It is quite simple - machine learning is very CPU time consuming task. C++ is still much faster than Java or C#, and with MPI you can parallelize this task for very large clusters.

I don't know about MATLAB efficiency, but it is probably much easy to use than C++ (especially for newbies)

It is quite simple - machine learning is very CPU time consuming task. C++ is still much faster than Java or C#, and with MPI you can parallelize this task for very large clusters.

I don't know about MATLAB efficiency, but it is probably much easy to use than C++ (especially for newbies)

Use whatever language you are most comfortable with. If you are obliged to do your own implementation of neural nets, then it is important to be familiar with the implementation language. There should not be a performance issue with neural nets.

I can only agree with Wojciech Borkowski. Often the main issues with machine learning is speed, and the execution of C++ compiled code might be orders of magnitude faster than Java. This is even more true for stochastic approaches, where you'd have to run the same experiment several times to obtain meaningful statistics on the effectiveness of the methodology.

I am going to go ahead and strongly suggest that you develop in Java over C++. Contrary to the posts here, C++ is actually now slightly slower than Java on most machines. (see: http://stackoverflow.com/questions/145110/c-performance-vs-java-c) The basic reason for this is that C++ has no universally accepted compiler and so the compilers all try to be jack of all trades and end up trading speed for generality as opposed to the compilers for Java that always run on the Java virtual machine allowing for more precise optimization.

Speed and performance are often issues with neural nets, as can easily be seen when reading the details of the SENNA project (see: http://ml.nec-labs.com/senna/) and in fact neural nets have fallen out of favor due to their large degree of computational intensity and are less common today than support vector machines.

The actual reason that MATLAB and C++ are much more common than C and Java is that neural nets use a lot of matrices and MATLAB is the fastest thing out there for matrix math, and C++ is mostly a language that is low level but has native matrix support unlike C and Java. Additionally, C++ has long been seen as the de facto standard when it comes to physics simulations and has a number of spectacular libraries for matrix math.

Additionally, there is a history with C++ and MATLAB for neural nets in particular which means that there is more legacy code for these languages making it easier to continue to develop neural nets in these languages.

tl;dr: Java is faster than C++, but C++ has a lot of fantastic legacy code base and matrix math libraries that make it an attractive language for developing neural networks. If you don't want to heavily rely on these however, it doesn't make sense to use C++ especially if you aren't familiar with it.

Robotics is strongly related to hard real time processes. In consequence you have to use a programming language, in conjunction with a appropriate OS, which allows you to control all related issues. Programming languages like java being interpreted in a virtual machine do not offer the related functionality. The core process at least have to be implemented in c++.

C++ provides less help and more freedom than do C# or Java. Note that all of these languages hgave become so complicated in recent years that very few professional programmers know or use most of the features in any of these languages. Someone should develop a tool that uses only the commonly used subset of one of these languages and perhaps call it Useful-C++ or Useful-Java or Useful-C#.

If you compile for specific targets and know how to use the language, C++ is definitely faster than Java and much faster than C#. But that is beside the point.

The main reason MATLAB and C/C++ are used is because of a large amount of existing libraries, ease of use and rapid protoyping (in case of MATLAB). Also, Python is now emerging as an important language for machine learning applications, especially through scipy, numpy, and theano.

If you do not have access to matlab, you can use Octave: http://www.gnu.org/software/octave/

It is an open source implementation of MATLAB that is something like 99% compatible with it. If you are using functions that are not in Octave, you can usually find them in the add-on packages in octave-forge: http://octave.sourceforge.net/

In addition, certain very specialized functions can be taken directly from the website of mathworks.com, however you should note that this is only legal if you have a valid MATLAB license. I have found that I can run virtually any MATLAB script in octave, it has never let me down. Good luck.

You should use the language that you are most comfortable with to start with, unless you are considering working on huge amounts of data. Matlab/Python is mostly used for prototyping, and Python seems to have larger library support, although it does not fly well with threads. C++ is the fastest OOP language in almost all tests to day. Most neural networks algorithms are doing some gradient descent and it is quite simple to optmize at compile time. There are also some implementations that are quite good. Checkout FANN and Shark. In c++ you can also directly interface to LAPACK and BLAS if you really want to go faster. BOOST also provides nice BLAS machinery with uBLAS. Bottom line it depends on which language you are best at, but the rule of thumb is C++ for performance, JAVA for ease of development.

