and timbre from the magnitude, fundamental frequency, and frequency distributions respectively. It is likely that this too happens through a biological form of Fourier or wavelet transforms.

The components of both systems are also “tuned” for specific inputs. Particular photoreceptors are more sensitive to certain frequencies of light. Likewise, specific areas of the basilar membrane are more sensitive to particular frequencies of vibration. In summary, there are many similarities in the processing of visual and auditory information at the physical and physiological levels of analysis.

Interestingly, there are many similarities between visual and auditory information at a cognitive level of analysis as well. For example, by presenting different information to each eye (retinal disparity) the brain can reconstruct a three dimensional spatial representation of the external world (although there are other methods used as well). Something similar occurs in auditory processing, in that binaural differences in sound onset time can be used to localize the origin of specific sounds (again, other methods are used as well). Both ears and eyes are stereo devices, and the differences between the two are used for orienting and spatial purposes by mechanisms specifically tuned for sensitivity to those differences.

The similar tunings of auditory and visual system components also result in similarities at a cognitive level. In the case of vision, most “visual detail” that we use to identify various objects comes from high spatial frequency information. In the case of audition, the “auditory detail” or “timbre” that we use to identify various sounds and voices comes from the relative amplitude and distribution of upper harmonics, which are also products of higher frequency information.

Lateral suppression can be observed both within the auditory and the visual systems as well. In the case of vision, this takes the form of mach bands, in which adjacent bars of increasing luminance show contrast effects at their edges. In the case of audition, the masking threshold is much lower for those sounds that coincide with the cutoff frequencies of a non-simultaneous noise mask; in other words, it becomes easier to hear the masked tone when it exists near the edges of the mask because of increased perception of contrast. In both cases, there are contrast effects that appear at the “edges” (whether auditory or visual) of stimuli.

There are also similar figure vs. ground phenomena in both visual and auditory stimuli. In vision, Rubin’s illusion shows how a bistable stimulus can appear to have one of two possible figures (face or vase) superimposed on one of two possible backgrounds (white or black). In audition, a similar effect occurs in pulsation threshold masking, when one auditory stimulus appears to be superimposed over another. Even though neither is truly a “background” sound, since both are pulsating, one is perceived as occurring continuously “behind” the other. Both these scenarios exhibit pattern completion or, to use the gestalt phrase, good continuation.

The subjective experience of odd harmonic frequencies also seems similar between auditory and visual stimuli. The presence of relatively strong odd-numbered spatial frequencies harmonics result in visual stimuli with sharp edges, such as bars. The presence of all harmonics, on the other hand, would tend to appear as a more smooth transition. The presence of many odd-numbered audible harmonics results in the characteristic square- or triangle-wave pattern, which can be described as somewhat rough or sharp sounding. In contrast, the presence of all harmonics gives rise to a much smoother sounding sine tone. In both cases, more complete harmonic information is perceived as “smoother.”

Finally, there are similarities between time/frequency trade-offs in audition and location/spatial frequency tradeoffs in vision. For auditory stimuli, short duration necessarily results in a wideband frequency analysis. To get more specific spectral information about a sound, it needs to be played for longer. Similarly, in vision, high spatial-frequency resolution comes at the expense of low resolution for the location of that information. These are balanced in a similar way as in auditory coding.

These similarities between audition and vision in the adult brain pose many questions related to infant synaesthesia. To what extent are these similarities in audition and vision driven by their early integration? What "general purpose" integrative mechanisms exist in early infancy that are eventually superseded by cortical integration between modalities (as adults show)? Is adult synaesthesia caused by limbic pruning failures in infancy, excess cortical neurogenesis during childhood, too little cortical pruning, or all of the above? And finally, how might this early cross-modal synaesthesia help set the stage for higher-level integration of modalities seen in conscious adults?

