Wired Minds
Conversations on cognitive neuroscience from the CSLB and friends.
07 February 2013
"Chance favours the connected mind"
20 December 2012
Hierarchies of temporal scales in perception
The brain seems to do a lot of stuff. We know this because cognitive neuroscience continues to probe ever finer into this stuff, and report back voluminously on the details. These details are of course essential, but it's helpful to take a step back every once in a while and think about the big picture.
This is what appeals to me about the recent work of Stefan Kiebel and colleagues (e.g. Kiebel et al. 2008, 2009). These authors offer an interesting and compelling perspective on perception, learning, brain structure, brain dynamics, and the relationship between agents and the world in which they live.
"We...motivate the hypothesis that the environment exhibits temporal structure, which is exploited by the brain to optimise its predictions. This optimisation transcribes temporal structure in the environment into anatomical structure, lending the brain a generic form of structure-function mapping."
Bullet point summary of the central ideas:
- brains / organisms / agents make predictions about the world
- ...and have an inherent tendency to minimize 'surprise' by trying to make correct predictions*
- structure exists in world at multiple temporal scales (millisecods, seconds, minuted, hours, days, etc.)
- this fact is 'exploited by the brain to optimize its predictions
- ...resulting in a 'transcription' of temporal structure in the environment into anatomical structure
- this transcription takes the form of a 'hierarchy' or 'gradient', where lower-level brain regions (closer to sensory systems) represent structure at faster temporal scales, and higher-level regions at represent structure at slower temporal scales
- a natural mathematical description of such a system is a hierarchy of dynamical systems at different time scales with the higher (slower) systems specifying the manifold on which the faster systems unfold
- (the primary 'result' of the paper is a proof-of-principle mathematical model showing how this could work, including a demonstration that the model is able to learn certain structures in bird song )
Figure 1 of Kiebel et al. 2009 |
Figure 4 of Kiebel et al. 2009 |
Initially I wasn't sure how the separation of temporal scales idea fits in with the conventional conception of the visual processing hierarchy, where progressively more abstract types of structure are processed at progressively higher brain regions in the occipito-temporal processing stream - all of which operates (at least in the majority of experimental paradigms!) on static retinal images, with no temporal structure at all. This was cleared up for me in the supporting materials, which contain (somewhat unusually) a lengthy literature review of evidence from different domains showing that the cortex does indeed show the hypothesised rostro-caudal gradient of representational time-scales. In short: properties of higher-order visual representations like orientation and motion invariance, non-retinotopicity, etc. are in general insensitive to the rapidly changing edges and contrasts of immediate sensory input, even if this rapid change is not typically included in experimental paradigms.
I think these ideas make the most sense in the auditory domain, particularly language processing, which is well-understood to contain structure at multiple temporal scales. The authors clearly recognize this, as birdsong is in many ways a simple spoken language model. In a more recent frontiers in neuroinformatics paper that is classed as a review but also describes some new results, the authors develop the model and the general argument in the direction of artificial speech recognition, making firmer contact with an already extant literature on hierarchical models in this area.
I think language researchers will be quite comfortable and familiar with the idea that higher order language regions track temporal dependencies at longer time scales; it's almost a truism. This is certainly how I understand the computational role of LIFG / Broca's region in sentence-level syntax, one of the main research foci at the CSLB. I'm not sure how the Zatorreian idea of right hemisphere regions being more specialized for longer-duration auditory features such as prosody and music fits in with the rostro-caudal gradient idea.
A final thought: perhaps there may be some relevance here to 'PASA' - the 'posterior to anterior shift in ageing (Cabeza, Davis, etc.). This refers to the observation that older adults generally engage more anterior and less posterior areas across a range of cognitive functions. This is generally understood in terms of 'compensation', utilizing more 'cognitive resources', or using different 'cognitive strategies'. I personally find these somewhat unsatisfactory as explanatory constructs. One possibility that could be explored, then, is that ageing brings with it a preference towards utilizing slower-changing aspects of the environment to inform perception, decision making, and action.
One thing is certain: whether or not the Kiebel et al. slow+fast dynamic hierarchies idea holds sway in teh long term, predictive coding (the less jazzy and more generic uncle of Fristonian free energy) is very en vogue at the moment, and I suspect will work itself into the core neuroscience dogma before too long.
