One-Dimensional Dynamics of Attention and Decision Making in LIP

@article{citeulike:2667697, abstract = {Summary Where we allocate our visual spatial attention depends upon a continual competition between internally generated goals and external distractions. Recently it was shown that single neurons in the macaque lateral intraparietal area (LIP) can predict the amount of time a distractor can shift the locus of spatial attention away from a goal. We propose that this remarkable dynamical correspondence between single neurons and attention can be explained by a network model in which generically high-dimensional firing-rate vectors rapidly decay to a single mode. We find direct experimental evidence for this model, not only in the original attentional task, but also in a very different task involving perceptual decision making. These results confirm a theoretical prediction that slowly varying activity patterns are proportional to spontaneous activity, pose constraints on models of persistent activity, and suggest a network mechanism for the emergence of robust behavioral timing from heterogeneous neuronal populations.}, author = {Ganguli, Surya and Bisley, James W. and Roitman, Jamie D. and Shadlen, Michael N. and Goldberg, Michael E. and Miller, Kenneth D. }, citeulike-article-id = {2667697}, comment = {* computational modeling study proposing that LIP population activity behaves with one-dimensional dynamics: a network with generic high-dimensional firing rate vectors, perturbed by transient inputs, rapidly decays to a single mode * makes a very specific theoretical prediction: ''slowly varying'' patterns of LIP activity will be proportional to spontaneous activity * the authors motivate their research by pointing out that single neuron dynamics are heterogenous and noisy, yet the processes underlying attentional shifting and decision-making must function with predictable dynamics * data is based on findings originally reported in 'Bisley and Goldberg 2006':http://www.citeulike.org/user/klouie/article/853140 and based on initial work reported by 'Bisley and Goldberg 2003':http://www.citeulike.org/user/klouie/article/90449 * in the experimental task, either a target for a possible subsequent saccade or a distractor was flashed in the RF of single LIP neurons; when the acitivity from these two different trialtypes were compared, there generally existed a crossing point where the transiently increased activity in distractor-in-RF trials would decay and cross the level of elevated delay activity in target-in-RF trials * the main previous result was that the population crossing time fell within the behavioral time of attentional ambiguity (at other times, attentional advantage/contrast sensitivity enhancement fell either at the target or distractor site) * the motivating data lay in individual neurons, where despite heterogenous transient distractor peaks '''''V''''', transient fall rates '''''k''''', and delay levels '''''D''''', there was a common crossing time for each monkey; this was defined by the equation '''''ln(V/D)~tk''''' * because '''''V''''' and '''''k''''' are presumably mediated by bottom-up mechanisms and '''''D''''' by top-down salience mechanisms underlying saccade planning, the authors suggest that the underlying network mechanism controlling this phenomena can mediate interactions between top-down and bottom-up salience * the authors propose a conceptual model where collective firing rate dynamics are represented by the movement of an ''N''-dimensional firing rate vector through ''N''-dimensional space; the experimental single-neuron results could be explained if any transient activity to a distractor rapidly decays to a energy valley that drives the firing rate vector through the point D before reaching the spontaneous activity point S * with the further assumption that the energy valley floor is straight between D and S, theis model makes several predictions: 1) multineuronal data should be one-dimensional over long timescales (after transients have settled), 2) this one dimension shoudl correspond to spontaneous activity, 3) D (or any point once the network has reached the one-dimensional regime) should simply be a scaled version of activity in S * ''note: seems to me the stronger requirement is that the one-dimensional valley should pass through the origin, which the authors don't mention'' * examining the data, it appears that network activity in both trialtypes is primarily aligned with the vector representing spontaneous activity (S), except during transient activity (target or distractor onset in the respective trialtypes) * accordingly, individual neurons show greater correlation to spontaneous activity during the delay but not the visual transients * using a linear firing rate model (details in the Supplementary Material), the authors show that a network with ''recurrent excitatory feedback'' (sparse, random, net