Optic nerve: from eye to optic chiasm. Optic chiasm: where some of the fibers cross. Optic tract: from optic chiasm to lateral geniculate nucleus (LGN) of the thalamus. Optic radiation: from LGN to primary visual cortex (V1).
Information is split apart and processed by separate subparts of the visual system. Already seen a good example of a parallel pathways: rods and cones. Magno- and parvocellular pathways are another example:
30 or more secondary visual areas after V1 in the occipital lobe and
in parts of the parietal and temporal lobes.
These areas are identified in the monkey brain based on both physiology and anatomy (FACT):
Two distinct streams (another example of parallel pathways). Some go to parietal lobe and some go to temporal lobe. Retrograde tracers from parietal/temporal lobes back to occipital find that inputs to parietal/temporal are segregated (parallel pathways). Parietal pathways colored here in yellows and earth tones. Temporal pathways colored in blues and greens.
What do these two pathways do? Clinical observations have provided us with most info on this. Lesions in these 2 different pathways result in very different symptoms.
Parietal: spatial orientation, attention (where
Temporal: recognition (what pathway).
Patients with temporal lobe lesions are aware that there's a problem and they develop strategies to compensate for it. Parietal patients often unaware of their deficits.
Direction selectivity in monkey MT (video)
Record from MT neuron while displaying dots moving in different directions. Audio track allows you to hear the spikes - each click corresponds to an action potential. Neuron is strongly direction selective, responds only when dots move in a narrow range of directions.
Motion sensitivity in human MT
Moving vs stationary dot stimulus evokes particularly strong response in a lateral region near the junction of the occipital and parietal lobes - believed to be the human homologue of monkey area MT.
What kinds of evidence can we use to infer that area MT is functionally specialized?
What kinds of evidence can we use to infer that human MT is homologous to monkey MT? Why is that important?
Columnar architecture: Move electrode vertically through thickness of cortex, find that most neurons have the same selectivity (e.g., same orientation preference and eye dominance). Ocular dominance columns: Move an electrode tangentially through the cortex, first find cells that respond to left eye inputs, then binocular (responsive to both/either eye), then right eye, then binocular, then left again, etc. Orientation columns: Move electrode tangentially in the orthogonal direction, first find cells selective for vertical, then diagonal, then horizontal, etc. Hypercolumn: chuck of cortex about 1-2mm on each side by 3-4mm thick. Contains neurons, all with the same receptive field location, but with all different orientation selectivities, direction selectivities, both (left- and right-) eye dominance.
Ocular dominance columns via optical imaging
Blood flow response measured carefully in animal studies. Open hole in skull, point video camera at it, collect images that reflect the relative amount of oxygenated ("red") versus deoxygenated ("blue") blood. Subtract activity evoked by left-eye stimulus minus activity evoked by right-eye stimulus. Get zebra stripes of left- and right-dominance. The stripes end exactly at the V1/V2 border.
Orientation columns via optical imaging
Analogous experiment but using orientation (e.g., vertical versus horizontal) instead of eye input (right eye vs left eye). Gives picture of the orientation columns. The ocular dominance and orientation columns are laid out in an organized fashion with respect to one another.
Note that these columns are at a spatial scale that is just below the spatial resolution of current fMRI techniques. By changing stimulus orientation you would shift activity from one column to the next, but the sum total activity measurable with fMRI would be the same.
Columnar architecture in other areas
Columnar architecture is a big theme in cortical physiology. Columnar architecture in MT for stimulus motion, neurons with similar motion preferences nearby one another, orderly progression from one motion direction to the next as you move through MT. Area IT has columnar architecture of complex shape/feature selectivity for object recognition.
Much is unknown about cortical circuitry (not surprising given that there is still a lot that is still unknown about retinal circuitry). Even considering just the pyramidal cells in V1, there are many different subtypes with different connections. But there are some general principles that we can rely on...
Excitatory & inhibitory neurons
A typical neuron will release mainly one type of neurotransmitter from its axon terminals (some may also release a slow acting neuromodulator as well). Odd fact that morphology of dendrites is predictive of whether this is an excitatory or inhibitory neurotransmitter.