Main refs:
21 September 2011
Walter J. Freeman's Neurodynamics and Neurophilosophy of Meaning
I've been interested in Walter Freeman's work for a while now, and having just read his book 'How brains make up their minds' (HBMUTM; note the double entendre) on holiday, I thought I'd have a go at summarizing some of his ideas. This is a pretty tall order, for at least four reasons: there are a lot of them, they are sometimes quite complicated, they are sometimes quite unconventional, and they are sometimes a bit loose. Here goes...
- Freeman's general theory of brain function ('nonlinear neurodynamics') and of how meaningful percepts are formed from sensory stimuli
- The running commentary contrasting three 'modes of intepreting experimental data from the neural and cognitive sciences about the nature of the mind': materialism, cognitivism, and pragmatism.
Most of Freeman's experimental work comes from intracranial EEG (also called 'ECoG' - 'electrocorticogram') recordings from olfactory cortex of rabbits when they detect and respond to conditioned stimuli (i.e. stimuli, normally chemical odourants, that they have been trained to associate with a reward). He considers that this association of a reward to the stimulus makes an otherwise irrelevant sensation (e.g. the smell of burnt wood) 'meaningful' to the animal. Behavioural saliency of this kind appears to be quite an important component of his notion of meaning; in fact at one point he says that the amplitude modulation patterns (see below) that characterize the meaning of an odourant to the animal actually disappear when that odourant is 'unconditioned' through a new training regime, or when it is fed to satiety, which remove the behavioural salience of the stimuli.
Whether one would actually need to condition stimuli in this way to invoke a salience-based response in humans is probably an open question. Perhaps simply following instructions in a visual object processing task would be enough to establish the salience of the stimuli; whereas for rabbits it is necessary just to get them to do the task? Clearly for people working on constructs labelled 'meaning' in other areas such as language or vision, it's very important that this terminological/experimental difference be borne in mind when reading Freeman's less anthropocentric use of the term. Irrespective of whether one agrees with every aspect of his theory of meaning in all its neuroscientific and philosophical intricacies (and I will not go into all of these here), however, I think most neuroscientists interested in brain systems at the meso- or macroscopic scale will find something interesting in this body of work.
- The state transition of an excitatory population from a point attractor with zero activity to a non-zero point attractor with steady-state activity by positive feedback
- The emergence of oscillation through negative feedbck between excitatory and inhibitory neural populations
- The state transition from a point attractor to a limit cycle attractor that regulates steady-state oscillation of a mixed excitatory-inhibitory cortical population
- The genesis of chaos as background activity by combined negative and positive feedback among three or more mixed excitatory-inhibitory populations
- The distributed wave of chaotic dendritic activity that carries a spatial pattern of amplitude modulation made by the local heights of the wave
- The increase in nonlinear feedback gain that is driven by input to a mixed population, which results in construction of an amplitude-modulation pattern as the first step in perception
- The embodiment of meaning in amplitude modulation patterns of neural activity, which are shaped by synaptic interactions that have been modified through learning
- Attenuation of microscopic sensory-driven activity and enhancement of macroscopic amplitude-modulation patterns by divergent-convergent cortical projections underlying solipsism
- The divergence of corollary discharges in preafference followed by multisensory convergence into the entorhinal cortex as the basis for Gestalt formation
- The formation of a sequence of global amplitude-modulation patterns of chaotic activity that integrates and directs the intentional state of an entire hemisphere
The fact that the patterns are 'global', or at least cover large areas, implies some kind of connectivity between the different regions. There is a very revealing section in HBMUTM (pgs. 116-117) where Freeman likens his findings to those of a long list of scientists who have studied / are studying 'global processes underlying the unity of perception and action' - including Karl Pribram, Antonio Damasio, Jack Pettigrew, Paul Nunez, Stuart Hameroff & Roger Penrose, Dietrich Lehmann, Urs Ribary & Rodolfo Llinas, Catherine Tallon-Baudy, Mathias Muller, Francisco Varela, Steve Bressler, Moshe Abeles, and several others. So if like me you have an idea of the typical work attributed to these scientists, then you can get a bit more of an intuition as to what kind of phenomena Freeman thinks he is studying.
Interestingly, however, he does follow this up with the following comment:
According to the Stanford encyclopedia of philosophy:
"The literature distinguishes “philosophy of neuroscience” and “neurophilosophy.” The former concerns foundational issues within the neurosciences. The latter concerns application of neuroscientific concepts to traditional philosophical questions. Exploring various concepts of representation employed in neuroscientific theories is an example of the former. Examining implications of neurological syndromes for the concept of a unified self is an example of the latter."
