Some questions that arise in Machine Learning involve the prospect of using feedback from Higher-order areas (downstream) to Lower-order areas (upstream), and using Global Knowledge for Local Processing. A desire to gain insight into those issues from Neuroscience ("how does the brain do it?") led me to some fascinating investigations into the Microcircuits of the Cerebral Cortex. This blog entry is a broad review of the field, in the context of the original motivating questions from Machine Learning.
Starting out with a quote from the “bible of Neuroscience”:
The thalamus comprises many projections to the cerebral cortex; hence the name “anteroom”. Most of the information going to the cortex goes thru the Thalamus. For example, the lateral geniculate nucleus, the main central connection for the optic nerve to the primary visual cortex in the occipital lobe, resides in the thalamus.
(Perhaps more familiar to many people is the “hypothalamus”, meaning “below the thalamus”, which is responsible for the central control of homeostasis in the body.)
Both the thalamus and hypothalamus are part of the Diencephalon, an inner part of the brain (considered by some, but not all, neuroscientists to be part of the brain stem)
The number of layers is the same in most mammals, but varies throughout the cortex. In the neocortex, 6 layers can be recognized although many regions lack one or more layers.
The neocortex is the newest part of the cerebral cortex to evolve (prefix neo meaning new); the other part is the allocortex, which has just 3 or 4 cell layers.
From the perspective of ML, my take is that the brain cortex first evolved in a smaller “search space” (with fewer cell layers), and then parts of it “colonized” a larger search space of evolutionary possibilities. Akin to training a smaller network, and then adding extra layers and using the previously trained state as a starting point for the new training.
Layer IV is the main target of thalamocortical afferents from the thalamus.
Layer VI sends efferent fibers to the thalamus, establishing a very precise reciprocal interconnection between the cortex and the thalamus. That is, layer VI neurons from one cortical column connect with thalamus neurons that provide input to the same cortical column. These connections are both excitatory and inhibitory.
Cortical microcircuits are grouped into cortical columns and minicolumns. It has been proposed that the minicolumns are the basic functional units of the cortex. Functional properties of the cortex change abruptly between laterally adjacent points; however, they are continuous in the direction perpendicular to the surface. There is evidence of the presence of functionally distinct cortical columns in the visual cortex, auditory cortex, and associative cortex.
Source: Paneri, S., & Gregoriou, G. G. (2017). Top-Down Control of Visual Attention by the Prefrontal Cortex. Functional Specialization and Long-Range Interactions. Frontiers in neuroscience, 11, 545. doi:10.3389/fnins.2017.00545 link
Selected quotes from a 2007 review article:
Source: Gilbert & Sigman (2007), “Brain States: Top-Down Influences in Sensory Processing.” Neuron, Volume 54, Issue 5, 7 June 2007, Pages 677-696 link
From a 2016 article:
Source: Huang JY, Wang C and Dreher B (2017) Silencing “Top-Down” Cortical Signals Affects Spike-Responses of Neurons in Cat’s “Intermediate” Visual Cortex. Front. Neural Circuits 11:27. doi: 10.3389/fncir.2017.00027 link
From 2012 article “The Cell-Type Specific Cortical Microcircuit: Relating Structure and Activity in a Full-Scale Spiking Network Model” (link) Excitatory (black) and inhibitory (gray) connections with connection probabilities >0.04 are shown.
Central to the idea of a canonical microcircuit is the notion that a cortical column contains the circuitry necessary to perform requisite computations and that these circuits can be replicated with minor variations throughout the cortex.
George and Hawkins have suggested that the canonical microcircuit implements a form of Bayesian processing (George, D., and Hawkins, J. , 2009. Towards a mathematical theory of cortical micro-circuits. PLoS Comput. Biol.5, e1000532. )
The most popular scheme — for Bayesian filtering in neuronal circuits — is predictive coding (Srinivasan et al., 1982; Buchs-baum and Gottschalk, 1983; Rao and Ballard, 1999). In this context, surprise corresponds (roughly) to prediction error. In predictive coding, top-down predictions are compared with bottom-up sensory information to form a prediction error. This prediction error is used to update higher-level representations, upon which top-down predictions are based.
