Altklausur + Self-Assesment

• Time Coding = detailed firing structure of response

• Frequency Coding = number of action potential per time unit

Within one module:

• aligned orientation

• similar grid scale

• spatial phase shifts between grid cells

• single grid cell module can not represent spatial location

-> without 1. or 2. one module could encode space

Neighboring grid cells:

• similar spatial freq

• similar orientation

• different phase

Correct: 3., 4., 7.

Rest is incorrect, bc

• cortex has 6 layers

• nuclei in the brain are clusters of neurons

• some animals (e.g. jellyfish) have neurons but no centralized brain

• right and left hemisphere are connected via corpus callosum

E = 58 x log10(100) = 58 x 2 = 116 (mV)

Due to inactivation mechanisms:

sodium channels open -> Na+ intake -> depolarization -> inactivation by blocking further Na+ preventing further depolarization and allowing neuron to return to its resting state of negative voltage -> it is closed until next spike

Action potential is propagated by saltatory conduction, where the electrical signal jumps between nodes of Ranvier (= gaps in the myelin sheath).

At each node, voltage-gated Na⁺ channels open, regenerating the action potential.

Conduction speed increases by reducing ion leakage and membrane capacitance along the myelin insulation.

Information is defined as the reduction of uncertainty.

-> neuron's response conveys more information when it reduces uncertainty about the external world

-> greater neural selectivity or variability in response to different inputs increases the information content, because it helps distinguish between possible causes of sensory input.

1. For equiprobable states uncertainty grows with the number that states

2. Uncertainty should be continuous in probability of states

3. “Branching” property (let’s discuss some other time)

Shannon’s entropy:

Entropy of a system = measure of uncertainty about it’s state

-> describes the capacity of a system to carry information

1. early developement: activity-independent organization mechanisms where brain wires itself using chem. signals

2. afterwards, activity-dependent plasticity refines these connections based on experience and neural activity

3. Spontaneous activity in immature circuits for axonal refinement even before sensory input is fully available

4. Over time, synaptic pruning and experience-driven modifications to further sculpt circuits

5. Activity and Connectivity influence each other:

   -> activity via synaptic plasticity forms conntectios & connection dynamics allow activity at different strengths

6. can happen at different scales: synaptic or circuit

Hyperpolarized and graded potential

because it actively & dynamically filters and modulates visual information before sending it to the cortex

simple cells = elongated spatial RFs with neighboring on/off subfields

-> selective for orientation

How it works:

-> best response when a stimulus is at the right orientation and location within RF

-> RF have distinct ON and OFF subregions

-> respond to light or darkness

Spatial contrast, temporal contrast, spectral contrast

Dorsal (parietal) stream (“where”): motion and spatial perception/depth → spatial location

Ventral (temporal) stream (“what”): color, form, fine detail → object recognition 

decode an animal’s location by recording from spatially tuned neurons like place cells or grid cells, which fire at specific positions

Decoding methods:

• Bayesian decoding (uses prior knowledge of each cell's tuning curve to estimate the most likely position)

• Population vector decoding (treats the firing rates as weighted vectors to compute a position estimate)

Principle of experience-dependent correlated activity = Neurons that fire together get wired to gether (Hebbs Postulate)

Context of replay of neural activity in sleep:

in slow-wave sleep, the brain replays activity patterns in a compressed and coordinated way, reflects the same sequence of place cell activation seen during waking behavior.

The correlated replay strengthens synaptic connections, supporting memory consolidation.

-> sleep helps stabilize and store experiences by reactivating and reinforcing learned patterns

Dedicated models:

assume that specific neurons or circuits are specialized to encode particular features or functions

-> fixed, labeled lines

Intrinsic models:

suggest that coding arises from the dynamics and interactions within a general-purpose network, without needing predefined feature-selective units.

-> emergent activity patterns

-> often more flexible and context-dependent

= discriminality of two nearby stimuli depends on ratio between intensities and not absolute magnitude

-> reflects how our perception scales with stimulus intensity.

Pacemaker–accumulator model =

Internal pacemaker generates pulses at a regular rate.

These pulses are sent to an accumulator when attention is directed to timing.

The total number of accumulated pulses is used to estimate elapsed time.

This model helps explain how we compare durations and why our sense of time can vary with attention or arousal.

The circadian transcriptional circuit is a molecular feedback loop to generate ~24-hour rhythms using a gene expression cycle.

The circuit is within the suprachiasmatic nucleus in the hypothalamus and gets input form the retina

Difference-of-Gaussians (DoG) model

= represents how retinal ganglion cells respond to light in their receptive fields

models the center (excitatory) and surround (inhibitory) regions as two overlapping Gaussian functions

By subtracting the surround Gaussian from the center Gaussian, the model captures the antagonistic center-surround structure

by emphasizing changes and differences in light rather than absolute intensity

• Center-surround receptive fields filter out uniform areas and highlight contrast, removing spatial redundancy

• Temporal adaptation allows cells to focus on changes over time, ignoring constant stimuli

• Lateral inhibition sharpens edges and decorrelates neighboring signals

= methods to reduce statistical dependencies between input signals to make information more efficient and less redundant

Methods:

1. Principal Component Analysis (PCA) – removes linear correlations by rotating the data to align with axes of maximum variance; outputs uncorrelated components but may still be statistically dependent.

