General

Life is characterized by a decrease in entropy and an increase our probability in predicting states of the world. We can do this by mapping these states into brain states and then remembering and using this memory for prediction. Life defies the second law of thermodynamics.

Transduction is the art of taking sensory stimuli and converting it into bioelectrical signals. We can do this with chemical, mechanical, and electromagnetic transduction and breaking these stimuli down into features (by using different types of receptors). These types of stimuli can be immediately filtered by weber’s law (the minimal detectable change in stimulus intense is proportionate to the absolute intensity).

We can make use of lateral inhibition and convergence/divergence in order to build a circuit that integrates the incoming information.

We can organize this data dependent on where it came from (topographic mapping) and forming spatial and population codes based off of that. We can also use time in a similar fashion.

All of these transduction mechanisms can be modified by the associative architecture of the brain (plasticity). This is what is responsible for the McGurk Effect – our brain is great at picking up correlations in our environment. This shows us that perception is not fully determined by the bottom up processing within that modality.

In regards to touch, we sense pressure, temperature, and pain through mechanosensation. There are a good bit of receptors that are concentrated depending on their use. For example, there are 20k mechanoreceptors in one hand. This allows us to detect features of the environment that are related to the texture/tactile states of objects.

For audition we are using receptors to sense air pressure. There are only a few receptors…3500 for each cochlea, which can detect pitch and loudness as features in the environment.

For vision we are using photoreceptors (most in the brain – 100million per retina) to detect electromagnetic waves. This is the case because the visual system is working with more complex stimuli and representing more points in space.

We can represent codes by labeled lines or via population codes, where one neuron could fire to multiple stimuli but the combination is different in accordance with other neurons. Most codes are represented in the brain in a spatio-temporal manner.

Audition

Our cochlea will decompose sound into underlying frequencies.

The physical property of sound waves is basically a cycle of compressed air moving from a source with a certain periodicity (one cycle). Given two sources of sound detection we can begin to talk about relationships of these cycles. That is, if a

Speech is defined as the spectral-temporal structure of a spectrogram.There are “FM Sweeps” that can make up intonation.

The outer structure of the ear (the peripheral auditory system) is responsible for helping to funnel compressed air to the tympanic membrane. The pinna will actually be useful in downstream interpretations of elevation of sound. The tympanic membrane will serve as an amplifier into mechanical pressure. Note that the tympanic membrane is large because of its ability to amplify the incoming signal.

The mechanical flow of pressure transpires as follows. The air gets turned into mechanical vibrations by the tympanic membrane which shift the inner ear bones along a pivot to hammer (stapes) onto the oval window, which converts the mechanical pressure into liquid waves that can then traverse the scala vestiuli, go around the helicotrema into the scala tympani and release the pressure by bulging out the round window.

The cochlea is the auditory transducing portion of the inner ear. The tectorial membrane and basilar membrane cut out a wedge on the side of the scalas which allows for a endolymph (High Potassium) fluid to hold up the tectorial membrane on top of outer hair cells. As the membrane shears atop the inner hair cells, the cilia atop them can shift towards their tallest cilia (kinocilium) and allows potassium to flow in, which depolarized the cell, but does not cause an action potential. Instead, the influx of positive ions from the endolymph in Scala media depolarizes the cell, resulting in a receptor potential. This receptor potential opens voltage gated calcium channels; calcium ions then enter the cell and trigger the release of neurotransmitters at the basal end of the cell. The neurotransmitters diffuse across the narrow space between the hair cell and a nerve terminal, where they then bind to receptors and thus trigger action potentials in the nerve.

Inner hair cells are the sensory ones and outer are the amplifiers. Inner hair cells are arranged in a single row whereas outer hair cells are in 3-5 rows. Inner hair cells are innervated by many fibers (They have divergence), whereas outer hair cells have one cell innervating many of them (convergence). The first action potentials are in the spiral ganglia of CNVIII.

The potentials of these cells can follow tone frequency, but only up to a point. That is, if a sound is coming at a specific frequency, the receptor potential will match that

frequency. But, if it gets too fast (2khz), then you lose the individual fluctuations of the receptor potential and it just stays in a “steady on”, “DC” state.

There is a cochlear tonotopy where the apex has low frequency sounds and the base (closer to the oval/round window) carries higher frequencies. This tonotopy persists even into its first synapse in the cochlear nucleus.

Similar to a fourier transform, we extract/ decompose compelx wave forms into their underlying frequency components. Each of the many fibers that are leaving the inner hair cells will have different responses depending on the frequency tuning of that inner hair cell. So, depending on the frequency make up of a sound, many hair cells that make up the underlying components will fire. You can see the tuning curve of thee cells by seeing the amount of db of a particular sound that needs to be present in order to illicit a receptor potential.

In what is known as phase-locking, cranial nerve neurons will spike at the crest of the frequency wave function.

Outer hair cells can actually receive efferent input along the CNVIII and contract the basilar membrane in an attempt to “tune” the system”. This is really useful when we are paying attention to just one particular range of frequencies “human speech” and want to ignore others. Furthermore, outer hair cells can increase the threshold so that we are not responding to noise, but moreso waiting for a more discrete signal. You would need a higher dB level in order to have a receptor potential. Stimulatino of the olivocochlear bundle (from the SOC) can tune the receptive field properties of auditory fibers.

Otoacoustical emissions are sounds generated by the ear itself (since the pressure is being let out of the round window). Because of the contraction of outer hair cells, the ear can make its own noises that go outward through the tympanic membrane. A loud sound to the ear should cause a contraction to protect the cochlea and then this will cause an otautical emission.

The central auditory system consists of brain stem nuclei (cochlear nuclei, olivary complex, and inferior colliculus). Then the MGN of the thalamus. Then in cortex, A1 (Broadmann 41).

CNVIII will send out bilateral connections immediately through the trapezoid body. Lesions after this point, in a unilateral fashion, will not impair discrimination, but mainly sound localization problems.

