This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revision Previous revision
active_learning [2018/03/15 11:26]
active_learning [2018/03/15 11:27] (current)
Line 53: Line 53:
 https://​arxiv.org/​abs/​1708.00489v3 Active Learning for Convolutional Neural Networks: A Core-Set Approach https://​arxiv.org/​abs/​1708.00489v3 Active Learning for Convolutional Neural Networks: A Core-Set Approach
-Our empirical ​study suggests ​that many of the active learning heuristics in the literature are not effective when applied to CNNs in batch settingInspired by these limitations,​ we define ​the problem of active learning as core-set selection, ie. choosing ​set of points such that a model learned over the selected subset is competitive ​for the remaining data points. We further ​present a theoretical result characterizing the performance of any selected subset ​using the geometry of the datapoints. As an active learning algorithm, we choose the subset which is expected to yield best result according to our characterizationOur experiments show that the proposed method significantly outperforms existing approaches in image classification experiments ​by a large margin.+Our empirical ​analysis showed ​that classical 
 +uncertainty based methods have limited applicability to the CNNs due to the correlations caused 
 +by batch samplingWe re-formulate ​the active learning ​problem ​as core-set selection ​and study the 
 +core-set problem ​for CNNs. We further ​validated our algorithm ​using an extensive empirical study. 
 +Empirical results on three datasets showed state-of-the-art performance ​by a large margin.