Machine learning (ML) has recently achieved a lot in areas where the standard assumptions about the data hold and the amount of training data available is large. However, it still faces many challenges in areas where we would need it the most. The commonly made independently and identically distributed (IID) assumption about data is rarely holds in practice, violating the guarantees about ML methods relying on the assumption. In this project, novel tensor-based ML methods turn the IID violations into our advantage so that both sample complexity and computational complexity are decreased, and the reliability of the prediction performance estimates are improved. Immediate applications include zero-shot learning in general, ML in domains with highly-structured data in particular. The track record of the team leaders, Pahikkala and Roos, in applied research already covers a large array of examples in which the preliminary steps of this research direction has been shown to be highly successful.
|Effective start/end date||01/09/2017 → 31/12/2021|
- 113 Computer and information sciences
- 317 Pharmacy
- 318 Medical biotechnology