The central auditory pathway is from CN VIII to cochlear nuclei then to the superior olivary nucleus, then the nucleus of the lateral lemniscus, then inferior colliculus, then MGN, then A1.

The inferior colliculus has a precise map of auditory space which was computed topographically by the olivary complex nuclei.

The primary auditory cortex is tonotopically organized where the apex of the cochlea (aka lower frequencies) is more anterior and base(higher frequencies) are more posterior.

Neurons of similar frequency preference are arranged in isofrequency bands (IFBs) across the primary auditory cortex (AI) of many mammals.

The tones that elicit maximal excitation also elicit maximal inhibition.

Neurons can be extremely selective as well, such as in the way that they can be sensitive to complex sounds and temporally ordered sequences…like particular songs that activate particular neurons in the songbird. These neurons have a non-linear response, usually. The response is not just a response to note A and then response to note B and combine them to have A+B, but instead it’s a unique non-linear combination of exclusively A following B that results in a large response.

There is typically a lateralization for complex sounds in humans. Right is usually more responsive to music and environmental sounds, whereas left is more for speech.

Sound localization relies on 3 independent, combinatorial mechanisms. Horizontal, azimuth, sound location relies on both interaural time delays and interaural level differences. Vertical, elevation, sound localization depends on the shape of the outer ear (pinna). Vertical elevation and decibels interact to form a dB specific detection of elevation by way of the pinna.

The idea of level differences is that if sound is coming from Left, then the left ear will be louder than the right ear due to the nature of sound absorption.

We can detect inter-aural time delays in a primitive way by examining the phase of a wave when it hits both our ears. If the phase is not totally in line (assuming the wavelength is long enough), then the sound did not come from directly in front of the perceiver. Then, you would be able to detect the phase discrepancy in the wave. However, if the frequency is too high, then it could very well be that the phase is in line, but its on a totally different wave at a later time.

Therefore, since it takes .5ms for sound to travel from one side of the head from another and a 4kHz tone has a period of .25ms, then within .5ms, we could have exactly the same phase of a wave. Therefore, we need frequencies basically lower than 4 khz and not exactly 2Khz. Basically…lower than 2khz would do the trick. And remember that’s what we really care about anyway since we do DC higher than 2Khz anyway. If we had a smaller head, then maybe we could detect the differences

for higher frequencies. If we had a large head, then we would need lower and lower frequencies if we wanted to detect these delays via this phase methodology.

It’s important to note that there is phase-locking of the action potentials of CNVIII with the peak of a sound wave of a pure tone. So…firing in response to a tone (under 2kHz) will fire at the peak of that tone’s wave function.

Another way to detect inter-aural delay detection is within the medial superior olive where there are “coincidence detectors”. Basically, the MSO receives input from both ears, but one ear will have a longer path. The axons connecting synapsing on neurons in the MSO will synapse on many neurons. Some of those neurons will have a shorter path from the left ear and the same neuron will have a longer path from the right ear. If that neuron were to receive input from both ears at the same time, then it must mean that the source came from the right ear (since it had to traverse a longer distance and got there at the same time, it means the signal must have gotten a head start…that is that the sound reached the right ear first).

When computing inter-aural level diffeences, we can see the computation with the following process. A strong stimulus excited the left ear and left LSO (lateral superior olive) and this stimulus will also inhibit the right LSO via MNTB (trapezoid body). Since sxcitation from the left side (in this case) is greater than the right side, its “net” will be less than that of the right side (since the left is inhibiting the right more strongly than the right is inhibiting the left) and this stronger signal can be sent to higher centers. This creates opponency.

We can also hear via conduction where the inner ear bones are vibrated. Hearing loss that involves a blocking of these inner ear bones or prevents them from converting air/mechanical pressure to water waves is known as conductive hearing loss.

Sensorineural hearing loss is when there is actual hair cell loss or brain stem lesions.

Weber’s test is done by clodding one ear and vibrating the skull. You should hear sound louder in the clogged ear because there will be an escape of sound out of the tympanic membrane (the wrong way) and it gets trapped by the clog in the ear and doesn’t escape and creates a partial feedback loop. This is useful for unilateral conductive and sensorineural hearing loss. A patient with conductive hearing loss in one ear would hear the tuning fork loudest in the affected ear. This is because the conduction problem of the middle ear would mask the ambient noise in the room, while the well-functioning inner ear picks up the sound via the bones of the skull causing it to be perceived as louder in the affected ear. A patient with a unilateral sensorineural hearing loss in an affected ear would hear the sound louder in the unaffected ear since the affected ear wouldn’t be doing anything.

The middle ear muscles (tensor tympani and tensor stapedius) are innervated by the V and VII cranial nerves respectively and alter the motility of the malleus and stepes respectively in an attempt to protect the inner ear against loud sounds.

CN V tensor tympanic malleusCN VII tensor stapedius stapes

Presbycusis is the gradual loss of hearing in the high frequency range. This phenomenon is used to help disperse loitering teenagers and not disrupt older customers.

Tinnitus is a continuous ringing at a specific frequency thought to be caused by otoacoustic emissions and outer hair cells possibly producing phantom sounds. Chronic tinnitus is associated with a centrally generated phantom sound. After hearing loss, the A1 can get reorganized to only represent a narrow band of frequencies and the phantom sound can be generated by a mishap in this remapping.

Sensorineural hearing loss can be aided with cochlear implants that can insert an electrode array into the cochlea and activate the CNVIII in response to specific frequeny mappings. However, with these, speech is compressed into a few channels instead of the wide range we need to understand these complex sounds, so speech might sound a bit garbled with too few channels.

Somatosensation

Functionally and morphologically distinct mechanoreceptors are in the skin. They are either encapsulated (typical mechanoreceptors) or “unspecialized” and unmyelinated (free nerve endings).

We can have either rapidly or slowly adapting. Rapidly adapting will respond only to the onset of a tonic stimulus.Slowly will keep firing throughout the duration of the stimulus.

Meissner’s are rapidly adapting. Merkel are slowly.

