Theory

How does sequential mode acquisition in deep linear networks accounts for progressive differentiation during semantic development.

The network decomposes task covariance into SVD modes. Modes with larger singular values (representing broad categorical distinctions) turn on early in training. Finer distinctions (smaller singular values) turn on later, reflecting staged cognitive maturation (“aha” moments).

How does sequential mode acquisition in deep linear networks account for temporary illusory correlations (overgeneralizations) during semantic development.

When broad modes activate before fine modes, the model uses partial low-rank solutions. It assigns features associated with a general category (e.g., “can fly”) to all items within that broader branch (e.g., predicting that a “salmon” can fly), creating temporary overgeneralizations that correct only when finer modes activate.

Define the concept of a “double dissociation” using the classic clinical cases of Phineas Gage (frontal lobe damage) and Patient H.M. (medial temporal lobe damage).

Double dissociation occurs when two distinct lesions have different effects on two distinct variables

1. Gage lost self-control/planning but kept memory

2. H.M. lost episodic memory but kept self-control/working memory

3. Together, Manipulation A affects X not Y, while B affects Y not X

Explain structurally and logically why demonstrating a double dissociation is considered necessary when attempting to map cognitive functions to specific brain architectures.

1. A single dissociation only proves necessity

2. It does not rule out the “task difficulty” confound, where a harder task falls away first under generalized damage

3. Double dissociation proves parallel/independent neural systems, eliminating the difficulty confound

Describe the hub-and-spoke model of semantic cognition.

1. Spokes are modality-specific sensory/motor regions

2. The Hub is in the anterior temporal lobe (ATL)

3. It is transmodal, linking all spokes

Contrast how a concept like “coffee” would be represented in a purely distributed model versus the hub-and-spoke model

1. Purely distributed: Visual input of coffee connects directly to smell and motor actions

2. Hub-and-spoke: Sight, smell, and word all independently activate a shared ATL pattern, which projects outward to other spokes

Why is a transmodal hub computationally necessary for true semantic generalization?

1. Distributed models rely on superficial correlations

2. A hub creates a computational bottleneck that learns deeper, non-linear conceptual similarities (e.g., whales vs. fish) beyond raw sensory overlap

Contrastive (CLIP)

1. Two separate encoders map inputs to a shared space

2. Softmax loss maximizes cosine similarity for matches and minimizes it for mismatches

3. Mimics hub slightly but suffers from a “modality gap” keeping modes separated

Generative (Denoising Autoencoders)

1. Corrupts/masks input, maps through bottleneck (encoder), rebuilds clean input (decoder)

2. Uses reconstruction error (e.g., MSE) loss

3. Highly biologically plausible; encoder mimics forward connections, decoder mimics top-down predictions

Contrastive model and a generative predictive model learn multimodal representations. Detail their loss functions and discuss how well each architecture mimics the biological transmodal semantic hub. State which of the models is a better fit for the hub-and-spoke architecture.

1. CLIP learns by distinguishing instances; autoencoders learn data distributions to fill in missing parts

2. Autoencoders better mimic the ATL because they force true transmodal fusion to infer missing modalities

The McGurk illusion requires automatic integration of visual and auditory phonemes. Based on the hub-and-spoke model, hypothesize what would happen if the McGurk illusion were tested on a patient with advanced semantic dementia (severe anterior temporal lobe atrophy).

The illusion would be weakened or removed.

why:

1. The McGurk effect relies on transmodal binding of visual and auditory spokes

2. Semantic dementia destroys the ATL transmodal hub, preventing the merging of conflicting sensory streams

3. The patient would likely perceive the raw auditory signal without visual top-down correction

Gestalt principles of vision: Law of Similarity, describe what the property entails

Elements that share visual characteristics (e.g., color, shape, size, orientation) are perceived as belonging together or forming a unified group or pattern.

Gestalt principles of vision: Law of Similarity, describe why it is useful within a visual system.

Enables fast surface/texture segmentation and feature grouping, allowing the visual system to distinguish distinct objects or regions from heterogeneous backgrounds.

Gestalt principles of vision: Law of Pr¨agnanz, describe what the property entails

Complex or ambiguous visual inputs are interpreted in the simplest, most stable, and least computationally complex structure possible

Gestalt principles of vision: Law of Pr¨agnanz, describe why it is useful within a visual system.

Acts as a global parsimony heuristic (prior) that resolves ill-posed perceptual inverse problems, preventing over-interpretation of noisy visual input.

Gestalt principles of vision: Law of Proximity, describe what the property entails

Objects or features that are spatial neighbors are grouped together into coherent perceptual units compared to those spaced farther apart.

Gestalt principles of vision: Law of Proximity, describe why it is useful within a visual system.

Exploits spatial locality in physical scenes, assuming that nearby visual features likely belong to the same physical object or entity.

Gestalt principles of vision: Law of Continuity, describe what the property entails

Visual elements arranged along a smooth line or continuous curve are perceived as belonging together, favoring smooth paths over abrupt directional changes.

Gestalt principles of vision: Law of Continuity, describe why it is useful within a visual system.

Facilitates boundary detection and contour integration across occlusions or noisy environments, preserving object identities across intersecting shapes.

Gestalt principles of vision: Law of Closure, describe what the property entails

The visual system fills in missing contour information or gaps in an incomplete figure to perceive a whole, enclosed shape or boundary.

Gestalt principles of vision: Law of Closure, describe why it is useful within a visual system.

Crucial for robust object recognition under partial occlusion, shadows, or degraded visual input, maintaining object constancy.

Functional roles, visual processing properties, and downstream projection targets of the dorsal and ventral visual streams: Ventral Stream Identification

1. Known as the “What” pathway or vision for perception.

2. Specialized for object identification, fine detail, and recognition.

3. Projects from primary visual cortex (V1) down through the inferior temporal (IT) cortex.

Functional roles, visual processing properties, and downstream projection targets of the dorsal and ventral visual streams: Dorsal Stream Identification

1. Known as the “Where” or “How” pathway or vision for action.

2. Specialized for spatial location, motion, and guiding physical motor interactions.

3. Projects from V1 up through parietal areas

Functional roles, visual processing properties, and downstream projection targets of the dorsal and ventral visual streams: Functional Contrast & Accessibility

1. Ventral stream processing is consciously accessible (conscious perception).

2. Dorsal stream processing operates largely unconsciously to guide real-time motor control.

Explain why damage to the dorsal stream results in optic ataxia while damage to the ventral stream leads to visual agnosia

1. Optic Ataxia (Dorsal Damage): Impairs the ability to use visual guidance to execute targetdirected motor movements (reaching/grasping), even though object recognition remains intact

2. Visual Agnosia (Ventral Damage): Impairs the ability to consciously perceive, describe, or recognize objects, while the patient retains the ability to accurately guide motor interactions toward those same objects.

Simple-cell receptive fields in primary visual cortex (V1) can be modeled as oriented Gabor filters with a Gaussian spatial envelope modulating a sinusoidal grating.

1. Sinusoidal Grating: Determines the cell’s preferred orientation, spatial frequency, and phase selectivity, acting as an edge/stripe detector.

2. Gaussian Spatial Envelope: Restricts the filter to a finite subregion of the visual field, defining the local spatial boundaries of the receptive field.

Detail the trade-off governed by the spatial envelope size regarding frequency/orientation tuning versus spatial localization.

1. Large Spatial Envelope: Provides narrower (more selective) frequency and orientation tuning, but yields poorer spatial localization.

2. Small Spatial Envelope: Provides superior spatial localization, but results in broader (less selective) frequency and orientation tuning.