Generated Deck + Lecture 4

Decoding maps brain activity to stimulus labels (Brain → Stimulus), whereas Encoding maps stimulus features to brain activity (Stimulus → Brain).

An encoding model specifies how features are transformed into neural responses and can predict responses to novel inputs, whereas decoding only indicates that information is present.

The receptive fields of orientation-selective neurons in the primary visual cortex (V1).

Because the model cannot generate an image from scratch; it predicts brain activity for a set of candidate images and selects the one that matches the observed activity best.

The candidate set consists of real YouTube clips, so the reconstruction inherently reflects the statistics and correlations of real-world video (e.g., shapes and motions), even if the model doesn't explicitly encode them.

A function (usually Gaussian) describing how a neuron's firing rate peaks at a "preferred numerosity" and drops off as the quantity moves away from that preference.

A spatial organization where neurons with similar preferred numerosities (e.g., 2 and 3) are located physically adjacent to each other, creating a gradual gradient across the cortical surface.

By comparing the voxel's fMRI time course to predicted time courses generated by models with different peak numerosities; the model yielding the best fit defined the preference.

The maps do not show the same tuning or topographic organization, indicating these regions encode perceptual quantity (how many items), not abstract symbolic concepts.

It is a network of frontal and parietal regions that responds to difficult tasks (math, logic, working memory); fMRI shows that activity in these regions scales with task load (e.g., number of items in memory).

Each unit processes only a small, local neighborhood of inputs (pooling), mimicking the restricted "receptive fields" and local connectivity of biological visual neurons.

Early layers detect simple features like edges and orientations, while deeper layers detect complex conjunctions, shapes, and eventually whole object categories (e.g., faces).

It reveals that higher layers look for specific complex patterns (like eyes or animals) and will hallucinate these structures into an image if forced to maximize their activation.

The network separates content (low-level features) from style (high-level correlation statistics) and applies the style statistics of an artist (e.g., Van Gogh) to the content of a new photograph.

Standard DCNNs are largely feedforward (input → output), whereas the human brain relies heavily on recurrent (horizontal) and feedback connections.

"Machine learning for cognitive modelling: biologically inspired models," which focuses on using machine learning to simulate and understand how the human brain processes information.

Statistical Similarity (Decoding), which looks for patterns to guess inputs (e.g., "Is this a cat?") without explaining the underlying brain mechanism.

Encoding, which builds models based on known brain properties to explain how information is processed and written into the brain.

Decoding asks "Can we guess what image you see?" (output focus), while Encoding asks "How does the brain process this image?" (process/mechanism focus).

Structure (organizing the model like the brain), Encoding (learning how info is written), and Prediction (predicting reactions to new stimuli).

Visual space, Object number (numerosity), Event timing, Object identity, and Deep Convolutional Neural Networks (DCNNs).

It starts in the primary visual cortex (V1), where cells are "orientation-selective" (picky about angles).

They combine signals from LGN cells (which have circular ON/OFF fields) and only fire if they see a line at a specific position and angle.

Shutterstock
Explore

The computer looks at images and brain scans, breaks images into edges, and uses linear regression to calculate "weights" for how much each brain chunk (voxel) likes specific edges/angles.

It uses the learned weights to predict brain activity for a new image and compares that prediction to the real scan; the best match is the identified image.

The model identified the correct image much better than random guessing, even if the person only saw the image once (single trial).

The model uses a library of real-world clips (priors) to make guesses, assuming things like "faces are skin-colored" or "sky is blue" even if the brain scan is incomplete.

Numerosity (number of items), Size (physical dimension), Timing (duration/frequency), and Task Load (mental effort).

A neuron that has a "preferred numerosity," meaning it fires most strongly for a specific number (e.g., 3) and less for others.

Getty Images
Explore

It is typically bell-shaped, peaking at the neuron's preferred number and dropping off as the number gets further away.

The AI made errors by guessing numbers close to the real one (e.g., mistaking 4 for 3), implying that the brain representations for these numbers are physically next to each other.

A specific network of brain areas (including the Intraparietal Sulcus and frontal areas) that activates during difficult cognitive tasks like math or programming.

That the brain organizes abstract concepts (like math/numbers) into physical topographic maps, just like it does for senses (vision/touch).

Ultra-high-resolution fMRI (7 Tesla) combined with biologically inspired modeling while participants viewed a changing dot pattern (1 to 7 dots).

Just as piano keys are laid out low-to-high, brain activity physically moved across the parietal lobe surface as the number of dots increased, proving a topographic map exists.

The Preferred Number (which number the neuron loves) and the Tuning Width (how picky or broad the neuron's preference is).

Because fMRI measures blood flow (which is slow) rather than electricity (which is fast), so the prediction needed a "delay" to match the data.

A "rainbow" pattern where neurons preferring 1 dot are next to those for 2, then 3, all the way to 7, showing a continuous physical layout.

Yes, they found Timing Maps (Harvey et al., 2020), where neurons preferring short durations sit next to those preferring long durations.

The visual quantity maps did not respond; these maps only care about the visual impression of quantity (dots), not learned symbols.

In a specific zone deep in the Intraparietal Sulcus (IPS), physically sandwiched between the "visual quantity" map and the "visual symbol" area.

That humans "recycled" older visual processing areas and connected them in a new way to create a "calculation zone" for math.

The difficulty of the task, specifically how many items (bars) the participant had to remember (e.g., 1 vs. 6).

Topographic maps found in the frontal/parietal lobes where neurons are organized by preferred difficulty (Easy Task neurons next to Hard Task neurons).

The idea that harder tasks don't just make the same neurons work harder; they recruit different neurons located elsewhere on the difficulty map.

Because humans are experts at it, but it is computationally complex compared to simple lines or numbers.

Specific patches that respond very strongly to faces but very weakly to other objects (like houses), making them easy to "decode" but hard to explain mechanistically.

The ability to recognize an object (like a cat) as the same thing even if it is upside down, in the dark, or seen from a different angle.

The theory that the brain builds objects in stages (Lines -> Shapes -> Objects). It failed because researchers never found "Shape/Geometric" neurons in the middle stages (V2/V4).

They use complex textures and non-linear patterns. Layer 1 detects edges (like V1), and Layer 2 detects textures/correlations (like V2).

They automatically developed internal "neurons" that matched real biological neurons (e.g., specific face detectors) without being explicitly told to.

A process where the network is "inverted" or run backward; the AI is asked to generate/modify an image to maximally excite a specific neuron, creating psychedelic results.

It suggests drugs might over-activate feedback loops, causing the brain to project its internal expectations/priors onto reality, just like Deep Dream forces clouds to look like dogs.

DCNNs are primarily Feedforward (Input -> Output), whereas the human brain has massive Feedback (Top-Down) connections where thoughts influence vision.

Using known brain biology (like maps or layers) to build models provides better explanatory power than just looking for statistical patterns.

Building AI with Recurrent Connections (loops and feedback) to truly mimic the human brain's intelligence.

1. The New Tool (Encoding vs Decoding), 2. Proof of Concept (Vision), 3. Main Discovery (Abstract thoughts like Math have physical maps).