To predict sensations, the brain must be equipped with a generative model of how its sensations are caused. Indeed, this led Geoffrey Hinton and colleagues to propose that the brain is an inference (Helmholtz) machine (Hinton and Zemel, 1994;Dayan et al., 1995). A generative model describes how variables or causes in the environment conspire to produce sensory input. Generative models map from (hidden) causes to (sensory) consequences. Perception then corresponds to the inverse mapping from sensations to their causes.
In predictive coding, representations (or conditional expectations) generate top-down predictions to produce prediction errors. These prediction errors are then passed up the hierarchy in the reverse direction, to update conditional expectations.
The generative model therefore maps from causes (e.g., concepts) to consequences (e.g., sensations), while its inversion corresponds to mapping from sensations to concepts or representations. This inversion corresponds to perceptual synthesis, in which the generative model is used to generate predictions. Note that this inversion implicitly resolves the binding problem by explaining multisensory cues with a single cause.
Diagram source: Canonical Microcircuits for Predictive Coding (2012)
Note: intrinsic connectivity = within a cortical column ;
extrinsic connections = between columns in different cortical areas
The above is a simplified schematic of the key intrinsic connections among excitatory (E) and inhibitory (I) populations in granular (L4), supragranular (L1/2/3), and infragranular (L5/6) layers. The excitatory interlaminar connections are based largely on Gilbert and Wiesel (1983).
Forward connections denote feedforward extrinsic corticocortical or thalamocortical afferents that are reciprocated by backward or feedback connections. Anatomical and functional data suggest that afferent input enters primarily into L4 and is conveyed to superficial layers L2/3 that are rich in pyramidal cells, which project forward to the next cortical area, forming a disynaptic route between thalamus and secondary cortical areas (Callaway, 1998). Information from L2/3 is then sent to L5 and L6,which sends (intrinsic) feedback projections back to L4 (Usrey and Fitzpatrick, 1996). L5 cells originate feedback connections to earlier cortical areas as well as to the pulvinar, superior colliculus, and brain stem.
In summary, forward input is segregated by intrinsic connections into a superficial forward stream and a deep backward stream. In this schematic, we have juxtaposed densely interconnected excitatory and inhibitory populations within each layer.
Starting out with a quote from the “bible of Neuroscience”:
From Principles of Neural Science, 5th edn (Online book location 1435.3 / 5867). Emphasis and note added by me:
Sensory pathways are not exclusively serial; in each functional pathway higher-order areas project back to the lower-order areas from which they receive input. In this way neurons in higher-order areas, sensitive to the global pattern of sensory input, can modulate the activity of neurons in lower-order areas that are sensitive to local detail.
For example, top-down signals originating in the inferotemporal cortex might help neurons in V1* to resolve a detail in a part of the face.
For example, top-down signals originating in the inferotemporal cortex might help neurons in V1* to resolve a detail in a part of the face.
(*) V1 is the Primary Visual cortex
Thalamus : passageway to the Brain Cortex
Thalamus means “anteroom”. It's a brain region that features prominently in the brain circuits described in later sections.
The thalamus comprises many projections to the cerebral cortex; hence the name “anteroom”. Most of the information going to the cortex goes thru the Thalamus. For example, the lateral geniculate nucleus, the main central connection for the optic nerve to the primary visual cortex in the occipital lobe, resides in the thalamus.
(Perhaps more familiar to many people is the “hypothalamus”, meaning “below the thalamus”, which is responsible for the central control of homeostasis in the body.)
