Limits...
Is predictive coding theory articulated enough to be testable?

Kogo N, Trengove C - Front Comput Neurosci (2015)

View Article: PubMed Central - PubMed

Affiliation: Laboratory of Experimental Psychology, Brain and Cognition, University of Leuven (KU Leuven) Leuven, Belgium.

AUTOMATICALLY GENERATED EXCERPT
Please rate it.

Predictive coding theory (Srinivasan et al., ; Mumford, ; Rao and Ballard, ) claims that the function of the hierarchical organization in the cortex is to reconcile representations and predictions of sensory input at multiple levels... It does this because the dynamics of neural activity is geared toward minimizing the error: the difference between the input representation at each level and the prediction originating from a higher level representation... First, they are fed forward to the higher level(s) where they influence the neural activities of the higher level representation(s)... The resulting predictions are in turn fed back to the lower level... The same applies to recordings of the activity of a population of neurons such as those obtained via fMRI: is an increase of the fMRI signal due to an increased error signal or to changes in the input representation? And does the process of reconciliation between the lower level representation and the prediction result in silencing of error neurons and if so, is this detectable in the data? The last question is particularly crucial because it has been suggested that reduction of neural signals at the lower level can be explained in terms of error minimization (Murray et al., ; Summerfield et al., ; den Ouden et al., ; Alink et al., ; Todorovic et al., ; Kok et al., )... Predictive coding theory is inspired by a systematic pattern of connectivity, both within individual areas of neocortex and within the feedforward and feedback projections between areas, specific to layer location and type of source and target neurons (Maunsell and van Essen, )... Hence, there are four main variables per level, μv, μx, ξv, and ξx. (Each of these variables is multi-dimensional, according to the dimensionality of the input representation at each level.) By analysing the sequential processes in Equation 1 and the known neural types and their connections in neocortex, they pointed out the “remarkable correspondence” between the sequential processes in the equations and the neural architecture... Although they suggested the involvement of inhibitory neurons in L1 earlier, among the diversity of distinct types of inhibitory neurons (Petilla Interneuron Nomenclature Group et al., ) many of them can “provide strong mono-synaptic inhibition to L2/3” (page 699) and there are no clear reasons given why the L1 inhibitory neurons should take the role of reversing the sign of g... Furthermore, they did not explicitly specify the function of the sign reversal by inhibitory neurons in Figure 5... Moreover, they also pointed out that (page 699) “feedback connections can both facilitate and suppress firing in lower hierarchical areas. ” How can this dualistic effect be exhibited by this circuit? Predictive coding theory is currently a highly influential theory for cognitive function and behavior, and one of the plausible theoretical frameworks that may explain the signal processing architecture of the cortex.

No MeSH data available.


(A) Bastos et al. (2012) proposed that neurons in layer 6 represent expectation of cause, μv, and expectation of state, μx, which send out feedback signals to the lower level. In their diagram, this output signal is expressed as a function g (red). (B) When this signal arrives at the lower level, the feedback signal is expressed as −g (red) in their proposal without any explanation of the reversal of the sign. Note that, to compute the error, the subtraction is done between the lower level representation signal, μv, and the prediction factor g, (ξv= μv − g) and, hence, the negative signal of g is necessary. However, if the neurons, μv and μx are pyramidal (excitatory) cells as proposed by Bastos et al. this subtraction cannot be performed. (C) The error, “representation – feedback,” can create either positive or negative values. However, the neuron that represents the error in the proposed circuit of Bastos et al. would not create action potentials when the error value is negative. Hence, the neuron is not capable to signal the error when the prediction factor g is larger than the representation signal, μv. To deal with the positive and negative error signals properly, “two distinct populations of neurons to signal errors, one for positive and another for negative errors” (Rao and Ballard, 1999) may be necessary. For example, the inhibitory neuron, η, shown here reverses the sign of the feedforward representation signal, μ(i), to compute the “negative” error, ξn (=g−μ(i)). The other inhibitory neuron, η, reverses the sign of g so that the “positive” error can be expressed as a neural signal in ξp (=μ(i) − g).
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4561670&req=5

