Simulating Human-Like Attention	
  
Deborah	
  Sandoval	
  
Prof.	
  Michael	
  Graziano	
  
Mo6va6on	
  
My	
  Project	
  
Implemen6ng	
  a	
  2D	
  Grid	
  A	
  Linear	
  Connec6on	
  
Similar	
  Projects	
  
Tes6ng	
  Behavior	
  of	
  Linear	
  Connec6ons	
  
Artificial Attention: Research on the human brain has allowed for
details on the interactions between neurons and neuron networks to be
thoroughly described, though much information on neuronal mapping is
still unknown. Can the behavior of neurons in the visual cortex be
simulated using a strict, biologically realistic approach? Is a bottom-up
approach the best approach?
Goal: Simulate human-like attention using the biological model and
analyze behavior and performance
Many large scale projects have attempted to simulate the entire
human brain and its complex connections using a biologically
realistic model. The most recent and popular projects are:
Blue Brain Project – modeling the neocortical column in the rat
brain; uses BlueGene/L supercomputer
-bluebrain.epfl.ch
Human Brain Project – understanding the human brain and its
complex connections by using a strict bottom-up approach; first
simulating mouse brain
-www.humanbrainproject.eu
These projects utilize a supercomputer to simulate parts of the billions
of neurons in the human brain.
1.) Create program that simulates neuronal competition and
suppression found specifically in the visual cortex and receptive field
from a strict biological model:
•  Implement linear connections and test behavior on various
combinations of input
•  Implement a grid connection and test behavior on various
combinations of input
2.) Analyze the cost and benefits of using a strictly bottom-up
technique
•  To what extent can the neurons and process be biologically
replicated?
•  What are the consequences of implementing more efficient (though
not biologically realistic) algorithms?
	
  
THE NEURON OBJECT
Input Level
Inhibition
Level
Activation
Level
Connections
Neuronal Competition/Suppression Algorithm
Input = Incoming Input
Inhibition = Input * INHIBITION_FACTOR
Activation Level = Input – Σ (Inhibition of Connections)
Major Differences from Biological Model
•  Connections: Details on connections vary by brain and many details on cortical
mappings are unknown. However, linear connections still replicate behavior (see
below).
•  Parallel vs. Sequential Calculations: Calculation of activation occurs in series, rather
than simultaneously; for large numbers of neurons (the visual cortex with near 140
million) this can significantly effect performance.
No Input
1D Input Tests on five linearly connected Neurons:
All Activated
One Activated
One Deactivated
Half Activated
Repeating
Observed Behaviors:
•  Edge Effects
•  Exaggeration of border (heightening of
activation before border, suppression of
activation after border) This effect is the
reasoning behind the optical illusion called
Mach Bands (see left).
Effects are found in Biological Model ->
Linear simulation a success!
Mach Bands: If looking at border between gradient sections, there appears to be a
dark aura on surrounding darker side of border and a white aura surrounding lighter
side. Neuronal competition/suppression facilitates this illusion and is a result of the
brain evolving to find borders in a visual environment important.
All Activated
Add two more connections to Neuron object to create a grid-like connection that
can map input of a 2D space (Note that human brains are able to convert a 2D
image to a 3D image)
Light Input Neuron Grid Network
Tes6ng	
  Behavior	
  of	
  Grid	
  Connec6ons	
  
Observed Behaviors:
•  Contrast of border:
Neurons on inner border
of circle have a higher
activation then the center
of the circle; Neurons on
outer border of circle have
a lower activation relative
to other no-input Neurons
Effects are found in
Biological Model -> Grid
simulation a success!
Analysis	
  and	
  Summary	
  
	
  
Advantages and Disadvantages of Bottom-Up techniques
•  Able to simulate a computational model of the biological attention process in
the human brain using the most basic unit - the neuron. Unfortunately,
mapping of neurons are complicated and are not fully understood to date.
Despite this, a small linear connection can still replicate this basic process of
attention.
•  Though this basic attention process can be simulated linearly, the complicated
connections of the actual human brain might be useful in higher level functions.
This must be considered when simulating the entire brain.
•  The computation power needed to replicate the 140 million neurons just in the
visual cortex are currently not ideal. Parallelism in calculations is definitely
needed to simulate speeds equivalent to the human brain.
•  We can learn how to simulate many processes by studying human brain, but
some techniques are centered greatly on physics and chemistry between and
within neurons. We can simply our computational models by excluding these
details.

