LECTURE 4

figure ground problem

challenge when brain has to decide which part of a visual/sensory scene is the main focus & which is the unimportant backdrop

Gestalt grouping principles

• brain naturally organizes individual visual elements into unified patterns, whole objects & meaningful groups → emerge from natural scene statistics (visual system encodes actual statistical regularities of the natural world)

• edge detection isn’t enough, visual system has to group subsets of edges into meaningful chunks (Gestalt)

1. proximity

2. similarity

3. continuity

4. closure

5. connectedness

why are we so tuned to mirror symmetry?

• a visual system that’s good at detecting symmetry is effectively good at detecting living things from asymmetric backgrounds

• strong cue for figure-ground segmentation

• humans are much better at detecting vertical mirror symmetry

why is object recognition so hard?

• a 2D retinal image is a projection of a 3D world

• image variability

image variability

differences & inconsistencies found across multiple images of the same subject, scene, or data type

• identity-preserving transformations

  • viewer variables (position on retina, size)

  • object variables (pose, background)

  • lighting variables (luminance, shading)

• within-class variability

  • new exemplars (instance within a class) of object (drawings, new type)

what did the computer classifier show?

recognition is easy when transformations are removed & only within-class shape variation remains as a clean signal, but becomes hard when real-world transformation variability is present

selectivity-invariance trade off

increasing a system’s tolerance to irrelevant changes reduces its ability to distinguish between fine-grained details

perceptual constancy

recognizing the same object but at a different size/scale

constancy using templates

1. each retina feeds into its own receptive field/template, tuned specifically to detect a spider at that particular size

2. each of these template neurons feeds into its own neuron, which fires if that specific template is matched

3. the outputs of these size-specific neurons are combined with an OR gate into a single downstream neuron

problems w/ constancy using templates

1. combinatorial explosion

2. no graded/parametric generalization

3. doesn’t scale to within-category variation

single parameter neurons

have different types of neurons containing info about different types of features (color cells, contrast cells, etc.)

binding problem

how do you attach (bind) the correct parameters to different objects?

• still debated

• perceptual grouping via attention

• temporal synchrony

grandmother cell

hypothetical single neuron that responds selectively to one specific concept

• one dedicated neuron per person, per object, per concept

• recorded from single-cell recordings in the human brain in epilepsy patients

  • found neurons that responded very selectively to specific individuals

• still inherits the same problems

the coding spectrum

how many neurons respond to a given stimulus & how many stimuli does each neuron respond to

1. local coding/representation: each individual stimulus/concept is represented by a single, dedicated neuron

2. distributed coding/rep: every stimulus is represented by a pattern of activity across the entire population of neurons; each individual neuron participates in representing many different stimuli (orientation tuning)