Mechanotransduction, unlike auditor conduction, will actually cause an action potential in the mechanoreceptors. Therefore, ganglion cells in the somatosensory system are both transducers and carriers of info to the CNS (since they are pseudounipolar).

Different fiber classes can be classified based on conduction velocity. Propriecption (from muscle spindles) are the fastest. Touch from merkel, meisser, pacinian, and ruffini, are 2nd with Ia and II fibers of the Alphabeta type.

Pain and termperature are really slow, using unmyelinated C fibers for slow pain and A-delta fibers for fast pain.

The density of these fibers in a region, the amount of convergence each neuron has downstream, and the cortical representation of that area determines the spatial resolution of mechanosensation for any portion of the body wall. The “two-point-discrimination” task can help determine this. If there is a high density of receptors, then you should be able to tell two points apart if they are close. However, if there aren’t many neurons then you might not be able to tell the same distance since one receptor is responsible for covering a larger area and two points might just activate the same receptor.

A-beta fibers carrying mechanosensation information will take the dorsal column tract (gracile and cuneate).

A-delta and C-fibers will ascend via the spinothalamic tract (after immediately crossing the midline and projecting up and down a few slices via lissauer’s tract).

The trigeminal pathway (from the face) has collections of cell bodies in the trigeminal ganglion which then synapses on the principal nucleus of the trigeminal complex, decussares and forms the medial lemniscus and heads up the trigeminothalamic tract to the VPM of the thalamus and then to S1.

The dorsal column pathway has gracile (lower) and cuneate (upper) tracts in the dorsal column for axons with cell bodies in the dorsal root ganglion. These will synapse on gracile and cuneat nuclei in the caudal medulla, decussate in the pyramids (as the internal arcuate fibers) and climb up the medial lemniscus to the VPL and then to S1.

The distribution of somatosensory information is organized in a somatoptic map with overrepresentation for certain areas (homunculus) that are shaped b experience. Feet and genitalia are more medial. Visceral are more tucked in towards insula. A rats barrel cortex is exactly represented in its V1 in the same topic mapping just inverted on both the x and y axes (similar to the retinotopic map). The brain stem and thalamus are responsible for this flipping. There is a vertical and horizontal spread of activity for neurons representing the whiskers in rat S1. That is, activation of Layer IV (main input layer from VPM of thalamus since its face) will spread to activity in the V with a slight time delay.

There are actually multiple maps in S1 along the postcentral gyrus around its curve. These additional maps represent rapidly adapting skin (pacininian), deep pressure joint position, deep tissue, join and muscle receptors, and rapidly and slowly adapting receptors. Together, these maps provide a complete picture of the body: what you are touching, the position of your joints, and which muscles are contracted that coincides with an external and internal state representation. So, merkel and

meissner will be projecting to different positions on S1 , but still in somatotopic organization.

There can be high-order feature detection as well where a neuron will only fire depending on the orientation of a stimulu or the directionality of a stimulus. These types of more complex neurons would be found in S2. As one moves into higher order somatosensory areas, there is an increased RF Size and complexity.

Lesions to S1 shows deficits in texture, size, and shape discrimination.

Lesions to posterior parietal cortex results in agnosias.

If you stimulate S1 directly, monkeys can make judgments based on intensities. Direct stimulation of S1 can substitute for real physical stimuli.

Some neurons in S2 are responsible for encoding “relative intensity” whereby the behavior of the neuron is not determined by the physical characterists of the current stiulu, but by the relative frequency in relation to the previous stimulus. So one stimulus could elicit two differen responses depending on the intensity of the stimulus that was presented beforehand.

Nociception(pain) involves specialized receptors such as free nerve endings that are sensitive to both thermal, chemical and mechanical stimulation. A delta are thermal and mechano and C are also polymodal.

Hyperalgesia is when repeated noxious stimulation can lead to nociceptors becoming responsive to mechanical stimuli. This is like if you got hurt in one area and then touching it, in a non-hurtful manner, it still transmits pain. This can lead to Allodynia where innocuous stimuli are now painful (sunburn). Histamines can decrease the threshold for nociceptor activation. In ventrally ediated hyperalgesia, dorsal horn neurons undergo sensitization.

Touch and pain pathways communicate. This is best illustrated by the gating of pain where pain is mediated because of a dynamica balance of C and Abeta fibers. Essentially, an Abeta mechanoreceptor can sprout collaterals that synapse onto inhibitory interneurons that go on and inhibit C fibers that are on their way up the anterolateral system. This would happen in the dorsal column right near lissauer’s tract. Since pain synapses immediately, it is the projection neuron for pain in the spinal cord that is getting inhibited by the non-synapsing Abeta fibers.

Referred pain is an example of convergence. Visceral organs will often synapse on the pain fibers coming from other regions and make it feel as though pain is coming from those parts of the body. Visceral and somatic nociceptors can converge onto the same dorsal horn neurons.

The anterior cingulate appears to be involved in the affective component of pain. fMRI reveals anterior cingulate and insula as being ultra activated during the perception of thermal pain.

Descending systems can modulate ascending pain signals by inhibiting them in the actual dorsal horn. The reticular formation, periaqueductal gray and raphe nuclei seem to be essential in this relay of top-down modulation.

Phantom limb syndrome can tell us a lot about somatosensory information by showing us that there is cortical reorganization and that if parts of the brain that represented a now long gone limb are being transferred to the face, then there may be some intermingling and crossed feelings/connections.

Visual System

Vision is in large part not what the retina “sees” but how the brain interprets the input the eye receives based on past experiences.

Feature search is defined as a parallel process in which the target and the distractors are maximally different, differentiated by a single property such as colour, shape, orientation or size. An example of this would be to pick out a red circle located within a group of black circles. This would cause a “pop out” effect and a bottom up effect.