Deborah_Sandoval

  • 1.
          SimulatingHuman-Like Attention   Deborah  Sandoval   Prof.  Michael  Graziano   Mo6va6on   My  Project   Implemen6ng  a  2D  Grid  A  Linear  Connec6on   Similar  Projects   Tes6ng  Behavior  of  Linear  Connec6ons   Artificial Attention: Research on the human brain has allowed for details on the interactions between neurons and neuron networks to be thoroughly described, though much information on neuronal mapping is still unknown. Can the behavior of neurons in the visual cortex be simulated using a strict, biologically realistic approach? Is a bottom-up approach the best approach? Goal: Simulate human-like attention using the biological model and analyze behavior and performance Many large scale projects have attempted to simulate the entire human brain and its complex connections using a biologically realistic model. The most recent and popular projects are: Blue Brain Project – modeling the neocortical column in the rat brain; uses BlueGene/L supercomputer -bluebrain.epfl.ch Human Brain Project – understanding the human brain and its complex connections by using a strict bottom-up approach; first simulating mouse brain -www.humanbrainproject.eu These projects utilize a supercomputer to simulate parts of the billions of neurons in the human brain. 1.) Create program that simulates neuronal competition and suppression found specifically in the visual cortex and receptive field from a strict biological model: •  Implement linear connections and test behavior on various combinations of input •  Implement a grid connection and test behavior on various combinations of input 2.) Analyze the cost and benefits of using a strictly bottom-up technique •  To what extent can the neurons and process be biologically replicated? •  What are the consequences of implementing more efficient (though not biologically realistic) algorithms?   THE NEURON OBJECT Input Level Inhibition Level Activation Level Connections Neuronal Competition/Suppression Algorithm Input = Incoming Input Inhibition = Input * INHIBITION_FACTOR Activation Level = Input – Σ (Inhibition of Connections) Major Differences from Biological Model •  Connections: Details on connections vary by brain and many details on cortical mappings are unknown. However, linear connections still replicate behavior (see below). •  Parallel vs. Sequential Calculations: Calculation of activation occurs in series, rather than simultaneously; for large numbers of neurons (the visual cortex with near 140 million) this can significantly effect performance. No Input 1D Input Tests on five linearly connected Neurons: All Activated One Activated One Deactivated Half Activated Repeating Observed Behaviors: •  Edge Effects •  Exaggeration of border (heightening of activation before border, suppression of activation after border) This effect is the reasoning behind the optical illusion called Mach Bands (see left). Effects are found in Biological Model -> Linear simulation a success! Mach Bands: If looking at border between gradient sections, there appears to be a dark aura on surrounding darker side of border and a white aura surrounding lighter side. Neuronal competition/suppression facilitates this illusion and is a result of the brain evolving to find borders in a visual environment important. All Activated Add two more connections to Neuron object to create a grid-like connection that can map input of a 2D space (Note that human brains are able to convert a 2D image to a 3D image) Light Input Neuron Grid Network Tes6ng  Behavior  of  Grid  Connec6ons   Observed Behaviors: •  Contrast of border: Neurons on inner border of circle have a higher activation then the center of the circle; Neurons on outer border of circle have a lower activation relative to other no-input Neurons Effects are found in Biological Model -> Grid simulation a success! Analysis  and  Summary     Advantages and Disadvantages of Bottom-Up techniques •  Able to simulate a computational model of the biological attention process in the human brain using the most basic unit - the neuron. Unfortunately, mapping of neurons are complicated and are not fully understood to date. Despite this, a small linear connection can still replicate this basic process of attention. •  Though this basic attention process can be simulated linearly, the complicated connections of the actual human brain might be useful in higher level functions. This must be considered when simulating the entire brain. •  The computation power needed to replicate the 140 million neurons just in the visual cortex are currently not ideal. Parallelism in calculations is definitely needed to simulate speeds equivalent to the human brain. •  We can learn how to simulate many processes by studying human brain, but some techniques are centered greatly on physics and chemistry between and within neurons. We can simply our computational models by excluding these details.