1. An Inferior Temporal
Cortex Model
for Object Recognition
and Classification1
Fernando Jesús García Hípola – Taller de RRNN
2. Introduction
• Aim of the Study: Simulating Temporal
Cortex 3D object classification through a
NN that imitates it.
• Virtual information is processed in two
major parallel pathways:
the Dorsal and the Ventral
• V1, V2,
V4
• Inferior
Temporal
Cortex
(IT)
5. 1
• Inicialización de los pesos wi
2
• Presentación en cada iteración un patrón de entrada x(t)
3
• Determinar similitud entre pesos de cada neurona y la entrada
4
• Determinar Neurona Ganadora (Mayor similitud -> Menor error)
5
• Actualización pesos de la ganadora y sus vecinas (f(x) de vecindad)
6
• Volver al paso 2 si no se ha alcanzado nº máximo de iteraciones, Si no
FIN
6. Kohonen SOM
Los Pesos se comparan con
el vector de entrada: Gana
la Neurona de la 2ª capa
más parecida (min error
entre sus pesos y la
entrada)
Se actualizan los pesos
según una función de
proximidad para asegurar la
vecindad de las clases
(Preservar Topología)
7. mnSOM
• Kohonen SOM only can deal
with vectors imputs.
• Tukunawa & Furukawa ->
mnSOM replace SOM imput
vectors by functional
modules
8. mnSom with RBF
Network Modules
• Radial Basis Function Neural Networks
• Advantages for Image Recognition:
No need of an Additional Additional
Algotithm for avoiding local minima
The network leans to store the objects
representation in the inner center
The use of RBF is more
neurophysiologically plausible
9. The NN for Image Recognition
• 10 objects as image input, in different degree of rotation:
3 objects with 5 straight segments.
3 objects with 4 straight segments.
4 objects with 3 straight segments.
10. 4. Adaptative Process:
All the weights and centers within the modules are updated
following the back-propagation algorithm
3. Cooperative Process
The Weights are calculated by the neighborhood function
2. Competitive Process
The module that minimizes the error is the winner.
1. Calculating Process
Random initialization of weights and calculate all outputs for
all the imputs in a single RBF unit
14. The NN for Image Recognition for 3D
Objects
• The previous hybrid NN works well for simple 2D
images, but... What about 3D images?
• Solution: Adding a pre-processing Module emuleting V1,
V2, V3 and IT biological neural areas, by processing
layers:
A retinal image is divided into small overlapping
patches
The patterns are filtered through a layer (S1) of
Gabor filters.
Position invariant detectors
16. The NN for Image Recognition for 3D
Objects
• 43D-Objects: Spherical objects and Spiky objects
• Objects were presented to the NN in projections at each 10º rotation
• Training: 200 epochs.
• Results follow: similarity between objects and similarity between the
different views of the same object