Conjunction search occurs when the target and the distractors share similarities in more than one single visual property such as size, colour, orientation and shape. An example of this can be seen if the target is a black horizontal line while the distractors are made up of white horizontal lines and black and white vertical lines. The target therefore shares orientation (horizontal) but not colour with some of the distractors but also shares colour (black) but not orientation with other distractors, The similarities with the distractors make the target harder to identify as there is no 'pop out' effect as seen in feature search. This would be more top-down modulated. Conjunction search will increase will display size whereas feature search won’t really.

We have a couple design flaws: a blind spot and the fact that light needs to travel down several cell layers before being recepted.

Most of the refraction (bending of light) is contributed by the cornea and relatively little by the lens. However, the lens refraction is runnable by the ciliary muscles. This is great because you want to change the focus on near objects my making the lens thicker and rounder (contraction of ciliary muscles) and flatter for sitant objects (relaxation of ciliary muscles).

The retina has 90 million rods, 4.5 million cones, 1.2 million ganglion cells (convergence!). A digital camera might have something like 10 million pixels.

Light comes in and passes all the cells until it hits the tips of the photoreceptors in the pigment epithelium. Photoreceptors synapse on horizontal and bipolar cells which then synapse on ganglion cells and amacrine cells.

Light produces a hyperpolarization of the photoreceptors. That is, light turns OFF the dark-current. Our photoreceptors aremore active in the dark and they get inhibited by light.

Light’s photon’s activate rhodopsin which uses G proteins to close channels and reduce transmitter release.

Rods are good for black and white vision and the night vision. Cones are for day vision and concentrated in the fovea. Rods are concentrated in the periphery and cones in the fovea (where they enjoy a one to one connection with bipolar and ganglion cells – no convergence).

The outer segment of photoreceptors maximize the likelihood of photons colliding with the photopigment.

Color vision can be reduced to the overlapping waves of the sensitivity functions of 3 types of cones. Short ones from blue. Medium ones for green and long ones for red. The medium and long slightly overlap, but only long goes into the red. If you lost long you’d lose red. But if you lost medium you wouldn’t necessarily lose green. The combination of RGB can make up the whole spectrum of visibile light.

On center ganglion cells will fire when light is in the center of its receptive field and not when its dark. The opposite would be the case for the off-center ganglion cell.

Lateral inhibition can be seen as a subtraction of center and surround. A center plus surround will still have activity in an on-center ganglion cell, but no where near as much as a center only. And off-center that received center plus surround would fire slightly because the surround is excitatory. An off center and an on center would basically fire the same to a center plus surround stimulation.

On and off bipolar cells are responsible for sign inversion / conservation. ON Bipolar express mGlu which are inhibitor for the bipolar cell in the presence of glutamate. So, if they are constantly being inhibited by a photoreceptor with its constant depolarized state and release of glutamate, then once photons hyperpolarize the center photoreceptor and deactivate the photoreceptor, the glutamate will stop and the on-center bipolar will have the opportunity to be disinhibited and synapse onto an on-center ganglion cell. With the off-center bipolar cells, they are activated by glutamate (AMPA) and thus are constantly firing when the stimulation is dark in the

center and get turned off when the photoreceptor gets turned off (in the presence of light in the center).

The key is that off center will mainly fire if the center is darker as compared to the surround. And an on center will only fire if its center is brighter relative to the surround.

So if we have a photoreceptor in the surround and one in the center and a stimulus comes on in the center, then the center cone will be inhibited. However, the surround will not be inhibited and will continue to fire. The surround will synapse onto the horizontal and excited the horizontal. This horizontal will then go on to inhibit the center (which is already inhibited), thus allowing for less glutamate to reach the on-center (which gets inhibited by glutamate) and ultimately end up with more firing in the on-center ganglion cell. If we were to have light in both the center and the surround, then the surround cone would be inhibited and wouldn’t excite the horizontal cell which wouldn’t go on to further inhibit the center cone and thus there would be a bit more glutamate reaching and inhibiting the on center bipolar, which would ultimately cause less firing in the on-center ganglion cell in the presence of light in both the center and the surround.

This center-surround architecture contributes to light adaptation/dynamic range because the response is relative (as opposed to absolute) light intensities. Therefore we can get the same discharge rate regardless of luminance. The same spot intensity produces dramatically different firing rates depending on the intensity of the background, though.

The same architecture can be used to edge detect. If an on-center ganglion has some light in the center and then some dark in the surround, then it will fire maximally. This type of situation can only really occur at edges/borders of light intensities/colors, etc.

The central visual pathway is Optic nerve that sends the nasal side of the retina projections to the contralateral side in the optic chiasm (temporal stays ipsilateral). Then it goes to LGN in the thalamus which is layered in a CIICIC format of contralateral/ipsilateral regardlessof visual field. And then via the optic radiations to striate cortex. Layers 3-6 are parvo-cellular(input from cones), whereas 1-2 are magnocellular(more concerned with “where” and thus, ruds). Koniocellular are between layers and thought to be responsible for relaying short wavelength cones. P ganglion cells have small receptive fiels, slower axons, are slowly adapting, and have color content. They have low contrast sensitivity and there are about a million of them. M ganglion cells have large receptive fields, fast axons, and are rapidly adapting. They have no color content and there are only 100k of them with high contrast sensitivity. The LGN of the thalamus will project to striate cortex with a ocular domaince, forming columns of preferred activity depending on which eye was the source of the information (eye. Not visual field).

Deficits along certain portions of this pathway will result in different problems. For example, if you were to cut the optic nerve coming from the right eye, you would only receive information from the left eye, but both fields would be present. If you lesion the optic chiasm, you would only have information about the non-periphery (because the temporal remains ipsilateral and center information goes in at an angle to the lateral side of the retina). If you cut the optic tract on one side, then you lose one entire, contralateral visual field. If you cut the upward loop on one side, you’d lose the inferior contralateral quadrant. If you cut right before synapses onto striate, then you would get loss of one visual field, but with macular sparing.

There are also projectons from retinal ganglion cells to the superior colliculus, pretectum, and hypothalamus (SCN).

