Afferent processing
Afferent processing is my name for the processing of data by the human brain.
The data can be from the external or internal senses or from the internal processes of the brain.
All external incoming signals are processed by the brain because that is all it can do to try to learn about the outside world,
and processing internal data about the body and the brain helps it to learn about and control itself.
Evolution has created a brain that processes data in order to learn and therefore to survive.
The end result is what can be described as a model of my world in my brain,
which includes my body and my brain itself.
My description of afferent processing spans four contiguous
levels of description in my
hierarchical structure of seven levels,
with the most important intermediate emergent concept
being the symbol schema, the representation of a concept in the brain.
The processing of data creates or strengthens afferent connections to create
symbol schemas and the connections between them, which together make up the model of my world,
but it also creates or strengthens important efferent connections
in the reverse direction, back towards where the data came from, which are vital for higher-level functions.
I have created a number of afferent processing examples
that show how this processing could take place based on the known capabilities of neurons
and synapses.
Contents of this page
|
Naming - details on my name for afferent processing.
|
The science of data processing - a brief history of the science of data processing in the brain and other names used.
|
Overview of my proposals - overview of afferent processing.
|
Detail of my proposals - details of afferent processing.
|
Conclusions - some final thoughts.
|
References - references and footnotes.
|
Naming
- I could not find a generic term that described the inward-bound processing of all data,
so I have called this “afferent processing”, “afferent” meaning “conducting inwards”,
as opposed to “efferent” which describes the signals going outwards,
back towards where the data came from.
- The adjectives “afferent” and “efferent” are often used when describing signals
to and from the nervous system outside the brain. For example, see
Afferent nerve fiber
and Efferent nerve fiber.
- The terms were commonly used in neuroscience some years ago describing the direction of signals within the
brain1,
but this usage seems to be less common now.
- I have called this “processing” because there is some transformation of the data.
- The transformation is best described as coincidence detection at its lowest level (more details below),
but at a higher level of description it can justifiably be referred to as
compression, pattern matching, statistical learning, extraction of invariance, abstraction, correlation and application
of symmetries3
or the making of a model.
- When referring to data moving in the opposite direction, I use the phrase
efferent connections - “connections”
because there is no real transformation of data in this direction.
- Many writers and researchers use other terms such as forward, feed-forward or bottom-up for afferent,
and backward, feedback or top-down for efferent, which (for me at least) gets very confusing - examples include
Jeff Hawkins4 and
Andy Clark5,
6.
I prefer to use “afferent” and “efferent” to avoid these potential confusions.
- Another source of potential confusion is that the term “feedforward”
is used in a slightly different way in artificial
neural networks
which aim to emulate what the brain does to a certain extent.
- In these web pages I try where possible to avoid using directional adjectives such as up and down
when referring to hierarchical processing; however, the analogy is very powerful - raw incoming data starts at
the bottom level and becomes more compressed, abstracted or invariant at each step until it reaches the highest level.
The science of data processing and other names used
- The processing of data, particularly sense data from the eyes, has been investigated by many scientists
over many years, and a lot of low-level details are well documented. At the other end of the spectrum of
levels of description
there are a number of high-level theories of consciousness proposed by people
working in different fields and using various names,
but there is a large explanatory gap between these two ends of the spectrum,
and there are very few proposals that span the gap.
- The following sections give some details of the research into the low-level details of the processing
of sense data, the names used, and includes some comparisons with my proposals.
- At the lowest level, the processing of signals from the eyes in the
visual cortex
has been investigated in great detail since the 1950s and many aspects are now quite well understood.
- Simple features such as edges, lines and movement
in the visual field are identified very early in this processing
(i.e. by neurons that are closely connected to the cells in the retina).
- Neurons that fire in response to a particular feature of this nature
in just one specific orientation are known as
simple cells.
- Neurons slightly further away from the retina in terms of the number of
connections that fire in response to a particular feature, but in any orientation, are known as
complex cells.
- It is broadly accepted that this processing is hierarchical, with more complex
features being identified and a wider field of view (receptive field) being processed,
by a smaller number of neurons, at each higher
level7,
8.
- In the first few decades of this study, connections that went backwards, from
the higher levels of the hierarchy back to the lower levels were ignored because their function was not
understood4,
but now it is well documented that there are just as many of these
efferent connections than the afferent ones, if not more.
One explanation of this is the
Reverse
Hierarchy Theory7.
- The relevance of sideways (lateral) connections was also not understood in the
early days, and theories that explained them were often ignored (e.g. the
V1 Saliency Hypothesis),
but it is now well documented that
lateral inhibition
is the starting point for attention and affects even the lowest levels of processing of data.
