Color in IT Cotex

Bonnie Cooper


model the hierarchy of color & hue signal propagation from LMS cone photoreceptors in the retina to the hue selective tuning observed in IT cortex

Visual Pathways

Modeling Color in the Cortex

LMS Cones

transform RGC image values to LMS cone weights




LGN Layer

Actual Stimulus Used:

LGN Single-Opponent Mechanism: \[ R_{mLGN} = \phi ( a_L( G(x,y,\sigma _L ) * R_L) + a_M( G(x,y,\sigma _M ) * R_M) \] \[ + a_S( G( x,y,\sigma_S ) * R_S) - a\prime _L( G( x,y,\sigma \prime _L ) * R_L) \] \[ - a\prime _M( G( x,y,\sigma \prime _M ) * R_M) - a\prime _S( G( x,y,\sigma \prime _S ) * R_S)) \]

V1 Layer

V1 Single-Opponent Mechanism: \[ R_{mV1} = \phi ( G( x,y,\sigma _{mV1} ) * R_{mLGN} ) \]

V2 Layer

  1. Single Opponent mechanism:

    \[ R_{singleopp.mV2} = \] \[ \phi ( G( x,y,\sigma _{mV2} ) * R_{mV1} ) \]


  1. Multiplicative Combination:


    \[ R_{multiplicative.mV2} = \] \[ \phi ( G( x,y,\sigma _{mV2} )* R_{mV1[L,M]} \] \[ * G( x,y,\sigma _{mV2} ) * R_{mV1[S]} \]

V4 Layer

Hue Selective Channels:

\[ R_{mV4,i} = \phi \left( \sum_{j = 1}^{14} w_{ij}( G( x,y,\sigma _{mV4}) * R_{mV2,j}) \right) \]






Simulation Time!


  • But the ratios of L,M & S cones vary a lot!
  • Add some gaussian noise to the LMS weights to simulate cone mosaic variety
  • Run the simulation 100 times
  • How do the distributions of peak responses at the earliest layer (LGN) compare to V4?

Simulation Results


LGN peak responses cluster tightly, presumably along approximate DKL color space axes


V4 responses have much higher variability and span hues inbetween LGN representations

Future Directions

Well that's great! a simple model managed to capture some aspects of hue representation. But there is a lot more ….

  • Build more realistic single-opponent channels at mLGN
  • in V2: are multiplicative channels the only way? …what about lateral inhibition?
  • would like to take the training wheels off this model & implement as a quasi-neural net….