There are three theories for the explanation of blindsight. The first states that after damage to area V1, other branches of the optic nerve deliver visual information to the superior colliculus and several other areas, including parts of the cerebral cortex. These areas might control the blindsight responses, but still many people with damage to area V1 don't show blindsight or only show it in certain parts of the visual field.Another explanation to the phenomenon is that even though the majority of a person's visual cortex may be damaged, tiny islands of healthy tissue remain. These islands aren't large enough to provide conscious perception, but nevertheless enough for blindsight. (Kalat, 2009)

A third theory is that the information required to determine the distance to and velocity of an object in object space is determined by the lateral geniculate nucleus before the information is projected to the cerebral cortex. In the normal subject these signals are used to merge the information from the eyes into a three-dimensional representation (which includes the position and velocity of individual objects relative to the organism), extract a vergence signal to benefit the precision (previously auxiliary) optical system (POS), and extract a focus control signal for the lenses of the eyes. The stereoscopic information is attached to the object information passed to the cerebral cortex.

Reaction times are generally faster in the auditory system.

Receptive fields in V1 start to get more complicated, like needing edges, bars, and movement in order to fire. Neurons in the primary visual cortex will respond selectively to oriented edges. Each point in retinotopic space is represented by different orientation detectors.

Orientation selective neurons receive aligned retinal input. This highlights how V1 neurons can have receptive fields that can be laterally inhibited by other neurons.

As for ocular dominance columns, this is across the surface of the cortex, each layer (of the column) will respond to either contralateral or ipsilateral input. There will

be some cells that respond to both eyes equally, but they are small in number compared to the monocular amounts. Animals that are preyed upon have more monocular because they always need to be looking throughout their surroundings. Predators need more depth so they have more binocular.

Complex cells will respond to multiple positions in their receptive field as well as to on and off stimuli. Multiple cells usually converge to create complex cells in V1. 3 simple cells with center surrounds could converge onto neuron that would maximally fire to a bar of light instead of just a circle. If those are oriented in a specific way, then that adds to even more complexity.

In V1 you can see ocular dominance and orientation selectivity maps that look like pinwheels. That is, if you labelled each neuron by a color associated with a specific orientation, you’d get pinwheels where some potions seem to represent almost all possible oritentations just by moving left and right across the cortex. What is interesting though, is that in rats, these columns don’w exist in any patterns. Just purely random in their selectivity. no groupings.

Following pathways out of V1, we usually see that the dorsal pathway is concerned with “where”. This takes the magno cellular information and processes it downstream, like in MT and parietal lobe. The ventral pathway will be more concerned with “what” and is mainly headed for IC and contains parvocelular information.

If you lesion this ventral stream you will have object discrimination deficits in so you’d be impaired on a non-matched to sample task.

With a dorsal stream lesion you’d have trouble with relational properties like landmark discriminations. You’d be bad with global locations. You would be able to identify objects, but not where they are in relation to one another.

In the Inferior Temporal Cortex we see neurons that are responsive to complex stimuli. These are not just a linear combination of lines but moreso what we would “call” certain object. If an object is embedded in an environment you would still, probably, get preferred activity. More like “concept” / platonic form cells. These neurons will typically have a toleration for firing. So, if you are focused on a fixation and the object is right on the fixation, then the neuron that encodes that object will fire maximally. But if you remain fixated and move it to the periphery, it will have a graded decrease in its preferential firing to that same stimulus.

Plasticity

“Rerouting” is a huge phenomenon where if you ablate the superior colliculus by lesioning the projection from inferior colliculus to MGN, you’ll see the optic nerve

start to project to the MGN as well… The rewired A1 will exhibit visual orientation tunings just like we saw in V1.

Ocular dominance columns are a great example of plasticity. They reflect thalamo-cortical segregation that emerges over development.

If you close one eye in cats during the developmental period in what is known as monocular deprivation and you record from V1, you find that virtually all the neurons are responsive to the eye that was open. Even if you just closed the eye for 2 months then opened it again for 3 years, you still see the ipsilateral preference. However, if you open for 1 year then close for 2 then reopen, it will look totally normal.

Whats interesting is that ocular dominance plasticity is not due to thalamic plasticity. The responses in the LGN are normal during the relay of retinal events.

There is less arborization in layer IV of V1 for the thalamo-cortical axons of the deprived eye.

If you open up more cortical real-estate, then you can enjoy beneficial effects and increased perceptual ability. A deprived eye will have much worse acuity compared to control because it has less cortical influence. However, the eye that was open, since it got to take over mor cortical real estate will have more acuity as compared to a control animal. Reopening that depreived eye will increase acuity in that eye, but also decrease in the other eye, since it got to overperform before due to lack of the closed eye’s influence.

Binocular neurons could be formed by the convergence of the left and right neurons in the II and III layers of cortex.

Associative plasticity contributes to formation of ocular dominance columns. As we grow, any two points on one eye will correlate locally more so than the corresponding points in the other eye. If multiple neurons converge on a cell and they all fire together, then those cells will get more plastic with the converged-on cell. Then, it becomes easier for those cells to activate the converged on neuron and thus that converged on neuron is more selective to those initial neurons that fired together onto it. Furthermore, if left eye and right eye inputs are firing together onto a cell they are firing slightly different patterns. However, growth will encourage more activity into this cell, even if seemingly random at first. Under the BCM theory, this will increase the threshold for LTP for pre and post synaptic cells. Thus once the threshold catches up with the input, whatever is randomly drive the cell the most (which will be from one eye only), then that will catch the threshold and be given LTP and then be taken away in an exponential fashion, not letting the other cells have the chance to catch up to the threshold.

When you look at a vertical bar, the same neurons in both the left and right eye will see the same thing and there could be a correlated reaction that then drive a binocular cell.

The effect of strabismus is consistent with the role of associative synaptic plasticity (correlations) in the formation of binocular cells. If we decorrelate the eyes, we should see less binocular cells. In strabismus, since the lateral rectus is awry, the eyes are misaligned. Thus, the focal points are uncorrelated between each point. Strabismic animals have less binocular cells.