- Much of this low-level processing almost certainly depends on connections
laid down in the developing brain that are governed by the genetic blueprint of DNA
(more on this below),
although they could easily change with learning and experience at any future point.
- The details of the processing of sound by the
auditory cortex
are less well documented, partly because there are no agreed definitions for higher-level representations in auditory
processing9,
but the evidence does suggest that the processing is hierarchical and that more general features are extracted at higher levels.
- Details of the processing by the others senses are even less known, but it is assumed that it is likely to
be similar to the processing of vision and sound.
- Sometimes the term perception is used to describe the processing of data, but the same term is also used
to describe the end result of the conscious perception of an object or concept, which I think can be confusing.
- I prefer to use perception to mean the conscious end result of afferent processing.
There can be afferent processing of incoming data that even gets as far as updating a symbol schema, but does not result in a perception.
- So I treat perception as a higher-level function than afferent processing because it depends on the
pre-existence of something that encompasses understanding of the data.
- The same applies to all the other ...ception words (see below for more on these).
- Sensation is a term sometimes
used for the early processing of sense
data10,
but it is usually restricted to the processing of data from the five commonly-recognised senses, and in normal usage is
a term that goes hand-in-hand with perception - there is no sensation unless there is a perception.
- Perceptual learning
is a term used by psychologists, but, as its name suggests, this is also related more to the
higher-level description of perception.
- Psychologists have also used the term
statistical learning11,
mainly when investigating how children learn, and normally applied only to sight and hearing (and particularly language acquisition),
but it could apply to general learning with any of the senses and should be applicable to adults as well as children.
- Statistical learning is similar to parts of the lower levels of description I outline here,
but described from a wider timescale perspective: if two events happen together once, it is a coincidence;
if they frequently happen together, it becomes a statistical probability.
- Statistical learning requires the involvement of prediction, which in turn depends on
efferent connections that are created by afferent processing.
- So statistical learning could certainly be said to be caused by what I describe as afferent processing.
- Psychologists and neuroscientists have investigated
Bayesian approaches to brain function,
which is a mathematical description of the way the brain may assess probabilities in order to learn
(there is more on this in the page on prediction).
- The Bayesian approach assumes that the rules laid down by
Bayes’ theorem
are used to update the probability of an event based on an occurrence of the event.
- However, it seems likely that the methods I outline here do generate broadly the same results as
Bayesian statistics would predict, because of the feedback and predictive effect of
efferent connections.
- The Bayesian approach has similarities to statistical learning discussed above, but with more mathematical rigour.
- Predictive coding
(or Predictive processing) is a relatively recent theory that considers the processing of data from the point of view of prediction
(there is more on this in the page on prediction).
- It encompasses Bayesian inference
and includes the making and updating of models, but gives few details on what the models consist of, or how they are updated.
- It defines prediction errors, sometimes called surprise, which is when the incoming data does not match the model.
- It also defined precision weighting as the amount of confidence (certainty or uncertainty) in the existing
model or schema verses the reliability of incoming data that may not match the model.
- Assessing the reliability of data is the most difficult aspect of this theory, because in many cases,
there is very little the brain could do to get this data. In practice, what is considered reliability is simply the
influence of the multi-level competition processes of attention.
- There is some recent evidence that the brain does not adhere to optimal learning
models12,
which may be a reason for some of our biases, particularly the well-known
confirmation bias.
- An addition to the theory is the
free energy principle,
a mathematical model of how this measure of confidence may be optimised using a principle that may apply to life in general,
not just the processing of the brain.
- The free energy principle encompasses perception, action, and some aspects of attention that may
apply to the processing in the brain.
- Attention is proposed to be a way in which prediction errors
(or surprise, or free energy) are minimised by the selection of the best fit with the model, and suppression
of data that does not fit.
- Action is also proposed to be an attempt to minimise prediction errors
(or surprise, or free energy) - if my model does not fit the input, I take action to make it fit if possible.
- Scientists working mostly in the area of
Artificial Intelligence
have developed many cognitive architectures
and artificial neural networks
that are all attempts to model the workings of the brain, mostly with a view to reproducing it on a computer.
- These models have methods for processing incoming data, but very few seem to be based on the known architecture of the brain.
- One model of processing that is designed to resemble the known architecture of the brain is Jeff Hawkins’
Hierarchical temporal memory and
Memory prediction framework.
- He uses the term abstraction
in the same way that I do, meaning the processing of sense data to create invariant (unchanging) representations.
- A similar theory has been proposed by
Max Bennett13,
who has some good suggestions as to why the neocortex is ideally suited to provide multiple layers of hierarchical processing.