2. Zero-phase Component Analysis (ZCA) – similar to PCA but preserves the original data orientation; results look more like the input, useful for early vision models.

3. Independent Component Analysis (ICA) – goes further by making components statistically independent, not just uncorrelated; good for separating mixed sources (e.g. natural scenes).

differ mainly in how far they go: PCA decorrelates linearly, ZCA also whitens but preserves structure, and ICA removes higher-order dependencies.

in the primary visual cortex (V1)

= neuron's selective response to a specific range or type of stimulus (e.g. orientation, frequency, location, or direction)

-> A neuron is "tuned" to the stimulus that elicits its strongest response

= way the brain represents information by combining the activity of many neurons, rather than relying on a single neuron

-> Each neuron contributes a part of the information based on its tuning curve, and the overall pattern across the population encodes the stimulus.

Coding precision in a population code depends mainly on:

1. Number of neurons involved:

   -> the more neurons, the higher the precision.

2. Tuning curve sharpness

   -> narrower tuning means neurons respond more selectively, improving accuracy.

3. Neuronal variability (noise)

   -> less noise leads to more reliable signals.

4. Correlation between neurons

   -> lower correlations typically increase coding efficiency.

1. Non-uniform distribution of tuning preferences, where neurons are unevenly tuned across stimulus space.

2. Noise and variability in neuronal responses skewing the decoded estimate.

3. Correlations among neurons that distort the combined signal.

4. Inaccurate or incomplete decoding models that fail to capture true neural dynamics.

a) In the striate cortex (V1), population coding occurs in the combined activity of orientation-selective neurons, where the pattern of firing across many neurons represents the orientation of edges in the visual scene.

b) In extrastriate cortex (like area MT), populations of direction-selective neurons encode the overall motion direction of objects by integrating signals from V1 and other areas.

20-30%

= rapid shift of a neuron's receptive field to a new location in anticipation of an eye movement (saccade).

-> It helps maintain stable visual perception despite frequent gaze shifts.

-> observed in areas like the parietal cortex and frontal eye fields.

24h time frame (night-day)

suprachiasmatic nucleus in the hypothalamus

hundreds of milliseconds to several minutes

Examples:

1. Temporal Production: The subject generates a specific time interval (e.g., pressing a button for 10 seconds).

2. Temporal Reproduction: The subject observes a time interval and then tries to reproduce it.

3. Temporal Discrimination: The subject judges which of two intervals is longer.

4. Temporal Estimation: The subject estimates the duration of a presented interval verbally or by other means.

= tendency of people to overestimate small stimuli and underestimate large stimuli, causing their responses to regress toward the mean of the presented range.

This bias reflects imperfect scaling of perceived magnitude.

using Bayes’ theorem:

where:

• P(M|S) = likelihood of measurement given stimulus,

• P(S) = prior probability of the stimulus,

• P(M) = marginal probability of the measurement.

This method is called Bayesian inference.

= neural mechanisms where the time is represented by the evolving activity pattern across a population of neurons.

(Instead of relying on a single pacemaker, timing emerges from the dynamic, time-dependent changes in population activity, allowing the brain to track intervals and temporal sequences.)

A state-dependent network consists of:

1. Neurons with intrinsic dynamics, whose current activity depends on their previous state.

2. Recurrent connections, allowing feedback and interaction among neurons.

3. External inputs that drive the network and interact with its internal state.

Together, these components create time-dependent patterns of activity that encode information based on the network’s evolving state.

by encoding temporal information in its changing internal activity patterns.

As the network evolves after a stimulus, different network states correspond to different elapsed times, allowing the brain to read out time intervals from these dynamic activity trajectories without needing a dedicated clock.

Possible advantages:

1. the ability to test for unique cognitive effects – certain things cannot be tested well in animals, example Numerosity, or speech.

2. training and communication – training is faster because one can communicate with the human subject and describe the desired task, this cannot be done in animals. In humans it avoids overtraining effects.

3. The whole brain – with the techniques that we have, we can measure at a mid-level of spatial and temporal scale brain activity patterns across the entire brain.

Possible disadvantages:

1. The measured signal is not a direct reflection of neural activity, but an indirect reflection of neural activity. This is particularly problematic for comparisons across different populations where a change in the vascular response would be expected (e.g. age).

