Artificial vs Biological Neural Networks: models and debates

1
Artificial vs Biological Neural Networks models
and debates

• A presentation based on
• Lehky Sejnowskis network model of
  shape-from-shading

Presented by Clara Boyd and Angelos Stavrou
2
A Brief Overview of Artificial Neural Networks

• Different Types
• (if the neurons of one of the net's layers may be
  connected among each other)
• Feed Forward
• Feed Back

• Different Learning Algorithm
• (A mathematical algorithm that a neural net uses
  to learn specific problems)
• Backpropagation
• Delta Learning Rule
• Forward Propagation
• Hebb Learning Rule
• Simulated Annealing

3
A Brief Overview of Artificial Neural Networks
Perceptron The Perceptron was first introduced by
F. Rosenblatt in 1958
Type Feedforward Neuron layers 1 input
layer 1 output layer Input value types
Binary Learning Method Supervised
4
A Brief Overview of Artificial Neural Networks
Multi-Layer-Perceptron The Multi-Layer-Perceptron
was first introduced by M. Minsky and S. Papert
in 1969
Type Feedforward Neuron layers 1
input layer 1 or more hidden layers 1
output layer Input value types
Binary Learning Method Supervised
5
A Brief Overview of Artificial Neural Networks
Backpropagation Network The Backpropagation Net
was first introduced by G.E. Hinton, E. Rumelhart
and R.J. Williams in 1986
Type Feedforward Neuron layers 1
input layer 1 or more hidden layers 1
output layer Input value types
Binary Learning Method BackPropagation
6
Hubel Wiesel

• Area V1 in the Monkey
• Receptive Fields (orientation selectivity to bar
  of light)
• Vision based on a set of EMERGENT properties
• Each cortical cell extracts a different feature
  of the visual image

Simple Cell
Complex Cell
7
Macrocircuitry Between Visual Areas

MT
1. Redundancy of Connections
PO
V3
VP
PIP
V2
2. Bidirectional Transport
V1
3. Hierarchical Organization
4. Parallel Pathways
8
Hierarchical Arrangement Of Visual Processing Stag
es
9
The Visual Pathway
Decisions Actions ( Conscious Awareness?)
Prefrontal Areas Premotor Areas
Higher Visual Areas (V2, V3, V4, Medial
Temporal)
Striate Cortex (V1/area 17)
Lateral Geniculate Nucleus
Retina
10
Microcircuitry V1 Organization
Layer Specific 1. Main Input
from different parts (I,P,M) of LGN terminate in
different
lamina (mostly lamina 4) 2. Other Inputs
(V2,V3,etc) avoid lamina 4 3. Resident Cells
characteristic for a given layer a) lamina to
lamina recurrent/colateral branches form
circuit b) projection axons exhibit lamina
specificity
Highly Localized Processing - most V1
projections dont go very far - more
vertical than horizontal
Many Synapses - convergence and divergence
- stellate cells/local interneurons pyramidal
neurons
11
(No Transcript)
12
(No Transcript)
13
(No Transcript)
14
(No Transcript)
15
Discussion and Open Questions

• Equivalency of models of Artificial neural
  networks to Biological systems (Strong / Weak)

• Learning using a Back propagation technique vs
  pure Feed Forward models of Hubel Wiesel

• How extensive is the inherited genetic knowledge?

16
Discussion and Open Questions

• Although 80 of the artificial neural networks
  work using Back propagation there is no strong
  biological support this rule.

• But is our knowledge of learning adequate?

• How the Feed-Forward network is created?

• Different modes of learning Feed-Forward vs
  Back-Propagation but same result?

• Intrinsic properties are necessary in any case
  of a biological network, evidence of prenatal
  neural networks

17
Discussion and Open Questions

• But is Back-propagation learning achieved by an
  outer and bigger environment/network?

• Master / Slave approach and Rule based Learning.

• Maybe the truth is a hybrid of genetically
  inherited knowledge and learning rules on
  hierarchical unstructured neural networks.