- Another idea is that compression (called compressionism by some) is the key to the processing of data in the
brain that leads to consciousness14.
This is similar to the way a computer compresses data, but the idea is that the brain not only compresses incoming data but it also
compresses data about itself, and that this is key to producing consciousness.
- Coincidence detection as a processing method by neurons has been proposed and
investigated15,
16,
but, as far as I know, the idea of it being the main driver of the processing of data has not been further followed up.
- To try to summarise the work of many thousands of researchers over many decades relating to the processing of data
by the brain may seem an optimistic task, but I think there is evidence to support the general consensus that:
- It is done in a similar way for all types of sense data.
- It consists of many recursive, hierarchical levels where each level carries out similar processing.
- Each level processes the data from the previous level.
- Each higher level has a wider receptive field.
- Each level compresses the data and extracts more general features.
- Each level uses prediction with the aim of reducing prediction errors.
- This means that each level uses top-down (efferent) signals as well as bottom-up (afferent) signals.
- The end result approximates to Bayesian statistics and therefore also the minimisation of prediction errors and free energy.
- Prediction means that perceptions can be resolved much more quickly than if done solely by afferent processing.
- The end result are neural representations or schemas of the incoming data that are constantly being updated.
- Connections between these schemas represent connections and relationships in the real world.
Overview of my proposals
- My proposals aim to include all the features listed in the six numbered points immediately above, but additionally aim to include:
- Low-level details of the common processing that the brain may use at each hierarchical level.
- How the processing of data from within the brain is carried out in exactly the same way as sense data is processed
(what I call cognoception).
- How the processing of data includes lateral inhibition, creating a competitive selection processes required to explain attention.
- Details of how efferent connections are created and updated, and why they are fundamentally important to higher-level functions.
- The definition of intermediate levels of description, including new concepts and behaviours, that can help to explain the hierarchical processing.
- Precise definitions of the symbol schemas that are the end results of the hierarchical processing and how they are used.
- Afferent processing is the processing carried out by the brain on all available data.
- Available data means all the electrical signals from external and internal senses as well as signals in the brain
generated by the processing of the brain. This includes:
- The well-known five external senses: sight
(ophthalmoception), hearing
(audioception), smell
(olfacoception), taste
(gustaoception) and touch
(haptic perception).
These together are often commonly known simply as
perception,
although some would use this term for all the other items below as well.
- Others sources that are often not thought of as senses, such as
balance (equilibrioception) and
heat (thermoception).
- The sense of the position of parts of the body
(proprioception).
- The senses of the internal state of the body
(interoception),
such as pain
(nociception),
appetite or hunger,
body temperature, and others.
- Internal brain processes, the processing of data related to things like memory, attention and free will.
(I call this cognoception, but this is a term I have invented because there does not seem to be a term that describes it.)
As discussed relating to perception above,
all these ...ception words describe a higher level view which may (or may not) be the result of the afferent processing of data.
Data from any of these areas may be processed by a small number of hierarchical levels, but may not make it as far as to update
a symbol schema, so will not be remembered, even subconsciously.
- The processing happens autonomously (automatically and unconsciously) all the time except
in the following circumstances:
- When I am asleep data processing is limited, and varies with the phase of
sleep, which is largely controlled by
neuromodulation.
- Some processing obviously still takes place even when I am deeply asleep,
otherwise I would not be woken by an alarm or someone calling or shaking me.
- Other processing that is critical to the working of the body also continues.
- If I am under the influence of any significant amount of a drug that is, or
emulates, a neuromodulator or is psychoactive (e.g. an anaesthetic, alcohol, medication or illegal drug)
processing may be limited or disrupted.
- If I am in a coma, processing does not happen at all except for the monitoring
of the crucial autonomous functions, such as body temperature and heart rate.
- Afferent processing involves massive parallel processing of the millions of signals that are coming in all the
time17,
18,
which is the only thing that the brain can do if it is to survive in the world.
- The brain is locked away inside the protective shell of the skull.
- The brain has no direct senses itself, there are no pain or movement
sensors inside the skull.
- It must rely on electrical signals from the various external and internal senses
to be able to learn about the world, including the body of which it is a part.
- The more it can learn, the more likely it is to be able to survive, hence
evolution has created an exquisite analysis machine using a large number of nerve cells
connected together in intricate ways.
- So this processing is how the brain learns, how it remembers, and is
the way that experience and intelligence are created.
- Afferent processing is a recursive and hierarchical process that results in
the creation of a hierarchical structure of increasingly invariant and compressed representations.
Identical processing is carried out on all incoming data, from internal and external senses, and
also from internal brain processes.
- There is good evidence that the processing is recursive and
hierarchical19,
20.