Learning rules essentially underly ocular dominance plasticity. In deprivation, the deprived eye will become weaker first and then the open eye will become stronger.

Similar effects like ocular dominance are emphasized by rat’s barrel cortex. If you trim all whiskers but one and record in another part of cortex that isn’t normally tuned for that whisker, you will see a representation for the remaining whisker as if that whisker was spread out over all neurons.

Overall, the plasticity criticap period that has been demonstrated is generally for Layer IV neurons. Layer II and III stay plastic much longer. That is, you are cementing thalamo-cortical, but not cortical-cortical.

Interesting note about plasticity with language is that you actuall lose the ability to hear the differences between certain sounds. You are not interested in minor variances. you don’t want to focus on mild differences, you want to increase your robustness. This is aking to increasing bias and decreasing variance. Learning a language is in part to learn not to discriminate irrelevant variances.

Plasticity manifests a lot with cortical map reorganization after denervation. You make certain finger representations large based on using them more often than not. You can measure this by calculating (using MEG) the distance from digit 1 from digit 5. Then, infer that 2,3,4 are between those and if they expand, then so should the distance from 1 and 5.

There is also cross-modal plasticity and compensation. For example, congenitally deaf cats have a lower threshold for detecting movement (their visual system is on overdrive). This increased visual performance then relies on parts of the auditory system. If you were to “cool” those regions, you wouldn’t be able to benefit from this advantage.

Most all of plasticity is due to temporal correlations. It’s the statistical nature of the world. Something touching one finger and then another happens more regularly than other sets of stimuli pairings.

If you train a monkey to have strong intradigit correlations (like what you would normally have with distal and proximal correlations within one finger), then you see

that many neurons turn out to have triple digit receptive fields. That is, if its distal (independent of finger) the neuron will fire. Neurons would respond equally well to a touch on any one of 3 fingers, say, so long as it was distal vs. proximal.

Furthermore, if you pair D2 and D3 whiskers, then you increase the ability of whisker D3 to activate neurons in barrel D2. Thus whisker pairing produces a shift in whisker dominance.

If we have a thalamocortical input coming in and this drive’s the primary whisker’s neurons….then this is locked after critical period. However, cortical cortical plasticity can still share information horizontally in layers II and II. So, whats underlying the new receptive fields seen by plasticity is the LTP of horizontal synapses. Furthermore, as seen by D-AP5 (an antagonist of NMDA), NMDA is responsible for LTP and thus increased plasticity. Not NMDA, no LTP, no increased plasticity.

Threshold of Just Noticeable Differences (JND) can be altered as a function of training within a particular hemi field.

Neurons that have the maximal derivative, the most change in response between any two orientation representations by that neuron, would be the most informative neuron. So, it is just not about which respond maximally to one stimulus, it’s the neuron that has the maximal derivative.

With training there will be no increase in the number of cells that are tuned to a trained orientation. Instead, the ones that have a little bit of overlap with that orientation change so as to maximize the derivative of their response in accordance with that particular orientation. The slope should be 0 at the trained orientation and the slope of the neurons at the non-trained orientations nearby would be greatest. The biggest increase in slope was on the neurons that neighbored the trained orientation. However, the change in the orientation tuning of these nearby cells’ tuning curves is asymmetric, where the slope is sharper on the side of the orientation tuning curve that is closer to the target orientation.

In V4, there is a decrease in the number of cells with a preferred orientation at the trained orientation.

Perceptual learning of contour integration is location specific. If you train a task in one part of the visual field, you don’t immediately witness that benefit in other parts of the visual field.

Training produces facilitation of V1 responses by adjacent collinear (iso-oriented) stimuli. Facilitation was partially task-dependent: weaker influence of multiple lines in the Trained Dimming task. Anesthesia abolishes training induced fascilitation. This is strong support for top-down location and task specific modulation of V1 responses.

Rats that are trained to “go” at 5kHz have expanded representation of 5 Khz in A1. Furthermore, rats trained to go for ANY 35dB tone sees an increase in areas devoted particular to that volume. So, there are overlapping dB and kHz regions in the cortex.

Two types of map plasticity occurs in A1 in order to optimize task solving. The percent of recording sites responsive to frequency training shows increase in number of sites tuned to 5kHz. The percent of sites according to intensity shows a shift in the distribution of neurons that respond to that particular loudness (since its not specific to a particular frequency).

Memory

The hippocampus is important for the formation of new declarative memories, but is not the site where long-term memories are stored.

Anterograde amnesia produced by temporal lobe lesions shows a dissociation between decarative and nondeclarative memories. New and old procedural memory is in tact even if the memory for learning those things is not retained. Puzzle solving and mirror drawing, motor learning are good examples of memory that remain after temporal lobe lesions.

Memory can be thought of as a storage of synaptic webs with weights that appropriately connect particular interrelated semantic concepts. So, if you “prime” nodes within a network of a semantic web, then it might be easier for you to pick up on memories that are within the more interconnected portions of that same web. This can help increase task performance even without conscious awareness. Lexical decision task is a great measure of this. When trying to figure out if a string of letters is a word, then related prime words, like “bone” before “dog” will decrease the reaction time to saying that “dog” is a word. These can be intact even with anterograde amnesia.

Our emmory was not designed to rapidly store large amount of unrelated pieces of information such as lists of names or numbers. This is why remembering someone was a baker is easier then remembering their last name is Baker. The profession of baking has a much larger semantic net where the word baker is stored, so it may be easier to activate baker by pure probability odds and the facts that it is much more richly encoded. As compared to just the name Baker which is an isolated node in a lonely network that lacks any serious interconnectedness.

The Framing effect elicits an interesting phenomenon where people will opt out of situations if given a “mortality” framework and told their chances of dying when compared to their chances of living in a “survival” framework even if the statistics are entirely match. The same applies with gambling. If you tell people how much they will lose as compared to keep, they will be more likely to gamble (loss averse).