Figure 1: (A) Bastos et al. (2012) proposed that neurons in layer 6 represent expectation of cause, μv, and expectation of state, μx, which send out feedback signals to the lower level. In their diagram, this output signal is expressed as a function g (red). (B) When this signal arrives at the lower level, the feedback signal is expressed as −g (red) in their proposal without any explanation of the reversal of the sign. Note that, to compute the error, the subtraction is done between the lower level representation signal, μv, and the prediction factor g, (ξv= μv − g) and, hence, the negative signal of g is necessary. However, if the neurons, μv and μx are pyramidal (excitatory) cells as proposed by Bastos et al. this subtraction cannot be performed. (C) The error, “representation – feedback,” can create either positive or negative values. However, the neuron that represents the error in the proposed circuit of Bastos et al. would not create action potentials when the error value is negative. Hence, the neuron is not capable to signal the error when the prediction factor g is larger than the representation signal, μv. To deal with the positive and negative error signals properly, “two distinct populations of neurons to signal errors, one for positive and another for negative errors” (Rao and Ballard, 1999) may be necessary. For example, the inhibitory neuron, η, shown here reverses the sign of the feedforward representation signal, μ(i), to compute the “negative” error, ξn (=g−μ(i)). The other inhibitory neuron, η, reverses the sign of g so that the “positive” error can be expressed as a neural signal in ξp (=μ(i) − g).

Mentions: In their model, the feedback signal, g, is sent from the layer 5/6 neurons (μv and μx) at the higher level. These are excitatory cells. It is, then, not clear how the subtraction can be made when this signal reaches the superficial layer at the lower level. Note that while the feedback signal sent from the higher level is g (Figure 1A corresponding to their Figure 5 right; at bottom), when it reaches the top layer at the lower level, it is -g (Figure 1B corresponding to their Figure 5 right; at top) without any explanation of the reversal of the sign. Although they suggested the involvement of inhibitory neurons in L1 earlier, among the diversity of distinct types of inhibitory neurons (Petilla Interneuron Nomenclature Group et al., 2008) many of them can “provide strong mono-synaptic inhibition to L2/3” (page 699) and there are no clear reasons given why the L1 inhibitory neurons should take the role of reversing the sign of g. Furthermore, they did not explicitly specify the function of the sign reversal by inhibitory neurons in Figure 5. Moreover, they also pointed out that (page 699) “feedback connections can both facilitate and suppress firing in lower hierarchical areas.” How can this dualistic effect be exhibited by this circuit? Note that certain formulations of predictive coding have been shown to be functionally equivalent to a biased competition framework (Spratling, 2008) in which the error signal is computed within the upper level rather than at the lower level. Therefore, it may be possible, that with the different mapping of variables to neuronal sub-types, the biologically implausible top-down inhibition for subtraction is avoided.


Is predictive coding theory articulated enough to be testable?

Kogo N, Trengove C - Front Comput Neurosci (2015)

(A) Bastos et al. (2012) proposed that neurons in layer 6 represent expectation of cause, μv, and expectation of state, μx, which send out feedback signals to the lower level. In their diagram, this output signal is expressed as a function g (red). (B) When this signal arrives at the lower level, the feedback signal is expressed as −g (red) in their proposal without any explanation of the reversal of the sign. Note that, to compute the error, the subtraction is done between the lower level representation signal, μv, and the prediction factor g, (ξv= μv − g) and, hence, the negative signal of g is necessary. However, if the neurons, μv and μx are pyramidal (excitatory) cells as proposed by Bastos et al. this subtraction cannot be performed. (C) The error, “representation – feedback,” can create either positive or negative values. However, the neuron that represents the error in the proposed circuit of Bastos et al. would not create action potentials when the error value is negative. Hence, the neuron is not capable to signal the error when the prediction factor g is larger than the representation signal, μv. To deal with the positive and negative error signals properly, “two distinct populations of neurons to signal errors, one for positive and another for negative errors” (Rao and Ballard, 1999) may be necessary. For example, the inhibitory neuron, η, shown here reverses the sign of the feedforward representation signal, μ(i), to compute the “negative” error, ξn (=g−μ(i)). The other inhibitory neuron, η, reverses the sign of g so that the “positive” error can be expressed as a neural signal in ξp (=μ(i) − g).
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4561670&req=5