- The most obvious evidence that processing of different sources of data is done in the same way
comes from examples where parts of the cortex that are normally used for receiving one type of sense data,
if that sense data is missing, can easily be reused for another type of
sense data21,
24.
- Another obvious piece of evidence that all data is processed in the same way is that,
whatever the source of the incoming data, it is translated into identical electrochemical signals
that are all processed by neurons that have similar functionality. Once this data is inside the brain, there
no independent way of telling what type of data it is.
- From a very early age, the coincidences that are detected and the hierarchies
that are built up means that certain parts of the brain deal with certain sources of signals.
- If this is true, then there must be a common functionality that is applied
to all data at all levels in the hierarchy.
- If this is true, and the appropriate structures exist within the brain to carry out
unlimited associative learning,
then I believe that it is inevitable that exactly the same processing will be carried out on
any other data that is available within the brain.
Detail of my proposals
- I divide afferent processing into four hierarchical levels of description:
- Afferent processing can be described as the application of
memory-enhanced coincidence detection
in a hierarchical and recursive fashion to the incoming sense data.
A description follows; you might find it useful to first look at the simplified
afferent processing examples 1, 2 and 3.
- Any two or more neurons in the brain, each connected to
sensory neurons,
that fire at the same time, or very closely together, are quite likely to have been triggered
by the sensing of a common (or invariant) source.
This source could be, for example:
- Part of a physical object that has an unchanging physical structure but perhaps
with colour and brightness variation,
a fixed boundary so that it moves as one object, and a specific location;
- An extract from a sound which has a specific pitch and timbre (tone),
and which originates from a specific location;
- Signals from an internal sense resulting from a movement of a specific muscle or
from some other event;
- Signals from an unchanging location, background or timing.
- The processing of these “coincidences” in incoming data can be broken down into five phases:
- This “coincidence” of two or more neurons firing at the same time will strengthen connections
between and within this first set of neurons, so that if the same thing happens again, the whole set is more likely to fire again.
- This is also likely to trigger the firing of a second set of one or more neurons that are afferently connected
(further away from where the sense data came in) to the first set, in a hierarchical fashion, and those afferent connections will also
be strengthened so that if the same thing happens again, they will be more likely to fire again.
- Any efferent connections
(in the opposite direction, back towards where the data came in) from the second set of neurons back to the first will also be strengthened,
because the second set is now firing at the same time as the first (or actually very shortly afterwards).
- If the same “coincidence” happens more times, the connections within the first set and between the
first and second set will be further strengthened.
- Multiple occurrences of the same “coincidence” will cause the connections to be made even stronger,
because efferent connections can create a loop and cause the activations to repeat and to become a short burst.
- On the other hand, if the “coincidence” doesn’t happen again,
those connections may be weakened and eventually “pruned” altogether,
which is one of the “housekeeping” jobs that happens during sleep.
- With many occurrences of “coincidences” from different areas or aspects of the object being sensed,
this process will be repeated in a recursive fashion, and multiple levels of a hierarchy will be built up.
- An inverse tree-like structure will slowly be created with many inputs at the bottom, and
each higher hierarchical layer responding to a larger set of sensory neurons.
- Each higher level is a more compressed and abstracted representation of a larger part of the object being sensed.
- The end result is a set of neurons and the connections between them that
“represents” the whole object being sensed, and this is what I call a symbol schema.
- When this point is reached, no more compression or abstraction is possible for this particular object, because
the unchanging or invariant core of the object is represented.
- Different occurrences of a sensing of the same object will almost certainly have some different aspects,
but these can be recorded as additional properties in different symbol schemas to which this one is linked.
- The processing described in the preceding paragraph happens in a matter of a few milliseconds, far too quickly for new synapses to be created.
- So the only way it can happen is by changes in the strength or reliability of existing synapse connections.
- In the longer term, possibly many hours later, these links and circuits will be replayed,
the newly-strengthened connections will be made more permanent, and new synapse connections may be created that
are “short-cuts” for newly-strengthened longer-route connections.
- This replaying is often said to happen during sleep,
although it is possible that similar processes can happen while awake as well.
- The replaying while awake is likely to happen as part of the normal flow of
thought from one thing to another. It may or may not be conscious
(which means it may or may not connect to the self symbol schema).
- Replaying during sleep may or may not be remembered as a dream.
- Although this description of how sense data is processed can explain how invariant representations
are created, in a real brain many connections are already in place at the time of birth, created from the blueprint
of DNA and developed by evolution because of the advantage they
provide23.
- A new-born baby can recognise a face as soon as it opens its eyes.
