A unified theory for the origin of grid cells through the lens of - - PowerPoint PPT Presentation

a unified theory for the origin of grid cells through the
SMART_READER_LITE
LIVE PREVIEW

A unified theory for the origin of grid cells through the lens of - - PowerPoint PPT Presentation

A unified theory for the origin of grid cells through the lens of pattern formation Ben Sorscher*, Gabriel C. Mel*, Surya Ganguli, Sam Ocko Grid cells Krupic et al. (2012), Derdikman (2014) Trained neural networks learn grid patterns Banino et


slide-1
SLIDE 1

A unified theory for the origin of grid cells through the lens of pattern formation

Ben Sorscher*, Gabriel C. Mel*, Surya Ganguli, Sam Ocko

slide-2
SLIDE 2

Grid cells

Krupic et al. (2012), Derdikman (2014)

slide-3
SLIDE 3

unconstrained nonnegative Dordek et al. (2016) Stachenfeld et al. (2014) Banino et al. (2018) Cueva & Wei (2018)

Trained neural networks learn grid patterns

slide-4
SLIDE 4

Trained neural networks learn grid patterns

slide-5
SLIDE 5
  • 1. Why are the optimal maps grids?
  • 2. What determines the optimal grid type - square,

amorphous, or hexagonal?

slide-6
SLIDE 6

Gradient descent as a pattern forming dynamics

slide-7
SLIDE 7

Gradient descent as a pattern forming dynamics

slide-8
SLIDE 8

Gradient descent as a pattern forming dynamics

1. Why are the optimal maps grids?

slide-9
SLIDE 9

Gradient descent as a pattern forming dynamics

1. Why are the optimal maps grids?

Translation invariance => Fourier modes

slide-10
SLIDE 10
  • 1. Why are the optimal maps grids?
  • 2. What determines the optimal grid type - square,

amorphous, or hexagonal?

Translation invariance => Fourier modes

slide-11
SLIDE 11
  • 1. Why are the optimal maps grids?
  • 2. What determines the optimal grid type - square,

amorphous, or hexagonal?

Translation invariance => Fourier modes

slide-12
SLIDE 12

Nonnegativity yields hexagonal grids

slide-13
SLIDE 13

Nonnegativity yields hexagonal grids

slide-14
SLIDE 14

Taylor expand constraint

Nonnegativity yields hexagonal grids

slide-15
SLIDE 15

Taylor expand constraint

Nonnegativity yields hexagonal grids

“Three-body interaction” between stripes 60o apart

slide-16
SLIDE 16

Taylor expand constraint

Nonnegativity yields hexagonal grids

“Three-body interaction” between stripes 60o apart

slide-17
SLIDE 17

Unifying mechanistic and normative models

Normative encoding models

Banino et al. (2018) Cueva & Wei (2018) Dordek et al. (2016) Stachenfeld et al. (2014)

Grid cell model RNNs

Skaggs et al. (1995) Zhang (1996) Fuhs and Touretzky (2006) Burak and Fiete (2009)

slide-18
SLIDE 18

Unifying mechanistic and normative models

Normative encoding models

Activity dynamics Gradient descent Pattern forming dynamics

Banino et al. (2018) Cueva & Wei (2018) Dordek et al. (2016) Stachenfeld et al. (2014)

Grid cell model RNNs

Skaggs et al. (1995) Zhang (1996) Fuhs and Touretzky (2006) Burak and Fiete (2009)

slide-19
SLIDE 19

Unifying mechanistic and normative models

Normative encoding models

Activity dynamics Gradient descent

Banino et al. (2018) Cueva & Wei (2018) Dordek et al. (2016) Stachenfeld et al. (2014)

Grid cell model RNNs =

Skaggs et al. (1995) Zhang (1996) Fuhs and Touretzky (2006) Burak and Fiete (2009)

Pattern forming dynamics