VL13 Deep neural networks

• explicit instantiation of theories

• often explains aspects of the theory instead of the whole theory

• represents data not phenomena

• direct observation of phenomena

• specific to context it is observed in

• predicted by phenomena

• robust, recurring features of the world

• stable generalization of empirical data

• e.g. panic disorder

set of linked propositions that explain the phenomenon

• abduction from phenomena

• there is a phenomenon that is to be understood

• is arises from a target system

1. select target system

2. causal diagram

3. formal models need to reliably

   1. explain

   2. predict

   3. control

• verbal

• mathematical —> formulas

• diagrams

• mechanical models —> DNA

• computational models

• no ambiguity —>computationally explicit hypotheses!

• falsifiable -> can be used to make predictions

• often critiqued to be too complex and black box but

  • complex systems require complex hypotheses

1. computational theory

   1. what is goal of computation? why appropriate? what is the logic of the strategy=

2. representation & algorithm

   1. how can theory be implemented

   2. representation for input + output

3. hardware implementation

   1. how can representation & algorithm be realized physically

Goal: model that represents all three levels

• input layer

• multiple hidden layers

  • with neurons/units

  • weights connect to next layer

• output layer

• input image of digit

• output which digit

—> neurons activated in different degrees

—> activation of next neuron depends on incoming weights

• after calculation of model: feedback what the correct answer was

• thats how it learns

• sliding window

• moves across image and sums up the results

• edge detector: high values mean light and dark are next to each other

• CNN filters: random in beginning and learning through feedbacks

• just like gradient in complexity of features across the ventral stream

• coarse to fine

• deeper levels have later peaks than first levels —> hierarchical just like spatio-temporal cortical dynamics of human visual object recognition

• show that recurrence is required —> feedback from deeper layers to primary levels

• DNN capture stages of processing in time and space

• adversarial examples

• no causation

BUT THEY ARE FALSIFIABLE

• when inputs are intentionally perturbed in a subtle way to make network make a mistake

  • e.g. adding noise to pictures —> ostrich

  • but: maybe ANNs are just more sensitive to adversarial attacks and need another mechanism that makes it more robust

  
  

• ANNs as computational language for expressing falsifiable theories about brain computation

• Reductionist POV:

theory -> hypothesis -> prediction -> test on data -> theory falisified -> theory rejected

but in reality there is often a set of hypotheses and not the whole theory needs to be rejected!

BELT is tested

• uses biologically inspired ANNs to model human behavioral and neural data

  
  

1. core

2. progressive degenerative:

3. empirical data

4. Belt:

1. core: fixed background assumptions that are not challenged from within the programme and not substantially altered without abandoning the programme

2. 
   

1. progressive degenerative:

   1. progressive programmes: generate successful predictions

   2. degenerative: do not generate novel predictions and fail to account for empirical data

2. 
   

1. empirical data: used to test belt hypotheses

2. 
   

1. Belt: auxiliary hypotheses that are experimentally tested and subject to change when new evidence comes up

1. behavioral agreement

2. agreement with neural data -> representational similarity analysis

3. in silico electrophysiology —> lesion studies

4. developmental agreement —> learning vs. training development