This tripartite comparison forms a consistent thread that runs throughout the entire book, and is regularly returned to in the light of the particular content being discussed at various points. With respect to the core scientific topics of the book, spatial AM patterns in sensory cortices:
According to the materialist view:
"these AM patterns reflect information processing....an odorant stimulus delivers information to the receptors, which process it by transducing it to action potentials. These pulses are transmitted to the bulb, where the information is bound into patterns and held, while it is being relayed by the tract to the cortex. The information stored in the cortex from previous stimuli is retreived and sent back to the bulb, where a comparison is made by correlation of the newly recieved AM pattern with each of a collection of retrieved AM patterns...The classification process is completed when the best match is found to identify an odourant. That best AM pattern is sent to other parts of the brain, where it serves to select and guide a fixed action pattern as a response to the stimulus."
for cognitivists:
"...each AM pattern represents an odourant. It is a symbol that signifies the presence of a source of food or danger. The receptor action potentials represent the features of the odourant, and the process by which the bulbar action potential are brought into synchrony through their synaptic interactions to represent an odour is feature binding. The integration of the features by a higher-order neuron makes it fire, and its activity represents that object that has the features.
and for pragmatists:
"...the AM patterns are an early stage in the construction of meaning. They correspond to the 'affordances' advanced by J.J. Gibson in ecological psychology, by which an animal 'in-forms' itself as to what to do with or about an odourant, such as whether to eat the food or run from the predator giving the odorant information. They cannot be representations of odorants, because it is impossible to match them either with stimuli or with pulse pattens from receptor activation that convey stimuli to the cortex....They cannot be information, because that is discarded in the spatial integration performed by divergent-convergent pathways...They reveal the wings of attractors that are selected by the sensory pulses, each having a crater in the olfactory attractor landscape.
...In colloquial terms, the ingredients received by brains from their sensory cortices with which to make meanings are produce by the cortices. They are not direct transcriptions or impressions from the environment inside or outside the body. All that brains can known has been synthesized within themselves, in the form of hypotheses about the world and the outcomes of their own tests of the hypotheses..."
(pgs. 91-93)
Signing off,
john
18 August 2011
Help, my brain won't normalise!
- Define a mask that covers voxels with damage
- Carry out warping ignoring voxels in the mask
- Warp calculated on intact voxels is extrapolated over the masked region
- Brett et al. (2001)
- Unified normalisation includes a segmentation step
- Warping is based on the segmented image: only considers tissue of interest
- This may effectively remove the lesion and make cost function masking redundant
- Ashburner & Friston (2005)
- Stricter regularisation penalises unlikely warps
- Makes shrinking of the lesion less likely
Strict | Medium |
10 August 2011
Perception and semantics in the ventral stream
Recognising what a visual object is not only requires that objects are processed visually, but also that the semantic knowledge associated with the object is evoked. As such, a comprehensive account of how we recognise what an object is requires bringing together theories of visual object recognition, and cognitive models of semantic knowledge within the same neurobiological framework. This is the approach we’ve been developing (see Taylor, Moss & Tyler, 2007 for review) - understanding not only the cognitive contributions of different brain regions, but also how meaning emerges across time (e.g. Clarke, Taylor & Tyler, 2011). Uncovering how we understand what we see requires the development of a comprehensive systems-level account of how we get from perceiving an object, to understanding what it is, and requires the synthesis of cognitive theories and neurobiological models - a fundamental component of cognitive neuroscience.
Clarke, A., Taylor, K.I., & Tyler, L.K. (2011). The evolution of meaning: Spatiotemporal dynamics of visual object recognition Journal of Cognitive Neuroscience, 23(8), 1887-1899.
Felleman, D. J., & Van Essen, D. C. (1991). Distributed hierarchical processing in the primate cerebral cortex. Cerebral Cortex, 1, 1-47.
Tanaka, K. (1996). Inferotemporal cortex and object vision. Annual Review of Neuroscience, 19, 109-140.
Taylor, K. I., Moss, H. E., & Tyler, L. K. (2007). The conceptual structure account: a cognitive model of semantic memory and its neural instantiation. In J. Hart & M. Kraut (Eds.), The neural basis of semantic memory (pp. 265-301). Cambridge: Cambridge University Press.
