Improving efficiency: downsampling or low-rank approximation methods Improving efficiency: downsampling or low-rank approximation methods

Kernel methods such as unsupervised MKL, that work batch-based, do not scale well with large numbers of instances, thus its training presents a high computational cost. To ease the burden of computation, we will look at techniques that focus on finding representative instances [1] within the dataset so as to avoid including the whole number of data entries. This will ideally result in very similar embeddings but with a much lower cost.

Preferred skills: comfortable with Matlab; notions of machine learning.

References:

[1] E. Elhamifar, G. Sapiro and S. S. Sastry, Dissimilarity-Based Sparse Subset Selection, in IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 38, no. 11, pp. 2182-2197, 1 Nov. 2016, doi: 10.1109/TPAMI.2015.2511748.