Challenges / Olfactory Cocktail Party

Status New
Type Industrial
Start 2015-02-25 02:00:00 CET
End 2015-03-31 02:00:00 CET

Registration is required.


Our dataset represents object recognition in the olfactory domain. The testing data comprises the activity of a simulated gas sensor "traversing" an environment containing multiple odorant sources, as well as a persistent background odorant. In the training data the sensor is traversing the environment with just one odorant source present in isolation at a time. The model should learn the signature of that object during training (which is learned in mixture with a background odorant that is different from the background that will be present in testing). This task is analogous to the capability of biological systems to learn the smell of an object in one environment, and recognize it in multiple different contexts. Video demonstrations of the simulated virtual world environment are linked, so the ML community can get some visual context for the environment. The CFD simulations are driven by computationally light but rather realistic computational fluid dynamic simulation in which the movement of the agent in the environment perturbs the convection field. An interesting 'next step' in this research will be the application of sensory networks trained by this data to enable neural sensorimotor control of the agent in the virtual worlds. All datasets are comprised of 32-bit floating point values representing sensor activation (input features) and odorant concentration (labels) at each point in time, with 4000 training and 1000 testing timepoints. This problem is analogous to invariant object recognition in clutter and background. However, in our setting the object is a chemical signature detected by the activation of an array of olfactory sensors across which each odorant produces a distinct pattern.


Copyright © 2008-2013 by TunedIT
Design by luksite