1 of 25

Vision

David Ulloa

2 of 25

Vision

David Ulloa

3 of 25

Vision is complicated

  • Continuous stream of moving, ambiguous input

​

4 of 25

Vision complicated

  • Continuous stream of moving, ambiguous input

​

5 of 25

Computational Neuroscience

  1. Develop models to represent images on the eye and how we process them
    1. Image -> Internal representation

​

​

  • Biologically plausible

6 of 25

Crash course on the eye

7 of 25

Crash course on the eye

8 of 25

Crash course on the eye

Retina

Optic Nerve

Back of the eye

Connectivity between adjacent photoreceptors

9 of 25

Receptive Fields

  • Localized group of neurons that fire in response to a light pattern

10 of 25

Receptive Fields

  • Localized group of neurons that fire in response to a light pattern
    • Center-surround antagonistic

11 of 25

Edge Detection

12 of 25

How do we represent this mathematically?

13 of 25

14 of 25

Base firing rate

Net firing rate

Correlation

15 of 25

In 1-D

Not limited by the center-surround!

16 of 25

Improving the model

  • Most stimuli are not static

​

​

  • Solution: introduce time dependency in our filter (and stimulus)

17 of 25

Spatio-temporal Receptive Field

“Temporal profile”

Convolution

18 of 25

Examples

19 of 25

Examples

20 of 25

Examples

21 of 25

Examples

Motion selectivity

22 of 25

Putting it all together

  1. Receptive fields on the retina

​

  • Correlate image and a filter

​

  • Added time-dependent stimulus and filter

​

  • Convolution to find

23 of 25

References

  • Bear, M. F., Connors, B. W., & Paradiso, M. A. (2001). Neuroscience: Exploring the brain. Lippincott Williams & Wilkins.

​

  • Adapted from 9.40 lecture notes

24 of 25

Discretization

25 of 25

Deep CNNs

Stack multiple layers for further processing

No time dependence!