Efficient Machine Learning for Time-Varying Data
Research makes AI more reliable with changing data
Machine learning can handle changing and incomplete data much better than previously assumed. This is evident from research by data scientist Luis Silvestrin, who developed new methods to analyze time series of sensor data - such as measurements from machines or medical equipment - more reliably.
In practice, this type of data changes constantly. Sensors are adjusted, conditions fluctuate, and important signals, such as malfunctions or medical complications, often occur only rarely. According to Silvestrin, standard machine learning methods frequently fall short as a result. They implicitly assume that data remains stable, whereas in reality, this is rarely the case.
The research shows that usable predictions do remain possible, provided that algorithms explicitly take this variability into account. Silvestrin developed techniques that handle limited and evolving datasets better. These methods proved effective in industrial applications and healthcare, among others.
Fewer malfunctions and better care decisions
The societal impact of these findings could be significant. In industry, companies can use the new approach to detect anomalies in machines earlier, even before malfunctions occur. This saves costs and prevents downtime. A concrete example is a warning system for overheating motors in conveyor belts. The new methods help ensure that rare problems are not missed, while simultaneously limiting the number of false alarms.
This also offers opportunities in hospitals. Doctors often have to make decisions based on limited and constantly changing patient data. The new techniques can, for example, assist in determining the right moment to remove a breathing tube in the intensive care unit, even when little new data is available.
AI that moves with reality
The core of the research is that artificial intelligence works better when systems adapt to change, rather than assuming a stable world. This is relevant in a time when more and more decisions depend on data that is incomplete, noise-sensitive, and dynamic.
Some applications of the new methods are already immediately deployable because they have been tested in realistic environments. However, further validation is required for broader application, depending on the specific sector. The research thus aligns with larger societal developments, such as the rise of smart healthcare systems, a more reliable industry, and the growing role of AI in complex, realistic situations.
More information on the thesis