2. The typical methods for measuring neural activity in animals are unethical in humans,

3. The methods are costly and time intensive.

4. The temporal and spatial resolution is to coarse to see individual neurons or circuits.

PET: spatial resolution on the order of 5-10 mm, and temporal resolution of minutes to hours

EEG: spatial resolution on the order of centimeters and temporal resolution of muss less than a ms

Single cell recordings: high spatial (μm) and temporal resolution (<< ms)

The second question cannot be addressed with human MRI because the spatial scale of the measurements (mm) is too coarse to be able to resolve a single neuron and the fMRI signal cannot resolve inhibitory and excitatory activity

The Larmor equation:

-> ω0 = larmor frequency of a particle (resonant frequency)

-> ɣ = particle’s gyromagnetic ratio (ratio between the particle’s magnetic moment to its angular momentum)

-> B0 = strength of the magnetic field,

Relevance for MRI:

-> allows precise tuning of radiofrequency pulses to match the resonance frequency of hydrogen nuclei in tissue

-> This is how the MRI system selectively excites and detects signals from specific regions, enabling spatial encoding and image formation.

1. A strong static magnetic field: required to set up the atomic particles to receive energy

2. A radio-frequency pulse (from a radio transmission coil): sending energy into the system at the particle’s resonant frequency to be measured.

3. Magnetic gradients in the 3 spatial planes: turned on and off at specific times within the sequence to either select the exicted space, or encode the spatial dimension through the frequency or the signal phase.

= damped sinusoidal signal, also called a free induction decay (FID) signal, with a frequency that is centered on the Larmor frequency for hydrogen protons.

Differences in brain tissue are measured by tissue specific differences in the speed of decay of the signal (or recovery of the longitudinal magnetization) that are reflected in the brightness of the individual voxels of the image.

A magnetic gradient, that is perpendicular to the 2D slice to be measured, is turned on at the same time as the radio frequency pulse is given. This changes the strength of the main magnetic field along the direction of the 2D slice. Then only those protons in the 2D plane where the Larmor frequency (based on the static magnetic field strength) matches the given radio frequency will be excited.

BLOOD OXYGEN LEVEL DEPENDENT (BOLD) signal.

Neuron activation

-> requiring energy from oxygen in blood

-> dilation of local blood vessels

-> more blood is send to that area than needed

-> increased amount of oygenated hemoglobin with less magnetic property than deoxygenated hemoglobin

-> MRI signals gets stronger in that region

-> change measured in fMRI

in longer:

(When neurons become active, they need energy, which requires oxygen. Oxygen is delivered by the blood. In response to this activity, local blood vessels dilate, sending more oxygen-rich blood to the area than is actually needed. This increases the amount of oxygenated hemoglobin, which is less magnetic than deoxygenated hemoglobin. As a result, the MRI signal becomes stronger (brighter) in that region. This change is what fMRI measures, and the system gradually returns to baseline after the activity ends.)

Study showed: both single-neuron activity and fMRI signals in the visual cortex of monkeys.

-> When the animals saw a checkerboard pattern, the same brain area showed both neuron firing and increased fMRI (BOLD) signal.

Further analysis revealed that the fMRI signal matched best with local field potentials (LFPs), which reflect local synaptic input rather than long-distance action potentials. This suggests fMRI primarily reflects input and processing in a brain region, not just output.

In the experiment, researchers presented a short stimulus once, twice, and three times, with short pauses in between. They subtracted the BOLD signal for one stimulus from that of two, and then that from three, and aligned the resulting curves. The curves looked nearly identical, showing that BOLD responses add up linearly, even if they overlap in time.

This linearity means we can use a single, standard BOLD response (called the impulse response function) to predict the brain's response to any stimulus pattern.

-> simplifies fMRI data analysis and modeling.

The hemodynamic response function (HRF) rises and peaks around 6 seconds, then falls and dips slightly below baseline around 12 seconds before returning to normal.

It’s typically modeled as the difference of two gamma functions—one for the rise, one for the undershoot.

In fMRI analysis, this HRF is used as an impulse response function, and researchers convolve it with the stimulus timing to predict the expected BOLD signal.

Yes, we can measure regional brain activity when we are at rest.

This activity is measured by examining how the time courses of brain activity correlate between different voxels.

The assumption made here is that the correlated brain activity between regions means the regions are connected, and not for instance, that some external stimulus caused this activity.

The most common set of brain regions (or network) found at rest is the Default Mode Network.

= each neuron produces a distinctive extracellular waveform (shape and timing) across multiple recording channels, from which features spikes from different neurons can be classified and separated, allowing identification of individual neurons' activity from the mixed signals.

They synchronize the activity of excitatory neurons creating rhythmic patterns

(-> regulates the timing and pacing of neural firing, creating rhythmic patterns that underlie oscillations like gamma and theta rhythms essential for information processing.)

= Hippocampal place cells change their firing in new or altered environments, reorganizing the spatial map to flexibly represent different contexts.

Excitatory synaptic input to the distal dendrite of a pyramidal cell generates a current dipole with a sink at the dendrite (where positive ions enter) and a source near the soma (return current).

-> creates a current dipole polarity oriented from the dendrites toward the soma.

= reactivation and replay of neural activity patterns from prior experiences during rest or sleep.

-> helps transfer memories from the hippocampus to the cortex, strengthening and stabilizing long-term memory storage.