excitatory connectivity) is sufficient to ensure a single slow network decay mode; such a network could drive one-dimensional dynamics in a single local patch with similar RFs * what about neurons from distant, nonconnected patches (which the authors assume distant RF LIP neurons to be)? if the different patches are constructed using the same underlying distribution of connectivity and input strength parameters (i.e. the same statistical but not exact properties) and are of sufficient size, they will have similar '''''V''''', '''''D''''', and '''''k''''' * to demonstrate that these dynamics are generalizable, the authors examine 'Roitman and Shadlen 2002':http://www.citeulike.org/user/klouie/article/89409 data from monkeys performing a reaction time random-dot motion discrimination task and show that ramping activity (Shadlen's integration) occurs largely in the same dimension as spontaneous activity (fixation activity before motion dot onset) * the authors conclude that such one-dimensional firing rate dynamics underlies a consistent attentional switching time in different trials as in 'Bisley and Goldberg 2003':http://www.citeulike.org/user/klouie/article/90449 and ensures that motion integration occurs at the same rate in different neurons as in 'Roitman and Shadlen 2002':http://www.citeulike.org/user/klouie/article/89409 * ''one assumption is that integration rate could be controlled by modulating, for example, the strength of recurrent excitation - stronger recurrent excitation would yield slower timescales'' * the authors make a final point that attractor models of persistent neural activity don't explain the Bisley and Goldberg results because identical stimuli lead to persistent decay activity in one case but not the other; they further suggest that an attractor state may actually arise from coupling between slow modes of multiple interconnected regions (DLPFC, FEF, SC, LIP) * ''note that distractor and target presentation may represent identical bottom-up inputs but different top-down inputs; perhaps a single LIP attractor is driven to one attractor (delay) with this top-down input and a different attractor without''}, doi = {10.1016/j.neuron.2008.01.038}, journal = {Neuron}, keywords = {attention, computational\_model, decisionmaking, lip, monkey, neurophysiology}, month = {April}, number = {1}, pages = {15--25}, posted-at = {2008-04-14 17:54:55}, priority = {2}, title = {One-Dimensional Dynamics of Attention and Decision Making in LIP}, url = {http://dx.doi.org/10.1016/j.neuron.2008.01.038}, volume = {58}, year = {2008} }

TY - JOUR ID - citeulike:2667697 L3 - citeulike-article-id:2667697 TI - One-Dimensional Dynamics of Attention and Decision Making in LIP JF - Neuron VL - 58 IS - 1 SP - 15 EP - 25 N2 - Summary Where we allocate our visual spatial attention depends upon a continual competition between internally generated goals and external distractions. Recently it was shown that single neurons in the macaque lateral intraparietal area (LIP) can predict the amount of time a distractor can shift the locus of spatial attention away from a goal. We propose that this remarkable dynamical correspondence between single neurons and attention can be explained by a network model in which generically high-dimensional firing-rate vectors rapidly decay to a single mode. We find direct experimental evidence for this model, not only in the original attentional task, but also in a very different task involving perceptual decision making. These results confirm a theoretical prediction that slowly varying activity patterns are proportional to spontaneous activity, pose constraints on models of persistent activity, and suggest a network mechanism for the emergence of robust behavioral timing from heterogeneous neuronal populations. KW - attention KW - computational_model KW - decisionmaking KW - lip KW - monkey KW - neurophysiology N1 - * computational modeling study proposing that LIP population activity behaves with one-dimensional dynamics: a network with generic high-dimensional firing rate vectors, perturbed by transient inputs, rapidly decays to a single mode * makes a very specific theoretical prediction: ''slowly varying'' patterns of LIP activity will be proportional to spontaneous activity * the authors motivate their research by pointing out that single neuron dynamics are heterogenous and noisy, yet the processes underlying attentional shifting and decision-making must function with predictable dynamics * data is based on findings originally reported in 'Bisley and Goldberg 2006':http://www.citeulike.org/user/klouie/article/853140 and based on initial work reported by 'Bisley and Goldberg 2003':http://www.citeulike.org/user/klouie/article/90449 * in the experimental task, either a target for a possible subsequent saccade or a distractor was flashed in the RF of single LIP neurons; when the acitivity from these two different trialtypes were