perceptual decision making is supported by a hierarchical...

18
Laura Gwilliams | NYU Perceptual decision making is supported by a hierarchical cascade in both biological and artificial neural networks Laura Gwilliams 11th July 2018

Upload: others

Post on 23-May-2020

8 views

Category:

Documents


0 download

TRANSCRIPT

Laura Gwilliams | NYU

Perceptual decision making is supported by a hierarchical cascade

in both biological and artificial neural networks

Laura Gwilliams 11th July 2018

Laura Gwilliams | NYU

The world is an uncertain place❖ Ambiguity❖ Noise

Laura Gwilliams | NYU

AI can categorise, too❖ Artificial intelligence has sought to solve a similar

problem in visual processing❖ Deep neural networks (DNNs) can label images

very accurately

Laura Gwilliams | NYU

AI and neural convergence❖ Correspondence has been found in terms of the representations employed by brains and DNNs

Yamins et al., 2014

Laura Gwilliams | NYU

AI and neural convergence

❖ Not so surprising, given that aspects of DNNs are modelled on vision neuroscience

❖ There is more to characterising a system than simply knowing the representations it uses:

❖ Architecture

❖ Computation

Laura Gwilliams | NYU

Research Question

What is the computational architecture of perceptual

decision making?

Laura Gwilliams | NYU

❖ 17 healthy adults❖ 306 channel MEG

❖ VGG19❖ 19-layer CNN❖ Image Classification

Time / Layer

Parallel Analysis

Laura Gwilliams | NYU

Question 1

How are processing stages organised — what is the

architecture?

Laura Gwilliams | NYU

MEGDecoding Scores

Laura Gwilliams | NYU

MEGDecoding Scores

Laura Gwilliams | NYU

MEGDecoding Scores

Laura Gwilliams | NYU

MEGDecoding Scores

DNNDecoding Scores

Laura Gwilliams | NYU

Question 2

What are the underlying computations?

Laura Gwilliams | NYU

What are the underlying computations?

P ( l

ette

r )

0.5

0.45

0.55

P ( l

ette

r )

Linear Evidence

P ( l

ette

r )

Categorical Percept

Laura Gwilliams | NYU

P ( l

ette

r )

0.5

0.45

0.55

LinearCategorical

P ( l

ette

r )

0.5

0.45

0.55

P ( l

ette

r )

0.5

0.45

0.55

******h4 h h h h h h h4 h h h h h h

What are the underlying computations?

Laura Gwilliams | NYU

What are the underlying computations?

Time (s)

MEG

H

4

VGG19

4

H

Task-specific trained CNN

4

H

linear categorical

Time (s)

MEG

H

4

VGG19

4

H

Task-specific trained CNN

4

H

h

4

Time (s)

MEG

H

4

VGG19

Task-specific trained CNN

linear

Time (s)

MEG

H

4

VGG19

Task-specific trained CNN

h

4

Laura Gwilliams | NYU

Conclusions

❖ We replicate the finding that hierarchical representations converge between biological and artificial NNs

❖ Performance-optimised models only partially predict neural responses during perceptual decision making

Laura Gwilliams | NYU

With big thanks to:

Funding: G1001 Abu Dhabi Institute

• My supervisors, Alec Marantz and David Poeppel, and everyone in the Neuroscience of Language Lab and Poeppel Lab!

• Collaborator Jean-Rémi King

@GwilliamsL