week 2: structural descriptions

• Marr
• Biederman
• Hummel and Biederman
• Hummel

stages of vision

stage I

slide 3

stage II

slide 4

stage III

slide 5

stage IV

slide 6

Marr: a recap

slide 7

what issue does this approach purport to address?

slide 8

is the result satisfactory?

slide 9

Marr to Biederman

slide 10

geons = "geometrical ions"

slide 11

geons are generalized cylinders

slide 12

putting it together

slide 13

how about a model?

slide 14

JIM (Hummel and Biederman, 1992)

slide 15

slide 16

some psychophysics

slide 17

some psychophysics

slide 18

slide 19

slide 20

slide 21

JIM II

slide 22

JIM III

slide 23

slide 24

grouping by synchrony of firing

local features interact to segment image into geons

slide 25

time course of processing

feature (surface) outputs at time t go to both the propositional and the array representations

slide 26

time course of processing

feature (surface) outputs at time t go to both the propositional and the array representations

slide 27

time course of processing

feature (surface) outputs at time t go to both the propositional and the array representations

slide 28

time course of processing

feature (surface) outputs at time t go to both the propositional and the array representations

slide 29

time course of processing

slide 30

slide 31

two issues

• geons (low-level vision)
• binding

slide 33

issue #1 (geons): assume a spherical chicken

Dickinson, Nevatia, etc.

slide 34

issue #2 (binding): LISA

slide 35

slide 36

slide 37

Holographic Reduced Representations (Tony Plate)

slide 38

slide 39

slide 40

slide 41

slide 42

slide 43

slide 44

slide 45

slide 46

comparing HRRs with the "classical" approach

slide 47

slide 48

slide 49

slide 50

HRRs: implications for the binding problem

"Hummel and Biederman (1992) discussed the binding problem and identified two main problems faced by conjunctive coding approaches such as Tensor Products (Smolensky, 1990). These are exponential growth of the size of the representation with the number of associated objects (or attributes), and insensitivity to attribute structure. HRRs have much in common with conjunctive coding approaches (they can be viewed as a compressed conjunctive code), but do not suffer from these problems. The size of HRRs remains constant with increasing numbers of associated objects, and sensitivity to attribute structure has been demonstrated in this paper."

[from Tony Plate's NIPS'93 paper]