ETIS / ANR Saturn - Self-organizing maps     

Login
Forgot your Password?


Self-organizing maps

The concept of self-organization is intrinsically bound to the definition of the Kohonen’s maps [2]. This is particularly interesting for an artificial vision application.

The self-organizing map (SOM) is able to order it’s elements according to the entry set. This behavior is used to allocate special areas of a computing substrata to the most suitable tasks. As a result of this behavior, we get a data driven allocation process that can be applied to any application fitting this model.

Intrinsically, it is the case of this robotic application where the vision processing is one of the main cognitive action of the robot. For instance, the artificial vision system of our robot stands on a spatiotemporal visual saliency model [3]. In this model, the information contained in the input frames are divided in two types: static and dynamic. The first step of the vision process consists in the extraction of the quantity of each of this type of information in order to balance the allocation through the SOM. This can be done thanks to a retina-like neural network where a first ON-OFF layer computes the magnitudes of the spatial gradient and a second layer computes the temporal gradient.

The robot is also aware of its actuators states. It takes this information into account through another neural network which feeds the third entry of the system.

As the developed SOM converges to an organization that fits the needs of the application, three main areas (in this case study) emerge from the SOM. Each of them are specialized in relation to the different inputs of the system. For instance, the first and the second areas are allocated to compute the static and the dynamic saliency map. The result of the vision maps is to extract the interest points in the image in relation with the shapes and motions in the scene. Finally, a third area is used as a sensorimotor map. This area will merge the informations from the first two areas and from the actuators of the robot in order to learn and to adapt it’s behavior relatively to its actions and its environment.

Binding from the entries of the neural network to the self-organizing map

The learning process of our SOM can be computed as a cost function minimization problem of the average similarity distance between the input vector and the weight vector.

The Manhattan distance in the network topology space between the elected neuron and the learning neuron is used as a criteria of weightiness. A classical “Mexican hat” distribution is then used as a degree of lateral modulation to activate the neurons in a fixed neighborhood and inhibit farther neurons.

This choices improve the reactivity of the system and ensure the topological coherency of the network during the adaptation step as illustrated in the following simulations.

 We developed a simulator of the architecture presented previously. This simulator can work either with real-time sensors informations from the robot or with recorded data. We presented to the controller architecture the data corresponding to a 3-stages robot indoor mission: A) stationary initialization, B) moving for room exploration and C) stationary landscape learning.

As we can see in the figure, the maps compete for the ressources (cells) available into the controller according to their inputs.

Binding from the entries of the neural network to the self-organizing map

The three stages of the robot mission are clearly visible. From the first to the 200th frame (stage A), the robot remains static and a few objects are moving in front of it. From frames 200 to 490, the robot is moving (stage B). The motion estimation map takes space to the static map. This ensure enough computing power to the motion estimation tasks. Finally (stage C) the robot remains stationary moreover there is no motion in front of it. Therefore the stationary map is gaining space from the dynamic one.


[2] T. Kohonen, Self-Organization and Associative Memory. Springer-Verlag, 1989.

[3] F. Verdier, B. Miramond, M. Maillard, E. Huck, and T. Lefebvre, “Using High-Level RTOS Models for HW/SW Embedded Architecture Exploration : Case Study on Mobile Robotic Vision,” Eurasip Journal on Embedded Systems, vol. 2008, p. 349465, 2008. [Online]. Available: http://hal.archives-ouvertes.fr/hal-00524580/en/