Tsunoda, K., Yamane, Y., Nishizaki, M., & Tanifuji, M. (2001). Complex objects are represented in macaque inferotemporal cortex by the combination of feature columns. Nature Neuroscience, 4(8), 832-838.
22 June 2011
Direct Electrical Stimulation: Insight into brain networks
I have just been going through a thought provoking new article by Hughes Duffau (2011), and thought I would provide a brief synopsis of the methods, and highlight a result or two. For those of you who are not very familiar with Direct Electrical Stimulation (DES) work, this would be one of the better places to start, as it is very well written and the reference section covers nearly the entire literature.
The first strength of the paper lies in the methods. Whereas the vast majority of human neuroimaging (MRI, EEG, MEG) attempts to measure cortical activity through the scalp, skull, and meninges, DES goes straight to the source: In patients undergoing brain surgery (awake patients to be specific), bipolar electrode tips deliver biphasic current pulses to the cortex and white matter of the human brain. While the current is being delivered, a patient’s performance on some task is being monitored. When the patient’s performance changes due to the stimulation, you have successfully mapped an “eloquent region.” In other words, that region’s neuronal activity plays a direct role in producing that behaviour.
Dr. Duffau's results using DES support a view of the brain that is very dynamic, and do not adhere to the localization hypothesis. To put it another way, complex human behaviours such as language rely on many different brain regions operating in concert. If one of the regions is damaged, it is possible to re-organize the circuit and preserve the function (as long as the white matter connections are left intact). Bear in mind that this is in relation to so-called frontal/executive functions (cognitive processing, language, emotion, etc), and basic sensory systems might operate far differently (not surprisingly, Dr. Duffau has published on this as well).
Below are a few references... I am including two other intriguing articles for those who want to know more about DES, and how the brain might be wired. As always, email me with any questions or just follow up with a comment.
Duffau H (2011) The “frontal syndrome” revisited: Lessons from electrostimulation mapping studies. Cortex (In Press).
Mandonnet E, Winler PA, Duffau H (2010) Direct electrical stimulation as an input gate into brain functional networks: principles, advantages, and limitations. Acta Neurochirurgica 152: 185-193.
Mandonnet E, Jbabdi S, Taillandier L, Galanaud D, Benali H, Capelle L, Duffau H (2006) Preoperative estimation of residual volume for WHO grade II glioma resected with intraoperative functional mapping. Neuro-Oncology 9: 63-69.
21 June 2011
DTI preprocessing in FSL: B-vector correction
Using FSL to pre-process DTI data is easy to do, and the use of a standard FMRIB pipeline is ubiquitous in the literature. An ongoing concern with this common pre-processing pipeline is how to appropriately deal with eddy currents. FSL uses a 12 degree of freedom affine registration as a method of eddy current correction. By definition, therefore, it corrects for both eddy currents AND gross subject motion. So far this seems great... however...
It is really important to alter your b-vector file based on subject motion (Leemans and Jones 2009), but one should NEVER correct b-vectors based on eddy currents, since these vary across the brain (randomly) depending on a variety of factors. This can have profound effects on not only tractography, but also general diffusion modelling (DTIfit). Current wisdom on the FSL forum suggests that using the downloadable script “rotbvecs” will solve your b-vector correction issue. The problem is that this will correct for both subject motion AND eddy currents.
Here is a simple solution that will allow you to keep using FSL for preprocessing, and keep you up to date on appropriate b-vector corrections:
1) Create a new eddy_correct script that only uses 6-degrees of freedom, and call it eddy_correct_6dof. To do this, just add “-dof 6” to the following line of code:
${FSLDIR}/bin/flirt –in $i –ref ${output}_ref –nosearch –o $i –paddingsize 1
Then run the rotbvecs script as normal (this changes the bvecs file to reflect subject motion).
If you still want to correct for eddy_currents, simply run the regular eddy_correct script, and don’t touch the b-vector file. This might seem like a time consuming extra step, but probabilistic tractography is incredibly sensitive to the initial quality of the data. If you are looking for more detailed instructions, just send me a quick email (look to http://csl.psychol.cam.ac.uk/people/) for my contact info.
Of course ... one can also simply choose not to use FSL to pre-process the data... some great alternatives are emerging.