most interesting machine learning methods require varying degrees of sophistication: from linear algebra matrix inversion, as many matrix/tensor decomposition techniques as you can think of are the bread and butter. you will need to consider a huge range of probability distributions and the ability to manipulate/sample form those is crucial. when you have lots of data, you will need to consider threaded computation and other big data methods that run over distributed servers. and very importantly -- though people forget this -- you will also need to visualise everything if you want to make sense of the data and algorithm (machine learning isn't magic after all and poor execution of ML algorithms on simple data will elicit unusable results).
i'm afraid i see very few advantages to low level languages unless you intend to build a bespoke and narrow implementation of one particular model. this is generally only required in deployment contexts, but for design and development... you'll struggle to do much machine learning of worth with low-level languages.
