Limits...
Anatomical and functional plasticity in early blind individuals and the mixture of experts architecture.

Bock AS, Fine I - Front Hum Neurosci (2014)

Bottom Line: As described elsewhere in this special issue, recent advances in neuroimaging over the last decade have led to a rapid expansion in our knowledge of anatomical and functional correlations within the normal and abnormal human brain.We discuss how lack of power in group comparisons may provide a potential explanation for why extensive anatomical changes in cortico-cortical connectivity are not observed.Finally we suggest a framework-cortical specialization via hierarchical mixtures of experts-which offers some promise in reconciling a wide range of functional and anatomical data.

View Article: PubMed Central - PubMed

Affiliation: Department of Psychology, University of Washington Seattle, WA, USA.

ABSTRACT
As described elsewhere in this special issue, recent advances in neuroimaging over the last decade have led to a rapid expansion in our knowledge of anatomical and functional correlations within the normal and abnormal human brain. Here, we review how early blindness has been used as a model system for examining the role of visual experience in the development of anatomical connections and functional responses. We discuss how lack of power in group comparisons may provide a potential explanation for why extensive anatomical changes in cortico-cortical connectivity are not observed. Finally we suggest a framework-cortical specialization via hierarchical mixtures of experts-which offers some promise in reconciling a wide range of functional and anatomical data.

No MeSH data available.


Related in: MedlinePlus

A mixture-of-experts architecture. The expert networks compete to learn tasks while the gating network mediates the competition. For every input (x), the gating network receives information about the performance of all of the expert networks (y1,2,3) involved in solving the task, and each expert network”s output is compared with the target output (y). The weights gating the output of each expert network (g1,2,3) are modified based on the relative performance of that expert network (compared to the other experts) for that input pattern. These gating weights not only determine the extent to which the output of each network contributes to the final output, but also modulate learning within each network such that more learning occurs within those expert networks that contribute more heavily to the final output.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4269126&req=5

Figure 4: A mixture-of-experts architecture. The expert networks compete to learn tasks while the gating network mediates the competition. For every input (x), the gating network receives information about the performance of all of the expert networks (y1,2,3) involved in solving the task, and each expert network”s output is compared with the target output (y). The weights gating the output of each expert network (g1,2,3) are modified based on the relative performance of that expert network (compared to the other experts) for that input pattern. These gating weights not only determine the extent to which the output of each network contributes to the final output, but also modulate learning within each network such that more learning occurs within those expert networks that contribute more heavily to the final output.

Mentions: The ME modular architecture describes how a number of “experts” (or modules) compete to learn different tasks. For the purposes of illustration, we describe the classic mixture-of-experts (ME) model described by Jacobs et al. (1991), see Figure 4. This instantiation should simply be considered as an example: many other architectures make similar predictions (Yuksel et al. (2012) for a review). The traditional ME architecture consists of two types of networks: expert networks and a gating network. Expert networks compete to learn tasks, while the gating network mediates the competition. The final output of the full network is a weighted average (with gating weights g1, g2, g3, … gn) of the outputs (y1, y2, y3, … yn) of each expert network. For each input, x, the gating network receives information about the performance of all of the expert networks in solving the task (finding the correct y for that x) and each expert network”s output is compared with the target output. Learning occurs in two ways. First, the weights gating the output of each expert network are modified based on the relative performance of that expert network (compared to the other experts) for that input pattern. This is implemented by forcing the activation of the output units to be nonnegative and sum to 1. Thus, the gating weights determine the extent to which the output of each network contributes to the final output, so that on future trials with similar input the most accurate expert will have a larger influence on the final response than less accurate experts. Learning also occurs within each network: this learning is also modulated by the gating weights such that more learning occurs within those expert networks that contribute more heavily to the final output.


Anatomical and functional plasticity in early blind individuals and the mixture of experts architecture.

Bock AS, Fine I - Front Hum Neurosci (2014)

A mixture-of-experts architecture. The expert networks compete to learn tasks while the gating network mediates the competition. For every input (x), the gating network receives information about the performance of all of the expert networks (y1,2,3) involved in solving the task, and each expert network”s output is compared with the target output (y). The weights gating the output of each expert network (g1,2,3) are modified based on the relative performance of that expert network (compared to the other experts) for that input pattern. These gating weights not only determine the extent to which the output of each network contributes to the final output, but also modulate learning within each network such that more learning occurs within those expert networks that contribute more heavily to the final output.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4269126&req=5

Figure 4: A mixture-of-experts architecture. The expert networks compete to learn tasks while the gating network mediates the competition. For every input (x), the gating network receives information about the performance of all of the expert networks (y1,2,3) involved in solving the task, and each expert network”s output is compared with the target output (y). The weights gating the output of each expert network (g1,2,3) are modified based on the relative performance of that expert network (compared to the other experts) for that input pattern. These gating weights not only determine the extent to which the output of each network contributes to the final output, but also modulate learning within each network such that more learning occurs within those expert networks that contribute more heavily to the final output.
Mentions: The ME modular architecture describes how a number of “experts” (or modules) compete to learn different tasks. For the purposes of illustration, we describe the classic mixture-of-experts (ME) model described by Jacobs et al. (1991), see Figure 4. This instantiation should simply be considered as an example: many other architectures make similar predictions (Yuksel et al. (2012) for a review). The traditional ME architecture consists of two types of networks: expert networks and a gating network. Expert networks compete to learn tasks, while the gating network mediates the competition. The final output of the full network is a weighted average (with gating weights g1, g2, g3, … gn) of the outputs (y1, y2, y3, … yn) of each expert network. For each input, x, the gating network receives information about the performance of all of the expert networks in solving the task (finding the correct y for that x) and each expert network”s output is compared with the target output. Learning occurs in two ways. First, the weights gating the output of each expert network are modified based on the relative performance of that expert network (compared to the other experts) for that input pattern. This is implemented by forcing the activation of the output units to be nonnegative and sum to 1. Thus, the gating weights determine the extent to which the output of each network contributes to the final output, so that on future trials with similar input the most accurate expert will have a larger influence on the final response than less accurate experts. Learning also occurs within each network: this learning is also modulated by the gating weights such that more learning occurs within those expert networks that contribute more heavily to the final output.

Bottom Line: As described elsewhere in this special issue, recent advances in neuroimaging over the last decade have led to a rapid expansion in our knowledge of anatomical and functional correlations within the normal and abnormal human brain.We discuss how lack of power in group comparisons may provide a potential explanation for why extensive anatomical changes in cortico-cortical connectivity are not observed.Finally we suggest a framework-cortical specialization via hierarchical mixtures of experts-which offers some promise in reconciling a wide range of functional and anatomical data.

View Article: PubMed Central - PubMed

Affiliation: Department of Psychology, University of Washington Seattle, WA, USA.

ABSTRACT
As described elsewhere in this special issue, recent advances in neuroimaging over the last decade have led to a rapid expansion in our knowledge of anatomical and functional correlations within the normal and abnormal human brain. Here, we review how early blindness has been used as a model system for examining the role of visual experience in the development of anatomical connections and functional responses. We discuss how lack of power in group comparisons may provide a potential explanation for why extensive anatomical changes in cortico-cortical connectivity are not observed. Finally we suggest a framework-cortical specialization via hierarchical mixtures of experts-which offers some promise in reconciling a wide range of functional and anatomical data.

No MeSH data available.


Related in: MedlinePlus