- A new-born baby also has a number of automatic
primitive reflexes,
and will unconsciously copy certain actions it sees, such as putting its tongue
out22.
- So there are certainly some areas of data processing that are not created by my proposed methods,
the best-known one is the first few layers of the processing of visual data that starts in the retina, but
there may be others.
- There is a very large step between examples of a few neurons being triggered
at the same time in response to a coincidence, to the building up of a set of
symbol schemas with relationships between them that represent a model of my world.
- It certainly is a slow process - consider the number of years it takes to build up a useful model
starting with the helplessness of a new-born baby.
- Imagine the brain of a baby who sees a red frisbee for the very first time.
- Connections will be made within their brain to their symbol schemas for red, round, and plastic, because they will have seen
a number of red balls, plastic toys and other round objects before.
- Moving their head and then seeing the same frisbee from a different view several seconds later,
they may at first not connect it with the first view of the frisbee, but it will also connect to the other symbol schemas.
- It could be many viewings later before they makes the connection, and from then on other viewings
are likely to make the connection much more easily and quickly.
- Afferent processing examples 4 and 5 give details
of how this might happen.
- Children have many more connections between brain cells than adults, and as they grow, connections are pruned.
The “prediction matching” that goes on must be the key to this pruning, with connections that are not used in
successful predictions being the ones that disappear over time.
- Babies do not have the previous experience to allow very much compression to happen,
so full details of every experience must be stored. This is why young children’s brains have many more connections.
As certain connections are strengthened with compression from further experiences, so unused connections are pruned,
and the total number of connections reduces considerably.
- So this is one of the reasons why young children need a lot of sleep - they have a lot of connections that need pruning.
- The full “processing” cycle is only activated for new experiences, new concepts or new aspects of existing concepts.
Efferent connections are used to predict sense data and also to fill in missing data.
So compression is used to its fullest extent in a new born baby, but subsequently mostly partially.
- How can the brain can possibly recognise a frisbee when there are potentially hundreds of
different angles, sizes, colours and lighting conditions that it could be seen in? And this is just one of
many hundreds or even thousands of things that need to be recognised and perceived.
- The main reason that it is possible is because of the huge numbers of neurons and synapses involved.
- There are 86 billion neurons in the brain, and even though only a small proportion of these are
employed in visual processing that means there are many millions involved.
- As an example, it is estimated that there are around 800 million neurons in the visual
cortex in total2.
- There are about one million separate input
streams17,
18, with 10,000 neurons per input
stream2
just in the first visual processing area (known as V1), with each neuron have potentially thousands of synapses.
- More importantly, any one neuron can be used for many different purposes,
since it has connections to many thousands of other neurons.
- It is the connections between neurons that represent knowledge and create intelligence;
each connection is a synapse, and there are around 100 trillion synapses in an adult brain.
- Higher-level processing depends on connections between symbol schemas that are created
and/or strengthened solely by learning and experience. This is the source of memory and intelligence.
- Having said that, only angles that are encountered form part of the symbol schema,
which is usually only realistic ones.
- That is why certain strange views of things are not immediately recognised, i.e. not unconsciously recognised,
but the conscious mind can usually redirect attention and perhaps even perform mental rotations to work out what the object is
(see everyday objects from unusual angles).
Conclusions
- Afferent processing needs to answer three separate but related questions of
“what”, “where” and “when”:
- What is the thing being sensed?
- This depends on the reliable and repeatable
nature of the world, which is how an invariant representative symbol can be created.
- Where is it located?
- This can be considered to be an attribute or property
of (the invariant version of) the thing being sensed.
- It changes over time, so is represented in separate symbol schemas,
and is recorded in absolute as well as relative terms;
in other words, the specific position of something (e.g. in the corner of the garden) as well as
the position of something relative to me and/or to other things.
- When is it or was it sensed?
- Time also needs to be tracked if something is changing over time.
- This is also separate symbol schemas and is also recorded in absolute and relative ways.
-
^
The Organization of Behavior
- Hebb 1949
downloadable here or see
GoogleScholar.
This classic work uses the terms afferent and efferent frequently (there are 80 occurrences of the word “afferent” and 16 of “efferent” in the book), although not all of these refer to pathways inside the brain. See, for example, the last paragraph of page xvi:
“We know a good deal about the afferent pathways to the cortex, about the efferent pathways from it, and about many structures
linking the two.”
-
^ ^
The Cognitive Neurosciences ed. Gazzaniga Fourth Edition MIT Press USA 2009
Page 315, first paragraph: “If one includes all the 250 million neurons in V1, the factor increases to 180, and for the whole visual cortex (nearly 800 million neurons), the number of neurons is almost 600 times the number of LGN neurons.