Loss aversion refers to people's tendency to strongly prefer avoiding losses to acquiring gains. Some studies suggest that losses are twice as powerful, psychologically, as gains.

Memory has many fallibilities. When you show subjects a sequence that depicts a car accident and then ask them questions about a stop sign (when there was actually a yield sign), they will update and thing that there was actually a stop sign.

With contextual fear conditioning, the hippocampus is necessary for responses to context whereas it is not responsible for responses to a cue.

The concept of moving memories from a working memory workplace into long term memory is known as consolidation.

The first type of consolidation is cellular and that process is blocked by protein synthesis inhibitors. If you deliver anisomycin during the US pairing, then rats will freeze significantly less during LTM(24 hours), but freeze the exact same amount during short term memory (4 hours). Thus, cellular consolidation is when short-term memories are subject to ‘erasure’ by interventions such as protein synthesis inhibitors.

Reconsolidation is a temporarily altered state of the memory trace following memory reactivation. This altered state is characterized by increased sensitivity to amnestic agents, such as inhibitors of macromolecular synthesis, possible because of enhanced plasticity of the neuronal circuit that encodes the memory trace or parts of it. If you pair a CS and a US and then when you give an electroconvulsive shock after the condition stimulus, you can actually lose the fear response.

Fear memories require protein synthesis in the amygdala for reconsolidation after retrieval. Anisomycin (protein synthesis inhibitor) injected into the amygdala after recall produces amnesia. Thus, reactivation of a memory brings it online and susceptible to this interference.

With systems level consolidation, we see that the time recall of declarative memory can become independent of the hippocampus. Training and testing for contextual fear will remain after 26 days even if there if there a hippocampal lesion.

Systems levels posits that there are unlikned nodes in cortical circuits. Learning creates links by way of the hippocampus, which serves as a hub. Over time, the branch stemming from the hippocampus can connect within itself and not necessarily need the hippocampus. Thus systems consolidation has been successful when memories are accessed through a nonhippocampal-dependent path over time.

Essentially, learning is due to changes in the strength of synapses or the connections between neurons. Associative synaptic plasticity and homeostatic synaptic plasticity

(homeostatic plasticity refers to the capacity of neurons to regulate their own excitability relative to network activity) are responsible for memory.

Habituation memory a form of learning in which an organism decreases or ceases to respond to a stimulus after repeated presentations. Thus, it is a type of non-associative memory. This is primarily caused not by a decrease in sensory neurons, but a decrease in motor neuron firing that must be due in part to the depreciation of an excitatory interneuron’s response.

Sensitization is a non-associative learning process in which repeated administrations of a stimulus results in the progressive amplification of a response.

Hebb insisted, with his postulate, that when an axon of cell A is near enough to excited cell B repeatedly within the same time period, then the efficiency of that connection increases. “Neurons that fire together, wire together”.

NMDA Receptors detect associations at the molecular level. Associative LTP is dependent on these NMDA receptors. If you use APV to antagonize NMDA, you lose associative LTP.

NMDAr-Depedent LTP involves actually pushing new AMPA receptors onto the post-synaptic terminal after Calcium can come in to the NMDA gate and cause a protein kinase cascade. Inserition of new AMPA receptors and morphological changes contribute to LTP.

Late LTP requires protein synthesis.

LTP is spike-timing dependent. If the pre-and post are within a certain window of post following pre, you will get LTP. If pre is too far after post, you’ll start to get LTD.

The hippocampus is necessary for spatial learning. In the worries water maze, hippocampal lesioned animals will use a thigmotradal strategy when they don’t have a hippocampus. They can still reach the dock, but its not from memory.

APV blocks spatial learning which shows that NMDA receptors are necessary for spatial learning via the hippocampus. If you insert APV, the animal will spend equal time in all 4 quadrants of the maze. However, if you don’t insert, the animal will spend the most time in the quadrant where they had previously found the platform.

PKMZeta is a kinase whos mRNA seems to be translated when LTP is induced. Once activated, it contributes to synaptic potentiation by increase the number of AMPA receptors in the postsynaptic terminal. Blocking this will reverse the effects of LTP if you inhibit it 22hr after learning. Enhancing PKMeta before learning will increase learning in a taste aversion task. Enhancing after learning will improve memory for the first test (some extinction present).

The amygdala seems to be crucial for emotionally modulated retention. The Amygdala orchestrates the response to emotional information by sending projections to brain areas involved in motor, autonomic nervous system and neuroendocrine areas of the CNS.

Neurons in the lateral amygdala can develop responses to the fear inducing stimulus.

Synapses from auditory thalamus to amygdala are potentiated in fear conditioned animals. You can see an increase in the gain of input stimulation. That is, the same stimulation will produce higher output current after learning a paired CS-US stimulus.

Decreasing CREB in the Amygdala impairs learning. CREB is a cAMP response element binding protein that serves as a transcription factor. You can rescue memories by placing CREB back into the amygdala by way of a virus to recover it in a certain percentage of cells. The rescue of fear condition can be accomplished by reintroducing CREB to the amygdala to CREB deficient mice.

Enhancing CREB in the amygdala improves learning in wild types. After injecting CREB, you can test for the expression of Arc (an immediate early gene that is a marker of activity). If you give the CS and see the percentage of neurons expressing Arc and, of those, which express the CREB virus, you’ll see that most of the cells that were active (as seen by Arc) were the ones that had the injected CREB. Thereby, you’ve biased the population of neurons that respond to the tone to be the ones that you’ve injected. When the animal sets out to learn something new, the neurons that are the most plastic (the ones with the most CREB) are then the most likely to respond to the CS and bond. Think of this as memory allocation. Biasing injected popularion to be the ones responsible for storing the memory since they are more cabale since they have more CREB and are thus more plastic. Preferentially ablating the CREB overexpressing neurons from this would then isolate the newly formed memory and delete that particular memory. You can do this by finding animals that have dipheria toxin receptors (which only gets expressed in the cells that have CREB(which none of them do unless you inject the virus). Then this virus that enhances the memory double serves as a selective kill-switch because if it’s the case that memory is stored in these CREB cells more selectively than others, you can delete the memory by inserting diphtheria toxin, which will only get uptaken by those receptors.

