What the other 85% of V1 is doing Bruno A. Olshausen Helen Wills - - PowerPoint PPT Presentation

▶

Aug 15, 2022 241 likes •649 views

What the other 85% of V1 is doing Bruno A. Olshausen Helen Wills Neuroscience Institute School of Optometry and Redwood Center for Theoretical Neuroscience UC Berkeley The standard model of V1 R e s p ons e P o i n t w i s e R e s p

SLIDE 1

What the other 85%

f V1 is doing

Bruno A. Olshausen Helen Wills Neuroscience Institute School of Optometry and Redwood Center for Theoretical Neuroscience UC Berkeley

SLIDE 2

The “standard model” of V1

Image I(x,y,t) K(x,y,t) Receptive field

linear

response

or /

Response normalization Pointwise non-linearity

neighboring neurons

r(t) Response

SLIDE 3

Lessons from the retina
Lessons from invertebrates
Vast overcompleteness of

V1

Non-linearities of cortical neurons
Difficulty of predicting neural responses to

time-varying natural images

Why I am skeptical of the standard model

SLIDE 4

Lessons from the retina

SLIDE 5

SLIDE 6

On vs. off cone bipolar cells

SLIDE 7

Rod bipolar cell is

f on-type only

Net convergence of rods to bipolar cells

SLIDE 8

AII amacrine cell links rod bipolar cells to ganglion cells

SLIDE 9

Lessons from invertebrates

SLIDE 10

Jumping spiders

SLIDE 11

Jumping spiders

SLIDE 12

Vast overcompleteness of V1

SLIDE 13

1 mm2 of cortex analyzes ca. 14 x 14 array of retinal sample nodes and contains 100,000 neurons

SLIDE 14

SLIDE 15

V1 output is overcomplete by a factor of 50:1

Parvo cell input fibers V1 output fibers (layer 2/3)

70 μ

SLIDE 16

Non-linearities of cortical neurons

SLIDE 17

i j Perisomatic thin branches Thin branch subunits i y1 y2 Distal apical thin branches 2-Layer model 3-Layer model (a) (b) (c)

Current Opinion in Neurobiology

Hausser & Mel (2003)

SLIDE 18

Difficulty of predicting V1 neural responses to time-varying natural images

SLIDE 19

SLIDE 20

5 10 15 20 25 30 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 time (sec) spike count (35 msec bins) a219, group 3, cell 4 ρ=0.362

18 ms 53 ms 88 ms 123 ms 159 ms 194 ms 229 ms 264 ms

Responses of V1 neurons are not well predicted by RF models

receptive field:

SLIDE 21

5 10 15 20 25 30 10 20 30

cell 1

5 10 15 20 25 30 10 20 30

cell 2

5 10 15 20 25 30 10 20 30

spikes/sec time (sec) cell 3

Responses of neighboring cells are heterogeneous

SLIDE 22

Proportion of cells studied Variance explained

1.0 1.0

~85% of V1 function not understood

~0.4 0.3-0.4

What is the other 85% doing?

SLIDE 23

There’s hope.

SLIDE 24

Silicon polytrodes

(Swindale, Blanche, Spacek)

SLIDE 25

SLIDE 26

SLIDE 27

What does a “day in the life of

V1” look like?

Explaining away (sparsification)
Phase
Figure-ground
Synchrony
Laminar distribution of function (microcircuit)

What to look for

SLIDE 28

! ! ! !

Feedforward response ("#)

! ! ! !

Sparsified response ($#) ! !

Explaining away

SLIDE 29

Phase

Phase-Quadrant Demodulation Code [0, 0] [1, 0] [1, 1] [0, 1]

Re Im

Figure 2: The phase demodulation process used to

Iris recognition (Daugman)

SLIDE 30

Time-varying phase encodes information about transformations

time time

coefficient index

amplitude phase

SLIDE 31

Modeling phase dependencies (Charles Cadieu)

sparse

SLIDE 32

Learned D (space domain)

SLIDE 33

Learned D (frequency domain)

SLIDE 34

Learned D

SLIDE 35

Figure-ground

SLIDE 36

V1 simple cells can represent amodal completion Sugita (1999)

SLIDE 37

Synchrony

LGN spikes are phase-locked to ongoing retinal oscillations (Koepsell, Sommer, Hirsch)

SLIDE 38

Distribution of function across laminae

SLIDE 39

The Unknown As we know, There are known knowns. There are things we know we know. We also know There are known unknowns. That is to say We know there are some things We do not know. But there are also unknown unknowns, The ones we don't know We don't know.

Feb. 12, 2002, Department of Defense news briefing

From: The Poetry of Donald Rumsfeld Hart Seeley, Slate Magazine