... It is estimated that in the foveal region, the density of neurons in V1 per unit solid angle of the visual field is 10,000 times the density of input fibers from the LGN. ... The number of neurons available for computations on the input from the fovea is truly phenomenal.”
Page 317, fourth paragraph: “...in the foveal region of V1, there are some 10,000 cortical neurons per input fiber from the LGN.”
V1 is the first and primary processing area for visual input. LGN is the Lateral geniculate nucleus, the area of the thalamus that sends data to the visual cortex.
-
^
Ibid. The Cognitive Neurosciences
Pages 316-8 discusses the subtle connection between “correlation”, which is the same as what I call coincidence detection, and “symmetry”. For example, a frisbee is still recognised as a frisbee even if it is seen in a different place (displacement symmetry) viewed from a different angle (rotational symmetry), in different lighting (contrast, shadow or colour symmetry), between the two eyes (binocular or stereoscopic symmetry) or even using different senses (feeling rather than seeing, say). “Invariance” is another way of describing this.
-
^ ^
On Intelligence and Google Book preview - On Intelligence
- Jeff Hawkins with Sandra Blakeslee, St. Martin’s Press, New York 2004
See some references to ‘feedback’ and, for example, pages 113-114:
“Moreover, there are as many if not more feedback connections in visual cortex as there are feedforward connections.
For many years most scientists ignored these feedback connections. If your understanding of the brain focused on how the cortex took input, processed it, and then acted on it, you didn’t need feedback. All you needed were feedforward connections leading from sensory to motor sections of the cortex. But when you begin to realize that the cortex’s core function is to make predictions, then you have to put feedback into the model; the brain has to send information flowing back toward the region that first receives the inputs. Prediction requires a comparison between what is happening and what you expect to happen. What is actually happening flows up, and what you expect to happen flows down.
The same feedforward-feedback process is occurring in all your cortical areas involving all your senses.”
-
^
Google Book preview - Surfing Uncertainty: Prediction, Action, and the Embodied Mind - Andy Clark, Oxford University Press 2016
e.g. pages 143-146 or some references to ‘feedforward’
-
^
Whatever next? Predictive brains, situated agents, and the future of cognitive science - Andy Clark 2013
doi: 10.1017/S0140525X12000477
downloadable here or see
GoogleScholar.
See for example note 8 on page 22:
“I have adopted the neuroanatomist practice of labeling connections simply as 'backward' and 'forward' so as to avoid the functional implications of the labels 'feedback' and 'feedforward.' This is important in the context of predictive processing models, since it is now the forward connections that are really providing (by conveying prediction error) feedback on the downward-flowing predictions...”
-
^ ^
View from the top: Hierarchies and Reverse Hierarchies in the Visual System - Hochstein and Ahissar 2002
doi: 10.1016/S0896-6273(02)01091-7
downloadable here or see
GoogleScholar.
Text accompanying Figure 1 on page 792:
“Classically, the visual system was seen as a hierarchy of cortical areas and cell types. Neurons of low-level areas (V1, V2) receive visual input and represent simple features such as lines or edges of specific orientation and location. Their outputs are integrated and processed by successive cortical levels (V3, V4, medial-temporal area MT), which gradually generalize over spatial parameters and specialize to represent global features. Finally, further levels (inferotemporal area IT, prefrontal area PF, etc.) integrate their outputs to represent abstract forms, objects, and categories. The function of feedback connections was unknown. Reverse Hierarchy Theory proposes that the above forward hierarchy acts implicitly, with explicit perception beginning at high-level cortex, representing the gist of the scene on the basis of a first order approximate integration of low-level input. Later, explicit perception returns to lower areas via the feedback connections, to integrate into conscious vision with scrutiny the detailed information available there. Thus, initial perception is based on spread attention (large receptive fields), guessing at details, and making binding or conjunction errors. Later vision incorporates details, overcoming such blindnesses.”
-
^
The Human Visual Cortex - Grill-Spector and Malach 2004
doi: 10.1146/annurev.neuro.27.070203.144220
downloadable here or see
GoogleScholar.
Page 661, second paragraph, second sentence:
“Here we propose an alternative framework to examine the human object recognition system as a hierarchical system. The underlying idea implies that object recognition is implemented in the brain through a series of processing stages, in which more global and invariant representations emerge up the hierarchy of the processing stream. The hierarchy is implemented through a gradual transition, from local representations that are closely tied to the retinal image to abstract representations that are closely linked to perception.”
-
^
On the emergence and awareness of auditory objects - Shamma 2008
doi: 10.1371/journal.pbio.0060155
downloadable here or see
GoogleScholar.