12/21/2005

Imitation vs Self-awareness: The Mirror Test

One benchmark for "self-awareness" in animals and people (and now robots as well) is whether they will perform self-directed actions when looking in a mirror. When a mark is placed on the forehead of a child, they will only begin to inspect it on their own forehead at the age of 3 or 4. Adult bottlenose dolphins perform similarly in equivalent tests designed for underwater use.

According to this discovery news article, Junichi Takeno and a team of researchers at Meiji University in Japan have observed similar behavior in a robot with a hierarchical neural network. By equipping the bot with a series of LEDs that light differently depending on the internal state of the robot (two red diodes when the robot is performing behavior it considers its own, two green when the robot acknowledges behavior being performed by the other, or one blue when the robot is both recognizing behavior in another robot and imitating it) they noticed that it behaved differently when viewing its mirror image than in the presence of another robot which imitated its actions precisely.

There are numerous problems with the discovery news article (not least of which is the claim that imitation is the best test of consciousness, whatever that means) and there also appear to be even more problems with Takeno's research (for one thing, it's trivial to differentiate self from other if the other imitates you precisely without a blue light) but it is nonetheless an interesting beginning for tests of self-awareness in robotics.

Edit: now that the story has been posted on slashdot, I imagine everyone has already seen this. Nonetheless, here is the original paper.

Reference Organizers

Two free and extremely cool services: Connotea and CiteULike. Using "bookmarklets," either service will automatically extract the citation and abstract of any paper you want and allow you to tag the reference however you like. From there you can browse your own tags, anyone else's tags, or export your references in BibTex or EndNote format. There are also social networking features such as group lists. Here's my list.

Hypnotic Lullabies

A friend asked "about the effects of vestibular stimulation on overall brain wave states (alpha, beta etc)? I'm trying to figure out if certain motions might actually help brain-wave entrainment (and particularly a reduction in brain activity). I know there are some playthings for hyperactive children that involve bouncing, swinging etc, that supposedly help the kid chill out."

It turns out to be a fascinating question.

This article by Talkowski, Redfern, Jennings and Furman in the Journal of Cognitive Neuroscience claims that the "vestibulo-ocular reflex" was "until recently ... considered automatic (i.e., free from the need for cognitive resources). However, patients with vestibular dysfunction often report mental fatigue, disorientation, and an inability to concentrate on cognitive tasks." Further they point to "clinical and empirical evidence suggesting an interaction between higher cognitive processes and the VOR and ocular motor systems. However, this interaction remains poorly understood."

One hypothesis would be that integrating sensory inputs from different modalities expends mental resources, and sure enough, even simple changes in posture can interfere with cognitive tasks: "combining standing and walking balance tasks with mental arithmetic, visuospatial tasks, reaction time (RT) tasks, word recall, and verbal response tasks" has significant effect.

So do these bouncing, swinging toys calm children through mental fatigue or sheer physical exertion?

I think there are reasons to suspect there is also a cognitive as opposed to purely physiological effect. Auditory and vestibular signals are both processed in the superior temporal gyrus. What follows is a a purely anecdotal argument: one can imagine that lullabies and a rocking cradle could stimulate the same area. In the same way, even popular depictions of hypnosis - swinging watch and all - could theoretically entrain dominant EEG's downward via vestibulo-ocular rhythms.

But such analogies are not really necessary. If you accept the connection between "mental fatigue" and lower average EEG rhythm (as in sleep) then the dots connect themselves: excess vestibular stimulation results in fatigue, which ultimately results in lower EEG rhythms overall.

12/19/2005

Falsity Mechanics

If you hide your roomates' keys, you know they'll probably look where they last saw them. This knowledge is called "Theory of Mind" and it involves dissociating what you believe to be true from what other people may believe. For adults, it's incredibly easy.

But for the majority of three year olds, this is unfathomable. They will fail miserably on a test of Theory of Mind (called a false-belief task) while the majority of four year olds will succeed.

What gives?