Both the thalamus and hypothalamus are part of the Diencephalon, an inner part of the brain (considered by some, but not all, neuroscientists to be part of the brain stem)
The Cerebral Cortex
The mammalian cerebral cortex, the grey matter encapsulating the white matter, is composed of layers. The human cortex is between 2 and 3 mm thick.The number of layers is the same in most mammals, but varies throughout the cortex. In the neocortex, 6 layers can be recognized although many regions lack one or more layers.
The neocortex is the newest part of the cerebral cortex to evolve (prefix neo meaning new); the other part is the allocortex, which has just 3 or 4 cell layers.
From the perspective of ML, my take is that the brain cortex first evolved in a smaller “search space” (with fewer cell layers), and then parts of it “colonized” a larger search space of evolutionary possibilities. Akin to training a smaller network, and then adding extra layers and using the previously trained state as a starting point for the new training.
The Cortical Layers
Connections "up" and "down" within the thickness of the cortex are much denser than connections that spread from side to side.
Layer IV is the main target of thalamocortical afferents from the thalamus.
Layer VI sends efferent fibers to the thalamus, establishing a very precise reciprocal interconnection between the cortex and the thalamus. That is, layer VI neurons from one cortical column connect with thalamus neurons that provide input to the same cortical column. These connections are both excitatory and inhibitory.
Cortical microcircuits are grouped into cortical columns and minicolumns. It has been proposed that the minicolumns are the basic functional units of the cortex. Functional properties of the cortex change abruptly between laterally adjacent points; however, they are continuous in the direction perpendicular to the surface. There is evidence of the presence of functionally distinct cortical columns in the visual cortex, auditory cortex, and associative cortex.
Studies mentioning “Top-down Signals”
A 2017 article gives experimental evidence (from monkeys with brain lesions) that:
“the prefrontal cortex (PFC) has long been considered a source of top-down signals that bias selection in early visual areas in favor of the attended features”
Source: Paneri, S., & Gregoriou, G. G. (2017). Top-Down Control of Visual Attention by the Prefrontal Cortex. Functional Specialization and Long-Range Interactions. Frontiers in neuroscience, 11, 545. doi:10.3389/fnins.2017.00545 link
Selected quotes from a 2007 review article:
All cortical and thalamic levels of sensory processing are subject to powerful top-down influences, the shaping of lower-level processes by more complex information.
The general idea of top-down influence is that complex information that is represented at higher stages of processing influences simpler processes occurring at antecedent stages. Whereas some of the earlier work on spatial attention—the most studied instance of top-down modulation—suggested that significant influences of attention are found only at high levels in the visual pathway, it is becoming increasingly clear that even at the earliest stages in cortical sensory processing the functional properties of neurons are subject to influences of attention, as well as other forms of top-down modulation.
The higher-order information may include learned, internal representations of the shapes of objects and of the abstract syntax of object relationships. It may also include information about behavioral context, which would include attention, expectation, and perceptual task.
These influences may not even be specific to cortex, but wherever one sees feedback connections, including thalamus. This study showed even stronger attentional effects in the LGN* than in V1/V2**. Top-down influences are not unexpected in the LGN since it receives input from many more V1 neurons, by orders of magnitude, than it receives from the retina.
