We address the problem of efficient acoustic-model refinement (continuous retraining) using semi-supervised and active learning for a low resource Indian language, wherein the low resource constraints are having i) a small labeled corpus from which to train a baseline 'seed' acoustic model, and ii) a large training corpus without orthographic labeling or from which to perform a data selection for manual labeling at low costs. The proposed semi-supervised learning decodes the unlabeled large training corpus using the seed model and through various protocols, selects the decoded utterances with high reliability using confidence levels (that correlate to the WER of the decoded utterances) and iterative bootstrapping. The proposed active learning protocol uses confidence level based metric to select the decoded utterances from the large unlabeled corpus for further labeling. The semi-supervised learning protocols can offer a WER reduction, from a poorly trained seed model, by as much as 50% of the best WER-reduction realizable from the seed model's WER, if the large corpus were labeled and used for acoustic-model training. The active learning protocols allow that only 60% of the entire training corpus be manually labeled, to reach the same performance as the entire data.

No abstract

Electronic Health Records (EHRs) give a lot of information regarding a patient's progress in health, who is admitted to an Intensive Care Unit (ICU). Sepsis is a critical condition suffered by a patient who, if not treated in a timely manner can cause casualties. Machine learning algorithms have evolved to utilize EHRs to help doctors detect the onset of sepsis. In this work, we present a random forest-based ensemble machine learning technique to work on patient data, also called vital sign input, from ICU. The data we used is published as a part of the Physionet Challenge 2019 [11]. The proposed technique performs well on data that contain a major chunk as missing values due to the sparsity of measurement taken in an ICU. We used a combined classifier and an early predictor approach to accomplish the task. The classifier does the job of classification when the early prediction is not possible due to a lack of data. While early predictor predicts the onset of sepsis based on the patient's information it received from previous recordings of vital sign inputs. A utility metric score is used to evaluate the early predictor. The score increases with early predictions and decreases with late predictions as well as false alarms. Our team named 'Tricog' finished 58th in the challenge with a utility score of 0.149 in the official phase on the full test set data.