Leslie, German and Polizzi believe they have an answer. According to their February 05 paper in Cognitive Psychology, correct performance on false belief tasks requires overcoming a prepotent tendency to attribute "true" beliefs as the beliefs of everyone. The system that represents the beliefs and desires of others purportedly comes online at 2 years, and only then can children begin the lengthy process of learning about these "hidden" mental states. Successful performance requires inhibition of the default attribution, which is naturally that other people believe what is true. A generic "selection process" does the inhibiting that ultimately correctly attributes the false beliefs to others.

There are several complications to this theory of mind mechanism (ToMM) and selection process (SP), many of which they address in their paper:

1) If the representational ToMM system comes online at 2 years, why does it take 2 more years before the SP permits children to succeed at the false-belief task?

2) Why is the task made easier when you add the word "first" in front of the prediction question, as in "Where will Sally look first?

3) Why doesn't adding "first" help autistic children in the same way?

4) And why is it more difficult to correctly say where Sally will look than to say where Sally thinks it is, but only in "avoidance" false-belief tasks? (In avoidance versions, the object is aversive and Sally doesn't want to find it, and control tasks suggest kids can understand the statement semantically)

The authors argue that the "look first" comment (#2 above) helps in prediction questions because it makes the the first location more salient and thus more likely to be selected among competing representations by the executive selection process.

Leslie et al also posit that ToMM offers only plausible contents to SP, such that there are never more than two candidate beliefs: the last location of the bait to which the believer had access, and the current location of the bait. Usually the prepotent true belief "pathway" is selected as the belief of others by default - indeed, in most cases, people's beliefs are accurate. In the case of the avoidance task, Leslie et al. invoke the concept of "inhibition of inhibition" to explain question #4 above.

Philip Branning, a grad student in Randy O'Reilly's Computational Cognitive Neuroscience lab, has some concerns about the validity of the tasks described by Leslie et al. Since we're not always agreed on some of these points, we thought it would be a good opportunity to open up the floor for debate. Welcome, and let the games begin!

Mind Games: Humans, Dolphins and Computers

Video games have been scapegoats for ADHD, anti-social behavior, teenage violence, and more. Yet Michael Posner claims that certain video games can have remarkably positive effects on children, based on research guided by neural networks and confirmed with humans.

As little as five days of training with Posner's game can improve working memory measurements of children aged 4 to 6, and increase their nonverbal reasoning scores on an IQ test. Other researchers have had similar ideas, going so far as to reverse common wisdom and actually manufacture video games for kids with ADHD!

But games aren't just for kids - an interdisciplinary project at Michigan State University called "Cognitive Games" has developed "personal cognitive trainers" that exercise attention, memory, language, visual/spatial functions, and executive functions in elderly populations. And an in-press study by Stan Kuczaj and Lauren Highfill shows that even dolphins design and play games, with over 317 distinct variants (including simplified versions for younger dolphins!) Games have also been used to enhance the development of artificial intelligence, from rock paper scissors to chess, Go, and, of course, thermonuclear war.

Despite some recent objections to the view that video games, MTV, rock music and other aspects of popular culture serve to corrupt and handicap youth, very little research has been done on the possible positive effects of these stimuli. Indeed, this lack of emphasis on the positive is somewhat pervasive in psychology, with few exceptions.

What is it about games that is so clearly important for development? And what kinds of games are the most effective at developing intelligence?

The Tyranny of Inhibition

In the children's game called "Simon Says," players are to perform a specific action if and only if the action is preceeded by the phrase "simon says." The traditional and intuitive explanation for Simon Says is that successful players must inhibit responding unless the action is preceeded by "simon says." Similar explanations abound for a variety of tasks, including Stroop, False-Belief, Dimensional Change Card Sort, etc, all involving the use of suppression or inhibition.

However, several recent papers seem to suggest that this kind of "high-level" inhibition (as distinct from low-level concepts like lateral inhibition) may not be necessary to explain any behavior, despite how frequently it is invoked as an explanatory principle in cognitive psychology!