Neuron pictured above has little spines on its dendrites. Other cell types have smooth dendrites. Turns out that cells with spiny dendrites are excitatory (e.g., releases glutamate), and smooth cells are inhibitory (e.g., releases GABA).
All/most long-range connections from excitatory pyramidal cells. I.e., inhibitory inter-neurons make only local connections.
A given neuron will have many different postsynaptic receptors in its dendrites so that it receives both excitatory and inhibitory inputs.
Excitation & inhibition
Excitatory and inhibitory post-synaptic potentials (PSPs) in cat V1, due to visual stimulation. Intracellular, in vivo recordings.
Na/K pump sets up electrochemical gradient. Ion concentrations differ in/out. Internal potential is -70 mV relative to outside. Dynamic equilibrium with pump offsetting leak. Neurotransmitter binds to postsynaptic receptors and opens up hole in the membrane. That allows ions to flow either in or out with the gradient. Synaptic ion channels are opened by specific neurotransmitters and are selective for particular ions. Examples:
Canonical microcircuit hypothesis: each cortical area conducts computations of the same form using similar circuitry but different inputs to each area convey different functions. In spite of differences in cytoarchitecture, the types, arrangements, and connections of cortical neurons is similar throughout cortex.
In particular, there is clearly a columnar organization to cortical circuits. Neurons (both excitatory & inhibitory) within a column are highly interconnected, so that all the neurons within the column will tend to increase or decrease their firing rates together.
If there are canonical microcircuits, then we should be able to identify the elementary building blocks of neural computation. Won't take time to go into all of this, but here's a partial list of some of the computational functions that each part of a neuron can perform:
- decode spike train
- excitation & inhibition
- synaptic depression & facilitation
Dendrites combine inputs:
- add, subtract, multiply, divide/shunt
- coincidence detection
- spike generation
Axons simply transmit
Orientation selectivity and model
Receptive field and receptive field subregions. Linear, weighted sum of stimulus intensities. Excitation in central elongate subregion. Inhibition in flanking subregions. Preferred orientation: only excitation = big response. Orthogonal orientation: excitation and inhibition cancel out = no response.
This model is attractive because it allows for a complete description. Measure the linear weighting function for a given neuron, and you can predict that neuron's response to any visual stimulus. But...
Response saturation & phase advance
The linear model predicts that response should increase linearly with stimulus contrast (scale the input, scale the output). But that's not what happens. Graph plots PSTH (instantaneous firing rate) for monkey V1 neuron for moving sinusoidal grating stimuli. Each panel corresponds to a different stimulus contrast. As contrast is increased, response stops getting bigger (response amplitude saturation) and the response tends to happen sooner (response phase advance).
Consequence of response saturation is that things should look weird at high contrast...
Invariance with respect to stimulus contrast
Distributed representation of stimulus orientation in V1 at high and low contrasts. Width of the distribution is the same. Ratio of responses is the same.
Failure of invariance of invariance with saturation
Obviously, this is not how the brain works...
The linear model as predicts that response to the sum of two stimuli equals the sum of the two responses to the stimuli presented separately. But that's not what happens. Example: 50 sp/sec for preferred orientation, 2 sp/sec for orthogonal orientation. Superimpose the two. Model predicts 52 sp/sec, but neuron gives 35 sp/sec, almost half the predicted response. Called cross-orientation inhibition. The presence of the orthog stimulus, even though it evokes no response by itself, nonetheless inhibits the neuron's response to the preferred stimulus.
New model, begins with underlying linear stage. Followed by normalization stage, in which each cell's response is divided by a quantity proportional to the pooled activity of a large number of other cells. Activity of the large pool of cells partially suppresses each individual cell. Effect of this is to normalize (rescale) responses with respect to stimulus contrast.
Explains response saturation because divisive suppression increases with stimulus contrast. Explains cross-orientation inhibition because a given neuron is suppressed by many others including those with orthogonal orientation preferences. Simple enough that we can derive equations to fit data (see papers in reader). Havenít explained to you how we get phase advance but that falls out of the model too, when you develop it in detail (see papers in the reader).