Page 1142, second column, middle of second paragraph, to page 1143:
“...for animal experiments to unravel the mysteries of complex sound perception, it is essential that they encompass the study of auditory attention, memory, and plasticity, preferably while the animals are engaged in appropriately complex behaviors... Such approaches are common in experiments with animal vision, but are still rare in auditory research. One reason for this dearth is the absence of widely accepted orderly neural representations of higher auditory features (analogous to orientation, motion, and color selectivity in vision) that can serve as assays for the neural correlates of attention and learning in the auditory cortex. There is also a lack of clear understanding of the physiology and functional organization of secondary auditory fields beyond the primary auditory cortex (analogous to the V2, V4, MT, MST, FEF, and IT areas in vision). And finally, there is still much to be discovered about the anatomical connections within the auditory fields, and between the auditory fields and structures serving higher levels of representation, such as the prefrontal cortex.”
-
^
Cognitive Neuroscience: The Biology of the Mind - Gazzaniga, Ivry and Mangun, Fourth Edition 2014 Norton & Company USA
Page 164, Chapter 5 under the heading “Senses, Sensation and Perception”, first paragraph:
“Perception begins with a stimulus from the environment... Sensation refers to the early processing that goes on.”
-
^
Statistical Learning: From Acquiring Specific Items to Forming General Rules - Aslin and Newport 2012
doi: 10.1177/0963721412436806
downloadable here or see
GoogleScholar.
Abstract, page 1:
“in this article ...we review evidence and provide a unifying perspective that argues for a single statistical-learning mechanism that accounts for both the learning of input stimuli and the generalization of learned patterns to novel instances.”
-
^
Stubborn predictions in primary visual cortex - Yon, Thomas, Gilbert, de Lange, Kok and Press 2022
doi: 10.31234/osf.io/ndufg
downloadable here or see
GoogleScholar.
Page 12, last paragraph of “Discussion”:
“In conclusion, we have found that stimulus representations in primary visual cortex continue to be 'sharpened' in line with top-down expectations, even when the environment signals that 'expected' events are no longer likely. Evidence that our perceptual systems form such 'stubborn predictions' reveals important ways that learning and prediction in the human brain may depart from principles laid down in optimal models - possibly due to intrinsic constraints on our ability to estimate different kinds of uncertainty.”
-
^
An Attempt at a Unified Theory of the Neocortical Microcircuit in Sensory Cortex - Bennett 2020
doi: 10.3389/fncir.2020.00040
downloadable here or see
GoogleScholar.
Second paragraph of Introduction, pages 1-2:
“... the neocortex seems to be made up of repeating subunits called 'macrocolumns', each of which contains the same types of neurons, connectivity, and firing properties. This observation has led to the hypothesis that the neocortex is just a repeated replication of the exact same microcircuit and that there was an evolutionary benefit to this duplication. Additional support for this can be seen in rerouting studies, whereby rerouting visual input to auditory cortex seems to convert auditory cortex into a visual cortex, suggesting that the only difference between these two regions is the input they receive and not the computations they perform. This is further supported by the fact that the human neocortex increased in size by almost 3-fold over just the last 3 million years of human evolution, a time frame likely too fast for any new circuitry to emerge other than a duplication of existing circuits. This hypothesis suggests that the only difference between any two areas of the neocortex is the inputs it receives, and the location it sends its outputs - the actual computations themselves are the same.”
-
^
Compressionism: A Theory of Mind Based on Data Compression - Maguire, Mulhall, Maguire and Taylor 2015
downloadable here or see
GoogleScholar.
This very interesting paper contains a discussion on how compression can lead to intelligence, but goes a lot further in its claims (see self-awareness).
The following is a quote from an old draft page based on this paper on
wikivisually which no longer exists.
“Compressionism is the idea of representing intelligence and/or consciousness in terms of data compression. ...
Compressionism proposes that the awareness of a system is something that can be quantified in terms of the data compression it carries out: conscious systems are those who [sic] data compression is so sophisticated as to be feasibly irreversible, thus forcing observers to adopt the intentional stance to make predictions about behavior.”
-
^
Integrator or coincidence detector? The role of the cortical neuron revisited - Konig, Engel and Singer 1996
doi: 10.1016/S0166-2236(96)80019-1
downloadable here or see
GoogleScholar.