Figure 1: (A) Bastos et al. (2012) proposed that neurons in layer 6 represent expectation of cause, μv, and expectation of state, μx, which send out feedback signals to the lower level. In their diagram, this output signal is expressed as a function g (red). (B) When this signal arrives at the lower level, the feedback signal is expressed as −g (red) in their proposal without any explanation of the reversal of the sign. Note that, to compute the error, the subtraction is done between the lower level representation signal, μv, and the prediction factor g, (ξv= μv − g) and, hence, the negative signal of g is necessary. However, if the neurons, μv and μx are pyramidal (excitatory) cells as proposed by Bastos et al. this subtraction cannot be performed. (C) The error, “representation – feedback,” can create either positive or negative values. However, the neuron that represents the error in the proposed circuit of Bastos et al. would not create action potentials when the error value is negative. Hence, the neuron is not capable to signal the error when the prediction factor g is larger than the representation signal, μv. To deal with the positive and negative error signals properly, “two distinct populations of neurons to signal errors, one for positive and another for negative errors” (Rao and Ballard, 1999) may be necessary. For example, the inhibitory neuron, η, shown here reverses the sign of the feedforward representation signal, μ(i), to compute the “negative” error, ξn (=g−μ(i)). The other inhibitory neuron, η, reverses the sign of g so that the “positive” error can be expressed as a neural signal in ξp (=μ(i) − g).
Mentions: In their model, the feedback signal, g, is sent from the layer 5/6 neurons (μv and μx) at the higher level. These are excitatory cells. It is, then, not clear how the subtraction can be made when this signal reaches the superficial layer at the lower level. Note that while the feedback signal sent from the higher level is g (Figure 1A corresponding to their Figure 5 right; at bottom), when it reaches the top layer at the lower level, it is -g (Figure 1B corresponding to their Figure 5 right; at top) without any explanation of the reversal of the sign. Although they suggested the involvement of inhibitory neurons in L1 earlier, among the diversity of distinct types of inhibitory neurons (Petilla Interneuron Nomenclature Group et al., 2008) many of them can “provide strong mono-synaptic inhibition to L2/3” (page 699) and there are no clear reasons given why the L1 inhibitory neurons should take the role of reversing the sign of g. Furthermore, they did not explicitly specify the function of the sign reversal by inhibitory neurons in Figure 5. Moreover, they also pointed out that (page 699) “feedback connections can both facilitate and suppress firing in lower hierarchical areas.” How can this dualistic effect be exhibited by this circuit? Note that certain formulations of predictive coding have been shown to be functionally equivalent to a biased competition framework (Spratling, 2008) in which the error signal is computed within the upper level rather than at the lower level. Therefore, it may be possible, that with the different mapping of variables to neuronal sub-types, the biologically implausible top-down inhibition for subtraction is avoided.

View Article: PubMed Central - PubMed

Affiliation: Laboratory of Experimental Psychology, Brain and Cognition, University of Leuven (KU Leuven) Leuven, Belgium.

AUTOMATICALLY GENERATED EXCERPT
Please rate it.

Predictive coding theory (Srinivasan et al., ; Mumford, ; Rao and Ballard, ) claims that the function of the hierarchical organization in the cortex is to reconcile representations and predictions of sensory input at multiple levels... It does this because the dynamics of neural activity is geared toward minimizing the error: the difference between the input representation at each level and the prediction originating from a higher level representation... First, they are fed forward to the higher level(s) where they influence the neural activities of the higher level representation(s)... The resulting predictions are in turn fed back to the lower level... The same applies to recordings of the activity of a population of neurons such as those obtained via fMRI: is an increase of the fMRI signal due to an increased error signal or to changes in the input representation? And does the process of reconciliation between the lower level representation and the prediction result in silencing of error neurons and if so, is this detectable in the data? The last question is particularly crucial because it has been suggested that reduction of neural signals at the lower level can be explained in terms of error minimization (Murray et al., ; Summerfield et al., ; den Ouden et al., ; Alink et al., ; Todorovic et al., ; Kok et al., )... Predictive coding theory is inspired by a systematic pattern of connectivity, both within individual areas of neocortex and within the feedforward and feedback projections between areas, specific to layer location and type of source and target neurons (Maunsell and van Essen, )... Hence, there are four main variables per level, μv, μx, ξv, and ξx. (Each of these variables is multi-dimensional, according to the dimensionality of the input representation at each level.) By analysing the sequential processes in Equation 1 and the known neural types and their connections in neocortex, they pointed out the “remarkable correspondence” between the sequential processes in the equations and the neural architecture... Although they suggested the involvement of inhibitory neurons in L1 earlier, among the diversity of distinct types of inhibitory neurons (Petilla Interneuron Nomenclature Group et al., ) many of them can “provide strong mono-synaptic inhibition to L2/3” (page 699) and there are no clear reasons given why the L1 inhibitory neurons should take the role of reversing the sign of g... Furthermore, they did not explicitly specify the function of the sign reversal by inhibitory neurons in Figure 5... Moreover, they also pointed out that (page 699) “feedback connections can both facilitate and suppress firing in lower hierarchical areas. ” How can this dualistic effect be exhibited by this circuit? Predictive coding theory is currently a highly influential theory for cognitive function and behavior, and one of the plausible theoretical frameworks that may explain the signal processing architecture of the cortex.

No MeSH data available.