(*) LGN : the Lateral Geniculate Nucleus is a relay center in the thalamus for the visual pathway
(**) V1 is the Primary Visual Cortex
Source: Gilbert & Sigman (2007), “Brain States: Top-Down Influences in Sensory Processing.” Neuron, Volume 54, Issue 5, 7 June 2007, Pages 677-696 link
From a 2016 article:
(1) “lower-order” visuotopically organized cortical areas, some of which receive their principal or a substantial, direct thalamic input from the dorsal lateral geniculate nuclei (LGNd), send numerous “feedforward” associational projections to the “higher-order” visual areas
(2) beyond the primary visual cortices, information about pattern/form vs. motion is processed along two largely parallel “quasi-hierarchical” feedforward streams
(3) higher-order areas send numerous associational “recurrent” or “feedback” projections back to lower-order areas
(2) beyond the primary visual cortices, information about pattern/form vs. motion is processed along two largely parallel “quasi-hierarchical” feedforward streams
(3) higher-order areas send numerous associational “recurrent” or “feedback” projections back to lower-order areas
Source: Huang JY, Wang C and Dreher B (2017) Silencing “Top-Down” Cortical Signals Affects Spike-Responses of Neurons in Cat’s “Intermediate” Visual Cortex. Front. Neural Circuits 11:27. doi: 10.3389/fncir.2017.00027 link
Microcircuits of the Cerebral Cortex
From 2012 article “The Cell-Type Specific Cortical Microcircuit: Relating Structure and Activity in a Full-Scale Spiking Network Model” (link) Excitatory (black) and inhibitory (gray) connections with connection probabilities >0.04 are shown.
Central to the idea of a canonical microcircuit is the notion that a cortical column contains the circuitry necessary to perform requisite computations and that these circuits can be replicated with minor variations throughout the cortex.
George and Hawkins have suggested that the canonical microcircuit implements a form of Bayesian processing (George, D., and Hawkins, J. , 2009. Towards a mathematical theory of cortical micro-circuits. PLoS Comput. Biol.5, e1000532. )
The most popular scheme — for Bayesian filtering in neuronal circuits — is predictive coding (Srinivasan et al., 1982; Buchs-baum and Gottschalk, 1983; Rao and Ballard, 1999). In this context, surprise corresponds (roughly) to prediction error. In predictive coding, top-down predictions are compared with bottom-up sensory information to form a prediction error. This prediction error is used to update higher-level representations, upon which top-down predictions are based.
To predict sensations, the brain must be equipped with a generative model of how its sensations are caused. Indeed, this led Geoffrey Hinton and colleagues to propose that the brain is an inference (Helmholtz) machine (Hinton and Zemel, 1994;Dayan et al., 1995). A generative model describes how variables or causes in the environment conspire to produce sensory input. Generative models map from (hidden) causes to (sensory) consequences. Perception then corresponds to the inverse mapping from sensations to their causes.
In predictive coding, representations (or conditional expectations) generate top-down predictions to produce prediction errors. These prediction errors are then passed up the hierarchy in the reverse direction, to update conditional expectations.
The generative model therefore maps from causes (e.g., concepts) to consequences (e.g., sensations), while its inversion corresponds to mapping from sensations to concepts or representations. This inversion corresponds to perceptual synthesis, in which the generative model is used to generate predictions. Note that this inversion implicitly resolves the binding problem by explaining multisensory cues with a single cause.
Diagram source: Canonical Microcircuits for Predictive Coding (2012)
Note: intrinsic connectivity = within a cortical column ;
extrinsic connections = between columns in different cortical areas
The above is a simplified schematic of the key intrinsic connections among excitatory (E) and inhibitory (I) populations in granular (L4), supragranular (L1/2/3), and infragranular (L5/6) layers. The excitatory interlaminar connections are based largely on Gilbert and Wiesel (1983).
Forward connections denote feedforward extrinsic corticocortical or thalamocortical afferents that are reciprocated by backward or feedback connections. Anatomical and functional data suggest that afferent input enters primarily into L4 and is conveyed to superficial layers L2/3 that are rich in pyramidal cells, which project forward to the next cortical area, forming a disynaptic route between thalamus and secondary cortical areas (Callaway, 1998). Information from L2/3 is then sent to L5 and L6,which sends (intrinsic) feedback projections back to L4 (Usrey and Fitzpatrick, 1996). L5 cells originate feedback connections to earlier cortical areas as well as to the pulvinar, superior colliculus, and brain stem.
In summary, forward input is segregated by intrinsic connections into a superficial forward stream and a deep backward stream. In this schematic, we have juxtaposed densely interconnected excitatory and inhibitory populations within each layer.
Comments
Post a Comment