compared, there generally existed a crossing point where the transiently increased activity in distractor-in-RF trials would decay and cross the level of elevated delay activity in target-in-RF trials * the main previous result was that the population crossing time fell within the behavioral time of attentional ambiguity (at other times, attentional advantage/contrast sensitivity enhancement fell either at the target or distractor site) * the motivating data lay in individual neurons, where despite heterogenous transient distractor peaks '''''V''''', transient fall rates '''''k''''', and delay levels '''''D''''', there was a common crossing time for each monkey; this was defined by the equation '''''ln(V/D)~tk''''' * because '''''V''''' and '''''k''''' are presumably mediated by bottom-up mechanisms and '''''D''''' by top-down salience mechanisms underlying saccade planning, the authors suggest that the underlying network mechanism controlling this phenomena can mediate interactions between top-down and bottom-up salience * the authors propose a conceptual model where collective firing rate dynamics are represented by the movement of an ''N''-dimensional firing rate vector through ''N''-dimensional space; the experimental single-neuron results could be explained if any transient activity to a distractor rapidly decays to a energy valley that drives the firing rate vector through the point D before reaching the spontaneous activity point S * with the further assumption that the energy valley floor is straight between D and S, theis model makes several predictions: 1) multineuronal data should be one-dimensional over long timescales (after transients have settled), 2) this one dimension shoudl correspond to spontaneous activity, 3) D (or any point once the network has reached the one-dimensional regime) should simply be a scaled version of activity in S * ''note: seems to me the stronger requirement is that the one-dimensional valley should pass through the origin, which the authors don't mention'' * examining the data, it appears that network activity in both trialtypes is primarily aligned with the vector representing spontaneous activity (S), except during transient activity (target or distractor onset in the respective trialtypes) * accordingly, individual neurons show greater correlation to spontaneous activity during the delay but not the visual transients * using a linear firing rate model (details in the Supplementary Material), the authors show that a network with ''recurrent excitatory feedback'' (sparse, random, net excitatory connectivity) is sufficient to ensure a single slow network decay mode; such a network could drive one-dimensional dynamics in a single local patch with similar RFs * what about neurons from distant, nonconnected patches (which the authors assume distant RF LIP neurons to be)? if the different patches are constructed using the same underlying distribution of connectivity and input strength parameters (i.e. the same statistical but not exact properties) and are of sufficient size, they will have similar '''''V''''', '''''D''''', and '''''k''''' * to demonstrate that these dynamics are generalizable, the authors examine 'Roitman and Shadlen 2002':http://www.citeulike.org/user/klouie/article/89409 data from monkeys performing a reaction time random-dot motion discrimination task and show that ramping activity (Shadlen's integration) occurs largely in the same dimension as spontaneous activity (fixation activity before motion dot onset) * the authors conclude that such one-dimensional firing rate dynamics underlies a consistent attentional switching time in different trials as in 'Bisley and Goldberg 2003':http://www.citeulike.org/user/klouie/article/90449 and ensures that motion integration occurs at the same rate in different neurons as in 'Roitman and Shadlen 2002':http://www.citeulike.org/user/klouie/article/89409 * ''one assumption is that integration rate could be controlled by modulating, for example, the strength of recurrent excitation - stronger recurrent excitation would yield slower timescales'' * the authors make a final point that attractor models of persistent neural activity don't explain the Bisley and Goldberg results because identical stimuli lead to persistent decay activity in one case but not the other; they further suggest that an attractor state may actually arise from coupling between slow modes of multiple interconnected regions (DLPFC, FEF, SC, LIP) * ''note that distractor and target presentation may represent identical bottom-up inputs but different top-down inputs; perhaps a single LIP attractor is driven to one attractor (delay) with this top-down input and a different attractor without'' AU - Ganguli, Surya AU - Bisley, James AU - Roitman, Jamie AU - Shadlen, Michael AU - Goldberg, Michael AU - Miller, Kenneth PY - 2008/04/10/ UR - http://dx.doi.org/10.1016/j.neuron.2008.01.038 ER -