Online Machine Learning. 4.1.1 Online Learning Overview Traditional machine learning algorithms take a two-stage approach: batch training based on an annotated training dataset, followed by batch prediction on new datasets based on the model generated from stage one (see Figure 4.1 (a)). In contrast, online machine learning xxxxxxxxxx [00, 11] take an iterative approach (Figure 4.1 (b)). It learns one data points at a time, and based on external feedback, adapts its prediction model for subsequent data points. Online learning matches the data collection model of heathcare organizations, where new patient data is received on a daily basis. In IDEAL-X, when the predication model achieves a satisfactory accuracy, the system may be switched to run in batch mode. In both batch and online learning, the objective is to learn some function: F: X ->Y where X is the input dataset and Y is the set of predicted outcomes (also called labels). Given a sequence of input data (x1, y1), (x2, y2), ……(xn, yn), an online learning algorithm generate successive approximations to F that best capture the data points that have already been processed. Each approximation is calculated only on previous approximation and current data point. Previous data points need not be stored, which guarantees constant memory usage. The goal of online learning is to minimize the cumulative gap between predicted value and the real answers. This could be represented with a loss function: £= ∑ E(f(xi) - yi ) where f(xi) is the predicted value of iteration i, and yi is the real answer. Function E estimate the different between these two values. Online machine learning not only significantly reduces human’s effort for annotation, since user’s role evolves from annotator to reviewer over time, it also provides the mechanism for collecting feedback from human-machine interaction to continuously improved the system’s model.
