Abstract |
During the surveillance of an area using ISR assets in coalition operations, the same object may be observed with dif- ferent modalities by different coalition members. In those cases, the different members may reach different conclusions about the object. Understanding such systems is useful to design princi- ples for coalition collaboration. We model these environments as requiring classification problems done on overlapping but different set of features. As a specific case, we consider a binary classification problem with two features and ask the following question: suppose different observers observe different features, can they reduce the error on classifying random observations by collaboratively learning a model or performing inference collaboratively? If so, can we quantify accuracy improvements? We study this problem by considering two classes C1, C2 as bivariate Gaussian distributions with variables x and y; one observer observes variable x for both C1 and C2; the other observes variable y for both classes. We consider four strategies: 1) independent learning, independent inference; 2) collabora- tive learning, independent inference; 3) independent learning, collaborative inference; 4) collaborative learning, collaborative inference. Assuming all models are perfectly learned, we show the relation between these four strategies and the advantage of collaboration. We also analyze the problem assuming the amount of training data affects model accuracy. Although we formulate the problem using simple bivariate Gaussian distribution models, the ideas can lead to deeper insights into the advantage gained by collaborations in machine learning. |