4 ms·

Most machine learning models are nothing but souped up regression that swaps out standard polynomials for matrix-valued polynomials with some x'es replaced by n

Most machine learning models are nothing but souped up regression that swaps out standard polynomials for matrix-valued polynomials with some x'es replaced by non-linear activation functions like tanh(x) The magic is in the toolkits that allow one to build these complex expressions through net chain composition, and provide symbolic differentiation to easily calculate gradient on hugely complex net chain graphs. The avg person isn't going to benefit from ML unless you also teach them the old-hat statistical fitting methods that worked well enough before ML came on the scene. Things like SVD, MPCA: https://en.wikipedia.org/wiki/Singular_value_decomposition https://en.wikipedia.org/wiki/Singular_value_decomposition https://en.wikipedia.org/wiki/Multilinear_principal_component_analysis https://en.wikipedia.org/wiki/Multilinear_principal_componen... Old stats techniques like: https://en.wikipedia.org/wiki/Elastic_net_regularization https://en.wikipedia.org/wiki/Elastic_net_regularization Its not worth training a bunch of folk who cant go deep into the heavy mathematics, statistics, and epistomology of what ML really is in terms of manifolds. Just show them what ML can and cannot do, if its not possible to give them the comprehensive education. And besides, most people, once they find out ML is a shit ton of math and as profound and based as the theory of functions, shakes their head and trods off. They wanted magic soup and you gave them the biggest questions about existence..about patterns..forms..things..about consciousness, intelligence, and life, about meaning. Don't expect the avg lowbrow person to want to explore ML once they get hit with the reality of its implications, rigor, and depth.

zozbot123 8y ago
> Most machine learning models are nothing but souped up regression that swaps out standard polynomials for matrix-valued polynomials with some x'es replaced by non-linear activation functions like tanh(x) Well, not quite. The non-linear activation feature is found already in "generalized linear models" (as a 'link function'). The typical "deep learning" model is a hierarchical/multi-level version of the same - and just like any other hierarchical/multi-level model, the point of that is to account in a parsimonious way for increasingly-complicated interactions among the regressors/features. But I think your broader point is right; there is a lot of hype in ML that derives purely from a lack of familiarity with basic statistical principles.

zozbot123 8y ago

> Most machine learning models are nothing but souped up regression that swaps out standard polynomials for matrix-valued polynomials with some x'es replaced by non-linear activation functions like tanh(x) Well, not quite. The non-linear activation feature is found already in "generalized linear models" (as a 'link function'). The typical "deep learning" model is a hierarchical/multi-level version of the same - and just like any other hierarchical/multi-level model, the point of that is to account in a parsimonious way for increasingly-complicated interactions among the regressors/features. But I think your broader point is right; there is a lot of hype in ML that derives purely from a lack of familiarity with basic statistical principles.