Answer:
Best in what way? Best for the computer to learn? Well, that is determined by two things...the algorithm that the computer processor(s) run through to achieve the "machine learning", and the speed/efficiency of that algorithm - since it could be an algorithm that will work but perhaps will take so much time to arrive at a good result that it is ineffective in a practical sense.
Now, if you want to ask instead what is the best computer language in which to program algorithms...it is a controversial subject, but overall I would say programming it in a language that is translated well into the machine language which will be running the algorithm, since it must be the best that it can be -- for speed and efficiency. Then, the computer processor(s) speed is/are not hampered by the inefficiency of the choice of code generator or compiler/intrepreter (i.e., in general, the programming language).
However, I also suspect that you are asking from a personal point of view, namely, which language should you learn if you wanted to get into this machine learning field...I think majority of people working in the field like Prolog, Lisp, C++, and Java. Also, there are some people programming in Natural Language Programming (NLP) parsing and lexical analysis that prefer computer languages devoted specifically to NLP.
NLP and Machine Learning algorithms are often working hand in hand to sift through human language texts of all kinds to glean and summarize information and find relevant correlations that exist between all of the ideas and stories in those texts - some factual info, some opinions. Businesses like this stuff for data mining and finding patterns amongst people's likes and dislikes so that they can market to those preferences.
Also, many people like to discover correlations, for instance, that say when certain words are used in a financial article, the chances are that these certain types of stocks will rise in price. Thus data mining is a large reason for the popularity of machine learning.
NLP has been around for many years and is the effort behind what many people have sought -- to build a computer helper that they can talk with in a natural way - so much easier than all that typing and thinking of how to pose the question just so.
It is the next step in the evolution of an internet search engine -- one that learns the best answers and filters out all the extra stuff and does not present the "irrelevant".
My guess is that the programming language choice for machine learning algorithms will become less important than the design of the machine learning algorithms themselves. As time goes forward, the "programming" is being done with visual tools and automatic code generators, and implementing a design of a new machine learning learning algorithm is done with those high level tools where the designers don't necessarily know how it looks in the programming language or in the machine language.
An argument can be made that the "programmers" should know the efficiency of their designs of the machine learning algorithms on that hardware - but that is usually left to the people who are specifically designing the code generators - which is something far removed from the fun of designing the machine learning algorithms. Specialization is the nature of these fields. So, some designers will arrive on the scene that don't know anything about programming languages, they just know the algorithms of the specific domain and are constantly improving them using the Algorithm Studio software.