6&7 Visual System

Retino-geniculo-cortical pathway

Retinal ganglion cells (RGCs) - lateral geniculate nucleus of the thalamus (LGN) - primary visual cortex (V1)

opsins

photoreceptors hyperpolarize by light

• Fovea = primate specialization

• 99% of the cones reside in the fovea

• 1% of the retina, 50% of RGCs

• In fovea: cell and synaptic layers are displaced peripherally, fovea devoid of blood vessels

• Cones are less sensitive but faster than rods

• Rods are densest in areas adjacent to the fovea

region from which a stimulus can activate a neuron

for example center-surround receptive field:

RGCs do not simply respond to light but analyze spatial patterns

• Center response: determined by type of glutamate receptor in bipolar cells

• OFF bipolar: follow sign of photoreceptors, i.e. are hyperpolarized by light (= respond most to dark spot)

• ON bipolar: sign inversion between photoreceptors and ON bipolars, i.e. are depolarized by light (= respond most to bright spot)

• Photoreceptors (rods - most sensitive, most numerous; cones - less saturation, color vision, dense in fovea)

• Photoreceptors are hyperpolarized by light

• Parallel processing starts with bipolar cells (discussed: ON/OFF center, but there are many more different bipolar cell types, and even more ganglion cell types)

• RGCs send representation of extracted features via the optic nerve to various areas in the brain

dLGN = obligatory relay station before cortex

Macaque has almost as many LGN cells as V1 cells

—> Thalamus: First stage with massive neuromodulation of visual processing

• Stronger visual responses

• More desynchronized activity

• Better visual performance (until a certain point)

• what and where pathways

• highly selective neurons, e.g. neuron that responses to the image of a hand

Parasol ganglion cells show phasic response behavior: They respond transiently (temporär) to stimulus changes and thus reduce redundancy (das Vorhandensein von Überflüssigem) by removing temporal correlations

Small bistratified ganglion cells are specialized for the encoding of spectral information. They show S-ON/LM-OFF opponency and thus reduce redundancy by removing correlations in the wavelength domain

Efficient representation of sensory information

- Coding with minimal activity → sparse code

- Coding with minimal redundancy → factorial code

Finding Sparse Image codes

• natural images have a sparse structure

• can be represented by a small number of components out of a large set

• came up with basis functions which have a localized and oriented structure

Decorrelation 1: Zero-Phase Whitening

• ZCA decorrelates pixelwise, preserving the structure (higher-order statistics) in the outputs.

• W is the square root of the inverse of the stimulus covariance matrix. Filters are symmetrical, typically localized center-surround (DoG-like) filters

Decorrelation 2: PCA

• PCA decorrelates by rotating to axes of highest variance. The rotation matrix E is the matrix of eigenvectors of the covariance matrix, D the diagonal matrix of eigenvalues.

Decorrelation 3: Independent Component Analysis (ICA)

• constraint: components are independent (not like PCA)

• broad overlapping tuning curves

• each stimulus value is represented by several neurons - robust

• ubiquitous coding principle in nervous system

with a population vector

• example: encoding arm movement direction in the motor cortex

• responses in motor cortex are direction specific

• direction of the arm: max of tuning curves

Probabilistic models

• orientation/colour specific neurons

• Preferences are continuously distributed, tuning curves overlap strongly

• Popullation code

• Distributed representation with overlapping tuning curves

• Attenuation of responses of neurons tuned to stimulus values similar to the cont

• The extrastriate cortex located next to the primary visual cortex

• In primates, 30-50% of the cerebral cortex is concerned with the processing of visual information

• eye movements

Efficient Coding

• Neural processing in the retina reduces redundancy by decorrelating in space, time, and wavelength

• Efficient coding principles can explain orientation selectivity in primary visual cortex

Population Coding

• Population coding is ubiquitous coding principle in the nervous system

• Context-dependent processing can be implemented by response modulation in a population code

  
  

Visual Cortex

• Visual processing in extrastriate visual cortex is organized in two main processing stream

Difference-of-Gaussians (DoG) model

1. Through lateral inhibition, which removes spatial correlations (decorrelation in space).

2. Temporal contrast encoding by parasol cells reduces temporal correlations.

3. Spectral contrast encoding by small bistratified ganglion cells reduces wavelength correlations

• ZCA (Zero-Phase Whitening): Preserves spatial structure; uses symmetrical filters (DoG-like).

• PCA (Principal Component Analysis): Rotates data to axes of highest variance using orthogonal transformation.

• ICA (Independent Component Analysis): Aims for statistical independence of components, yielding localized and oriented filters similar to V1 cells.

In the primary visual cortex (V1)

It refers to the selectivity of a neuron to a specific stimulus value, typically shown by a tuning curve—the neuron’s firing rate as a function of the stimulus parameter (e.g., orientation)

Population coding is a neural representation where stimulus information is encoded across the activity of many neurons, each with overlapping tuning curves.

1. Sampling density (how many neurons cover the space)

2. Tuning width

3. Response variability (noise)

1. Population Vector: Weighted vector sum based on firing rates.

2. Maximum Likelihood (ML) Estimation: Selects stimulus that most likely produced the observed response.

3. Maximum A Posteriori (MAP) Estimation: Like ML but includes prior probability of the stimulus.

1. Contextual interactions can modulate neuronal responses, causing biases in decoding (e.g., tilt illusion).

2. Overlapping tuning curves and surrounding stimuli may shift the perceived stimulus away from the actual value.

a) Striate cortex (V1):

• Orientation and color coding with overlapping tuning curves.

b) Extrastriate cortex:

• Motion direction coding in dorsal stream areas like MT.

• Object/form recognition in ventral stream areas like V4 and IT.

30–50% of the cerebral cortex in primates

• Dorsal Stream ("Where"/"Action"):

  • Processes motion, spatial location

  • Involves large receptive fields, motion tuning, eye movement effects

• Ventral Stream ("What"/"Perception"):

  • Processes object form, faces, color

  • Involves V4, TE, STS

Receptive field remapping refers to the phenomenon where neurons temporarily respond to stimuli at their postsaccadic location before a saccade. It is observed in the dorsal stream, particularly related to mechanisms of visual stability (e.g., in LIP, MST).