In Egner & Hirsch's '05 paper in Nature Neuroscience, the authors use fMRI and a stroop-like task to demonstrate that inhibition of the fusiform-face area does not occur when faces are distractors in the presence of another stimulus, but that increased activation of this area occurs while attending to a face in the presence of other distractors. According to their logic, good performance in stroop tasks is caused by amplification of relevant information, rather than inhibition of irrelevant or inappropriate information. As Nieuwenhuis and Yeung put it: "a short-circuit in your desktop computer might conceivably cause random blotches of color to appear on your monitor, but you would not infer that the damaged region usually functions as a 'color inhibition' system. Instead, it would be more accurate to conclude that the damaged region—the video card—previously performed detailed, sophisticated computations that resulted in the delivery of meaningful, rather than random, patterns of color to your screen."

A new paper by Vogel, et al. mentions what I find to be a more useful explanatory concept than inhibition: selection efficiency. To use the example given above, while it may be inaccurate to conclude that the system normally performs color inhibition, it would be accurate to conclude that the system normally shows more efficient color selection. Vogel et al. rely on differential sustained EEG amplitudes for various loads on visual working memory, which asymptote at each individual's visual WM capacity, to show that those with higher visual WM capacity are the ones that can most reliably encode only the relevant items in a given display: they have a more efficient selection mechanism than those with lower memory capacity.

Metabolically speaking, these discoveries almost seem obvious: why would the brain evolve to spend resources inhibiting things it doesn't want to pay attention to, when instead it can simply select and amplify those things that should be attended to?

The applications for reconceptualizing working memory and inhibition are profound. Educationally, how can we enhance selection efficiency? Could ADHD be a disorder of selection efficiency, a disorder of the system that directs selection, or both?

12/15/2005

Intelligence Tradeoff

We often think that adults are more intelligent than children, and in some ways, this is certainly true. But in the May 2005 issue of Child Development, Anna Fisher and Vladimir Sloutsky have shown one way in which adults are not smarter than kids.

Fisher and Sloutsky presented 30 pictures of various animals to 5, 7, and 11 year-olds as well as adults. Within each age, there were three groups: baseline, induction, and blocked categorization. The baseline group was asked simply to study the pictures for a subsequent memory test. The induction group was told that all cats have "beta cells" and were then shown the same group of pictures, but asked to specify whether each animal had beta cells or not. The categorization group was a control for the induction group, in which participants were shown a picture of a cat and told it was young; they were then asked to specify whether each of the other animals was young or not from the same group of pictures. Finally, all three groups went through a recognition phase in which they had to identify the pictures that were previously presented from those that had not been presented.

Interestingly, 5 and 7 year olds showed no decrease in recognition accuracy relative to baseline, and yet 11 year olds and adults showed a dramatic decrease. The authors take this as evidence that 11 year olds and adults spontaneously perform category based induction, while younger children perform similarity based induction. A second experiment in which the youngest children were taught to perform category based induction showed the same performance decrement as 11 year olds and adults. The authors also surmise that greater categorization ability may cause decreased recognition of individual items.

So it's possible that children have a better memory for individual items than adults, because they are still forming categorical boundaries between items on the basis of perceptual features.

In this and other studies, the relative "intelligence" of children and adults critically depends on the nature of the test; in fact, both children and adults may manifest different forms of intelligence, with different purposes, and any comparison between the two is one of apples and oranges.

In fact, the purpose of "child intelligence" is just to set the stage for the emergence of adult-like intelligence. If we want to enhance child development, we shouldn't be encouraging them to think like adults: we should be encouraging their child-like cognition, and allow those mechanisms to do what they've evolved to do. In other words, nurture nature.

12/14/2005

Hello world!

This is a new blog dedicated to tracking the development of intelligence in both natural and artificial systems. This means looking at intelligence, what it means, and how it emerges over time in humans, monkeys, dolphins, chatbots, and neural network simulations alike.

Feel free to email paper or topic suggestions to me.