Case ID: M11-109P

Published: 2011-07-18 14:53:41

Last Updated: 1661889643


Inventor(s)

Shayok Chakraborty
Vineeth Nallure Balasubramanian
Sethuraman Panchanathan

Technology categories

Computing & Information TechnologyImagingMedical DevicesPhysical Science

Licensing Contacts

Shen Yan
Director of Intellectual Property - PS
[email protected]

Adaptive Batch Mode Active Learning for Evolving a Classifier

Due to the tremendous increase in the amount of digital
data, effective large scale data classification is playing an increasingly
important role. In active learning algorithms a ?classifier? is used to classify
the unlabeled data. To ensure reliable performance of the classifier, it must be
trained using a number of labeled examples, or the classifier?s ?training set.?
Such systems often rely on humans to manually label the training set. It is
impractical for human beings to hand-label large datasets, so in order to
optimize the labeling effort associated with training data classifiers, active
learning algorithms have been implemented which select only the promising and
exemplar instances for manual labeling. Current methods utilize the pool-based
strategy which only labels a single datum at a time after which the classifier
is retrained, but this is time consuming and inefficient.

Researchers at Arizona State University have developed a new
technology that incorporates batch mode active learning systems. This method
selects a batch of unlabeled data points simultaneously from a given body of
unlabeled data as opposed to the pool based method which selects only one at a
time. The classifier is retrained once after every batch of data points is
selected and labeled. The selection of multiple instances facilitates parallel
labeling increasing efficiency and productivity. The proposed technology
improves on current models by simultaneously solving for both the batch size as
well as the specific data batch to be classified. The batch size and data are
determined based on projected improvements in the classifier?s efficiency in
classifying unlabeled data.

Potential Applications


  • Video analytics
  • Face recognition software
  • Video processing
  • Medical imaging
  • Video surveillance and security
  • Gaming establishments

Benefits and Advantages


  • Utilizes batch mode active learning
  • Dynamically adapts to the complexity of data stream
  • Intelligently chooses the optimal batch size
  • Specific batch chosen to increase classifier efficiency