Discovery of Hybrid Ensemble Models Resilient to Input Resolution Deterioration

Recent advances in deep learning (DL) and ensemble learning frameworks significantly improved performance of machine learning (ML) models in different scopes. However, severe data limitations and absence of relevant problems for transfer and meta learning can drastically reduce capabilities of these techniques. Previously we have shown that hybrid approaches leveraging domain-expert knowledge, boosting and DL could discover robust solutions in cases with very limited data availability. Similar challenges are encountered in applications with reduced data sampling rate and resolution due to technological and operational restrictions such as longterm data collection by wearable biomedical devices and fully autonomous robotic systems. Multi-scale features discovered by DL or other empirical ML techniques from high-resolution data could become much less effective or unstable when used with lower resolution inputs. Direct training on low-resolution data may result in the selection of oversimplified and biased model with suboptimal performance. Here we argue that models resilient to input resolution deterioration can be discovered using our previously proposed hybrid framework where boosting-like algorithms are used for effective utilization and enhancement of existing domain-expert models with further non-linear combination of boosted ensemble components via DL or other ML algorithms. Our approach is illustrated in the biomedical context using cardio data from www.physionet.org.