I'm not averse to C++11 - I personally think it's a great language for certain tasks, and have used it extensively for performance-critical ML - see [1] for some public work.
The value proposition of Julia (and Python with Cython) is that you don't have to choose between the productivity you have from a dynamic language (IPython/IJulia [2] are incredibly productive environments for ML/data science IMO), and you have the ability to easily optimize hotspots in your code as you come across them - whether that be adding type annotations/disabling bounds checking in Julia, dropping into Cython, etc.
I think the biggest thing is that you would lose the REPL if you went entirely C++, which can be a big deal when you want to try things incrementally.
On the other hand, this code is written in such a low-level style (nested loops mutating arrays, all functions are top level with no env-dependence) that it's hard to see any other advantage in code like this over just using C++. In fact C++ can leverage closures now with little or no performance loss, unlike either of these languages, so hof-style programming becomes feasible even if high performance is a goal. Also I think there's some benefit to static typing for catching errors early (especially as the program gets larger), and documenting return types which I always appreciate as a code reader (!).
If it's sufficiently easy to link to C++ (or Rust), you could still implement parts in C++ (or Rust) and use the REPL for playing around, that might make a good combination...
Well, why not? Julia is very different from Python, if I am willing to use a completely different tool then the field is wide open and includes anything that has better performance on average
Julia's mental programming model for many things is not much different from Python: dynamically typed, pass by sharing, the implicit scoping rules are even quite similar. Sure, Julia has a few extra bells and whistles, like type annotations and dynamic multiple dispatch (which you can ignore and you have something very much like Python but faster), but still, Fortran and C++ are much more different from Python than Julia is.
Because the goal should be to use productive safer languages, while increasing the code quality of their toolchains.
Many languages that have native code compilers (AOT or JIT) can be used. It is just many seem to think only C and C++ are the only game in town.
Personally I look forward to improved code generation in alternative language compilers, so that we go back to the time where C and C++ were just two among many languages with available compilers one could choose from.
Matlab (and Octave) has terrible performance in every benchmark comparison I've ever seen. For example see the Julia benchmarks: http://julialang.org/benchmarks/
Have you only seen one benchmark or did you choose such a biased source for trolling?
How I know this graph has nothing to do with machine learning? It has JavaScript listed and JavaScript is on the faster end of the chart.
Do you really expect anyone to believe that a single threaded JavaScript is going to outperform python or R with highly parallel machine learning algorithms?
