I have coded a simple supervised ML classification using 10-20K data points for 25 samples. Linear ML models run quickly for example naive Bayes, linear regression and SVM linear on a small multi-core CPU desktop with 8 G RAM via scikit-learn.
However, non-linear models for example lasso regression and non-linear SVM via RBF and polynomial kernels "do not converge" within reasonable uptime, i.e. no more than 4 hours uptime.
Three questions:
- is this known behaviour and is there a pattern within the data set types?
- alternatively are their data sets that simply do not converge for non-linear models?
- would it be resolved via GPU *, such as the NVIDIA machines?
The rationale of the question is being able to justify that a shift onto a cloud platform will deliver the results sought.
* For example, a GPU accelerated "non-linear" ML model via either:
- libSVM SVM RBF kernel or;
- thunderSVM or possibly;
- falcon