First page abstract:
“Neurons can operate in two distinct ways, depending on the duration of the interval over which they effectively summate incoming synaptic potentials. If this interval is of the order of the mean interspike interval or longer, neurons act effectively as temporal integrators and transmit temporal patterns with only low reliability. If, by contrast, the integration interval is short compared to the interspike interval, neurons act essentially as coincidence detectors, relay preferentially synchronized input, and the temporal structure of their output is a direct function of the input pattern. Recently, interest in this distinction has been revived because experimental and theoretical results suggest that synchronous firing of neurons might play an important role for information processing in the cortex. Here, we argue that coincidence detection, rather than temporal integration, might be a prevalent operation mode of cortical neurons.”
-
^
Soma-axon coupling configurations that enhance neuronal coincidence detection - Goldwyn, Remme and Rinzel 2019
doi: 10.1371/journal.pcbi.1006476
downloadable here or see
GoogleScholar.
Page 2, Introduction, second sentence:
“Coincidence detection is a fundamental neural computation that allows the brain to extract information from the temporal patterns of synaptic inputs. In the cortex, neurons have biophysical specializations compatible with coincidence detection”
-
^ ^
Brain Mechanisms of Vision - Hubel and Wiesel 1979
doi: 10.1038/scientificamerican0979-150
downloadable here or see
GoogleScholar.
Page 43, second paragraph:
“The output from each eye is conveyed to the brain by about a million nerve fibers bundled together in the optic nerve. These fibers are the axons of the ganglion cells of the retina.”
-
^ ^
On Intelligence and Google Book preview - On Intelligence
- Jeff Hawkins with Sandra Blakeslee, St. Martin’s Press, New York 2004
Page 55, last paragraph: “Visual information from the outside world is sent to your brain via a million fibres in your optic nerve. ... Sounds are carried in via the thirty thousand fibres of your auditory nerve. ... Your spinal cord carries information about touch and internal sensations to your brain via another million fibres.”
-
^
Circular inference: mistaken belief, misplaced trust - Deneve and Jardri 2016
doi: 10.1016/j.cobeha.2016.04.001
downloadable here or see
GoogleScholar.
Although this paper is primarily about the possible causes of psychosis, it provides a good recent overview of the understanding of the way that sense data is dealt with.
Page 41, under the heading “In search of a neural mechanism”:
“The brain is inherently hierarchical and progressively constructs increasingly more abstract interpretations of its environment.”
-
^
A Theory of How Columns in the Neocortex Enable Learning the Structure of the World - Hawkins, Ahmad and Cui 2017
doi: 10.3389/fncir.2017.00081
downloadable here or see
GoogleScholar.
Page 12, under the heading “Hierarchy”: “The neocortex processes sensory input in a series of hierarchically arranged regions... the cortex learns multiple models of objects, both within a region and across hierarchical levels.”
-
^
Livewired - David Eagleman Canongate 2020
Page 37:
“When a person’s eyes are damaged, signals no longer flood in along the pathways to the occipital cortex (the portion at the back of the brain, often thought of as 'visual' cortex). And so that part of the cortex becomes no longer visual. The ships carrying visual data have stopped arriving, so the coveted territory is taken over by the competing kingdoms of sensory information. As a result, when a blind person passes her fingertips over the raised dots of a Braille poem, her occipital cortex becomes active from mere touch. If she gets a stroke that damages her occipital cortex, she’ll lose her ability to understand Braille. Her occipital cortex has been colonized by touch.”
-
^
Ibid. Livewired
Page 17:
“... babies will mimic an adult sticking out their tongue... fibers from your eye don't need to learn how to find their targets deep in the brain; they simply follow molecular cues and hit their goal - every time. For all this sort of hardwiring, we can thank our genes.”
-
^
How do neurons know? - Patricia Smith Churchland writing in Daedalus, the Journal of the American Academy of Arts & Sciences, Winter 2004
downloadable here or see
GoogleScholar.
Page 43, bottom of first column:
“Knowledge displayed at birth is obviously likely to be innate. A normal neonate rat scrambles to the warmest place, latches its mouth onto a nipple, and begins to suck. A kitten thrown into the air rights itself and lands on its feet. A human neonate will imitate a facial expression, such as an outstuck tongue. But other knowledge, such as how to weave or make fire, is obviously learned post-natally.”
-
^
Reorganization of Human Cerebral Cortex: The Range of Changes following Use and Injury - Elbert and Rockstroh 2004
doi: 10.1177/1073858403262111
downloadable here or see
GoogleScholar.
This wide-ranging and interesting paper gives a number of examples of reorganisation within the brain caused by a change to inputs.
Page 131 provides a “General Principles of Cortical Reorganization”: “The abolition of sensory input or deafferentation, as produced by amputation or segregation of a sensory nerve, results in the 'invasion' of adjacent cortical representations of intact parts of the sensorium into the cortical representation zone of the affected sensory part.”
Page last uploaded
Sat Mar 2 02:55:43 2024 MST