← All Days
Day 2 — Sun, Mar 22
What mechanisms power each visual stream?
- Book to read: "Principles of Neural Science" — covers intermediate & high-level visual processing in detail
Dorsal Stream — "Where / How"
Retina
V1
V2 / V3
MT / V5motion energy
MSToptical flow
PPCpredictive coding
Motor cortex Frontal eye field efference copy
Ventral Stream — "What"
Retina
LGN
V1fine detail, colour, high freq
V2contours, figure-ground, illusionary edges
V4shapes, colour constancy, curves
IT (posterior)complex shapes
Anterior ITrecognition
Amygdala Hippocampus emotional relevance & contextual memory
Dorsal Stream — "Where / How"
Works with M-cells (fast, transient, change-detecting). Operates in real-time with no memory.
- Path: Retina → V1 → V2/V3 → MT/V5 → MST → PPC → Motor cortex & Frontal eye field
- Mechanism 1 — Motion Energy MT / V5
MT neurons don't see objects — they only perceive motion vectors, a bank of directional filters tracking movement - Mechanism 2 — Optical Flow MST
Computes self-motion — how fast is the visual field changing around you? - Mechanism 3 — Predictive Coding PPC
Doesn't track the object itself, only the prediction error — where the object should be vs where it actually is - Mechanism 4 — Efference Copy Motor / FEF
The cancellation trick — brain subtracts self-generated motion so the world stays stable despite 3 eye movements per second
Ventral Stream — "What"
Works with P-cells (slower but persistent, sustained high-quality signals unlike M-cells which only respond to changes).
- Path: Retina → LGN → V1 → V2 → V4 → IT (posterior) → Anterior IT → Amygdala & Hippocampus
- V2 isn't passively observing reality — it's actively constructing it. "Perception is always a controlled hallucination"
- V4 computes colour relative to surroundings, not absolute — this is colour constancy
- After Anterior IT, signal splits: Amygdala asks "does what I'm seeing matter to me?" (emotional relevance), Hippocampus stores the context — when, where, with whom
- Interesting: only the ventral stream actually "sees" the object
- Mechanism 1 — Hierarchical Assembly V1 → IT
Like a CNN — V1 detects edges, V2 contours, V4 curves, IT full objects. Each level combines neurons from the previous one - Mechanism 2 — Invariance V4 / IT
Objects remain recognised despite changes in size, position, rotation (translational & rotational invariance) - Mechanism 3 — Population Coding IT
Similar to probabilistic neural networks — recognition outputs a probability distribution, not a binary answer - Mechanism 4 — Top-down Prediction IT → V4
More fibers run downward than upward — like backpropagation. IT tells V4 "I think this is a face, expect contours here." Makes recognition much faster via probabilistic priors