Learning and Perception

Many intelligent systems operate in a complex environment, and the tasks they must resolve require modeling of the environment. In some cases, when the environment is highly dynamic, problem solving can not be planned a priori and it involves collaborations between heterogeneous and autonomous processes (eg human and artificial actors). Another specificity of the considered applications is the lack of annotated corpus. We rely in this case on collaboration models between agents that achieve parts of the solution on the fly, allowing to advance in the resolution or learning on other parts of the data. Another part of our activity on environmental modeling is the application of classification techniques on raw data provided by sensors (typically a 3D point cloud provided by a sensor laser), and the recognition of moving objects with telemetric sensors and cameras, with a distinction between the recognition of generic object categories and recognition of individual object instances.

Finally, we are interested in the visualization of structured data. Ascending classification provides an effective aid in the analysis and modeling of data. However, when working with trees containing more than a few hundred terminal nodes, visual representation and exploration of the classification become difficult. This year we introduced a new visualization model, called “Stacked Trees”, which can simultaneously represent on the screen information coming from tens of thousands of objects.