NeuroComputation

The most straightforward way to think about how neurons get together and perform complex algorithms and encode memories is by a connectionist model. This is where we think of neurons very simply as weighted connections between nodes that can make eachother fire when they receive enough cumulative input. This can

help neurons perform logical functions. “This neuron will fire IF both these others fire or OR this particular neuron fires”.

The input/output of a neuron would just be the sum of its incoming weights on active neurons and its threshold.

Supervised learning is like with reinforcement. You know a label and you continually get exposed and told how and what things are. With unsupervised learning, you need experience in order to mold into what is most naturally optimal (like ocular dominance columns).

During the delay aspect of a working memory task an animal needs to keep in mind something. How is it possible keep a population of neurons online and representing information during a delay? Perhaps you can get the ativity in such a fashion such that the one pattern of activity stimulated itself in a perpetual manner (an attractor model). Once a pattern is achieved, it stays there (a fixed point attractor).

This brings us to auto-associative networkslike the Hopfield networks. These neurons can be though of as having recurrent connections with themselves and eachother. Weights can be adjusted such that if a target state is desired, the weight that caused it remained and the weight that hurt it gets a -1.

With hopfiled nets, you can have a weight matrix that gets multiple by its transpose in orer to get the memory necessary for that network. If you want to add another memory, you can just take the two resulting matrices and add them together. Combining the two allows you to store two memories. Furthermore, recurrent connections (with itself) would be set to 0 and any 0s would be -1.

When you feed in partial states (partial vectors of memories) into the hopefield matrix, you will get back the full vector if the Hopfield has the memory for that vector.

We can also use a delta rule as a form of backpropogation when building a network. You can know the target of your postsynaptic neuron and then compare that to the actual and use this difference as an error term that you adjust the presynaptic neuron by (alongside a scaling lambda). You can do this to build up Logical decision. For instance, you want both pre-neurons to fire and get a target firing in your post neuron if you want to build an AND gate.

However, this is relatively basic and to get at functions like an exclusive or function, you will need to add hidden layers so that you can incorporate and AND and an OR function in order to make a final Exclusive or function. You would backpropogate using the delta rule in a continuous contribution fashion back through all layers of the network.

Furthermore, with hopfileds, errors can occur with bad/over generalization. Furthermore they are not capable of view-invariant pattern recognition.

How can we use some of these ideas to explain the transition of cortical neurons from receiving input from both eyes to only on eye in layer IV?

Homeostatic plasticity refers to the notion that neurons can adjust their synaptic and cellular properties to up or down-regulate their activity levels to achieve some established “set point” of average activity. This is the basic theory of BCM which relies on a sliding threshold of LTP/LTD depending on the history of the cell’s activity. BCM strives to achieve homestasis by increasing the neuron’s threshold for firing based on its previous activity. This is because a neuron that never fires or is always firing is not processing information and is essentially a waste.

Hebb did not account for “when do synapses do down” and “what prevents all neurons from responding to the same stimulus”. BCM takes care of this with its continuously updated neuronal threshold.

The change in synaptic weight is the preactivity multiple by (post activity – some threshold). For LTP, the post needs to be activated at that particular threshold. If the post is above, you get LTP. Below, LTD.

The threshold is a variable that adapts over time depending on how constantly that neuron is firing. A neuron that rarely fires will have a low threshold. The same amount of post synaptic activity can yield a lot of LTP, nothing, or LTD depending on the threshold of that neuron. If we lose a hand, then our threshold for the somato neurons for that hand should go down making it easier to induce LTP in the future, which could be responsible for cortical remapping.

This changing threshold can explain the emergence of neurons with preferred eye input. If we have input into a neuron from 4 synapses and some of those synapses come from the left and some come from the right and we show stimuli back and forth to each eye, then eventually only the synapses that were encoding for one of the eyes will, by chance, win out. This is whichever synapse is reaching the threshold of the post synaptic neuron. Once it reaches that threshold, it gets a little LTP, the threshold increases and this starts a snowball effect that leaves the other synapses behind. The synapses firing at the same time would be rewarded with LTP (which would only be from one eye). Thus, one eye gets the preferential input and the other synapses get pruned away because they will not catch up to the threshold without the LTP that the other synapses enjoyed.

This random weight initialization can lead to selective neurons with a BCM model, but how are they arranged with spatial selectivity? How do these form ocular dominance column…..lateral inhibition. Neighboring neurons will help excite eachother whereas neurons further from eachother will inhibit eachother. If you

mix this concept with the BCM mode,, you can find groupings emerge. The same can be done for orientation selectivity.

With hopefield netwoks, memories are recalled as the network converges to a fixed-point attractor. With the delta rule, no dynamics and no timing and temporal processing involved. However, we interact with our environment in a dynamic and time-varying fashion. So, how does the brain do it? Integrate and fire neurons, perhaps. In a combinatorial cascade of inhibition, excitation, and temporal summation, a neuron can function as a computational device that depends on its current state and a tug-of-war between temporally locked events.

Short term facilitation (short term LTP) can change the timing of the excitation of a neuron in response to another couple of neurons.

Equations:

BCM∆Wpro , post=λ (Vpre ) (Vpost−Thr )

Thr= Act2

Set

Hebbian

∆Wij= λ(Vi)(Vj)

Hopfield Nets

Vj=∑i

N

Wij (Vi)

Wij=∑k

M

T ik (T j

k)

Where Vj greater than 0 will make Vj equal to 1 and less than 0 will make Vj be 0.

Delta Rule

∆Wpro , post=λ (Vpre ) (Tpost−Vpos )

Where T is the target activation desired and V is the actual.

Input-Output Function

Vj=∑i

N

Wij (Vi)