Response vs contrast
Ratio of reponses to preferred and non-preferred orientations is constant over full range of contrasts, in spite of response saturation. The pool in the denominator is the same (or at least very similar) for a whole population of neurons. Their responses are all rescaled by the same amount.
Linear summation of inputs, combined with rescaling. Rescaling is key because neurons have a limited dynamic range, limit to how strong a response can be. See the same computation at each of several stages in the visual pathways.
Retina: The light/dark adaptation mechanisms in the retina effectively re-normalize or divide by the average intensity in the image to compensate for (or factor out) the level of illumination. This helps the visual system achieve perceptual constancy: white looks white and black looks black regardless of the level of illumination. Also allows the brain to process visual information without having to deal with the million-fold range of possible intensities.
V1 neurons sum inputs from the retina. Then contrast normalization throws away some of the information about stimulus contrast. This helps the visual system to achieve perceptual constancy: invariant (ratio of responses) representation of stimulus orientation, pattern, motion, etc.. Also, allows later visual areas to process information about shape, motion, etc. without having to deal with contrast.
MT neurons are believed to sum inputs from a subset of direction-selective neurons in V1, then rescale. This throws away information about stimulus pattern, but achieves invariant representation of motion.
Simultaneous brightness contrast
This illusion can be explained by what we know about the visual processing in the retina. Retinal responses depend on the local average image intensity. On the right, the background is black so the average intensity there is pretty small; retina divides by a small number yielding a brighter percept. On the left, however, the background is light so the average intensity there is pretty large; retina divides by a large number yielding a darker percept.
Simultaneous contrast contrast
Analgous thing with contrast. When texture patch is surrounded by a high-contrast pattern, the bright points of the texture patch appear dimmer, and simultaneously, its dark points appear lighter. To explain this illusion, we assume that perceived contrast is a monotonic function of V1 activity. In this particular case, we are assuming that the perceived contrast of the texture patch is mediated by those neurons with receptive fields centered on the patch. The normalization model posits that each neuron is suppressed by the pooled activity of a large number of cells, including those with receptive fields in a surrounding spatial region. The responses of these surrounding cells increase with background contrast. Suppression is stronger (and perceived contrast of the texture patch is lower) when there is a high contrast background.
Target pattern is harder to detect when superimposed on a high contrast background. Can be explained by contrast normalization. Assume subject monitors the response of the V1 neuron that is most sensitive to the target, and that target is correctly detected when that neuron's response is greater than some criterion. Response of that neuron increases with target contrast, but suppressed by superimposing the background (masker) pattern. In presence of masker, target must have higher contrast to evoke criterion response.
Task: 2AFC, which one has higher contrast? Determine the contrast increment that is just barely detectable (e.g., 75% correct).
fMRI responses in V1 are consistent with contrast discrimination thresholds
The smooth curves were simultaneously fit to both the fMRI and psychophysical data under the hypothesis that a contrast change is detectable when the brain activity increases by a criterion amount. At low contrasts, the slope of the V1 activity is steep, so a small contrast increment evokes a criterion response increment. At high contrasts, a much larger contrast increment is needed to evoke a criterion response increment. Variants of this hypothesis have served as the basis for interpreting psychophysical data for over a century (Fechner). The fMRI measurements provide additional data that help to constrain the interpretation.
Conclusion: contrast discrimination performance is limited by signal-to-noise at or before V1.
Signal detection theory
Hypothetical internal response curves: the horizontal axis is labeled internal response and the vertical axis is labeled probability. The height of each curve represents how often that level of internal response will occur, for that stimulus condition.
Decision rule: on each trial, you get a draw from each of the two internal
response distributions and pick the one with the largest response. When
the contrast difference is small (left), there will be a lot of errors
because it is quite likely that the response to the lower contrast will
exceed the response to the higher contrast. When the contrast difference
is large, there will be few errors because the higher contrast will almost
always evoke the greater response. The threshold (75% correct) corresponds
to when the separation between the two distributions equals their spread
(middle). I.e., when the internal response increases by a fixed criterion