C# was build by Microsoft to ease the life of an average programmer. Most of the feature of C# are build to result in lesser code, by the programmer. the debugger has some layers to convert the C# to the intermediate language. this makes the end resultant code larger than its equivalent c++ code. Moreover C# is managed code. But in C++ you can still write pointers. C++ is more system centric than C#.

...I thought that c# was written by Microsoft to take on Sun's Java...

Maybe, C# is very similar to Java on the level of abstraction and domain of usage, but in opposite to Java work only on Microsoft Windows (there is open port to Linux, but rather without worldwide / business importance.)

Hi John and WOJCIECH

The idea of "write once and run anywhere" was tossed by Java. this however did not had much of practical or business advantage. There are many performance pitfalls and OS dependency with this concept.

Microsoft on the other hand developed a different approach. They said :"write in any language but run on windows". So Microsoft brought most of the languages on CLR. In Short: if you can make a programming language produce MSIL code, It will run on .NET framework. What it needs only, is a interpreter (Any language to MSIL , please Check the below list).

If you check the architecture of .NET and Java They are quite different. please refer the following link

http://arunraj.orgfree.com/arunraj/images/DotNetArchitecture.gif

For Java:

http://www.write-technical.com/126581/session1/lecture1/index.8.gif

The following Languages are supported in .NET:

A#: CLI implementation of Ada.

Boo: inspired by Python.

C#: Most widely used CLI language,

C++/CLI: A version of C++

Cobra: A CLI language with static and dynamic typing

Component Pascal: A CLI-compliant Oberon dialect.

ClojureCLR: A native implementation of Clojure (a Lisp dialect)

DL: Language using the Common Language Runtime (CLR)

Eiffel: Purely object-oriented language

F#: A multi-paradigm CLI language supporting functional programming

IronJS: An open-source CLI implementation of ECMAScript, built on the DLR.

IronPython: An open-source CLI implementation of Python, built on the DLR.

IronRuby: An open-source CLI implementation of Ruby, built on the DLR.

IronLisp: A CLI implementation of Lisp.

J#: A CLI-compliant implementation of Java.

JScript .NET: A CLI implementation of ECMAScript version 3,

L#: A CLI implementation of Lisp.

Managed Extensions for C++: A version of C++ targeting the CLR. Deprecated in favor of C++/CLI.

Managed JScript: A CLI implementation of JScript built on the DLR. Conforms to ECMAScript version 3.

Nemerle: A multi-paradigm language similar to C#, OCaml and Lisp.

Oxygene: An Object Pascal-based CLI language.

P#: A CLI implementation of Prolog.

Phalanger: An implementation of PHP with extensions for ASP.NET

Phrogram: A custom CLI language for beginners and intermediate users produced by The Phrogram Company

PowerBuilder: Can target CLI since version 11.1.

Synergy DBL .NET: an object oriented CLI compliant implementation of DBL and DIBOL produced by Synergex.

Team Developer: SQLWindows Application Language (SAL) since Team Developer 6.0.

VBx: A dynamic version of Visual Basic .NET built on the DLR.

Visual Basic .NET (VB.NET): A redesigned dialect of Visual Basic.

Visual COBOL: A CLI compliant implementation of COBOL with object-oriented extensions.

Windows PowerShell: An object-oriented command-line shell.

Others:

COBOL for Microsoft .NET.

Perl for Microsoft .NET.

Eiffel for Microsoft .NET.

Python for Microsoft .NET.

Pascal for Microsoft .NET.

Mercury for Microsoft .NET.

Mondrian for Microsoft .NET.

Oberon for Microsoft .NET.

Salford FTN95 (Fortran) for Microsoft .NET.

SmallTalk for Microsoft .NET.

Standard ML for Microsoft .NET.

Dyalog APL for Microsoft .NET.

C++ is more difficult, but using C++ I can make almost all that I can imagine on particular hardware (from OS to high level expert system). Using any of .NET languages, with many different syntaxes, I can explore only what is possible using .NET. Similarly using Java I can do only, what is possible using JVM... And every run I have to pay costs of usage meso-level portable pseudo machine code.

" see http://stackoverflow.com/questions/145110/c-performance-vs-java-c"

Slater Victoroff

This thread is not "considered a good" :-)

"locked by Lasse V. Karlsen Oct 2 '11 at 12:28

This question exists because it has historical significance, but it is not considered a good, on-topic question for this site, so please do not use it as evidence that you can ask similar questions here. This question and its answers are frozen and cannot be changed."

Dear Mirko,

You are right - if you write application for one ad hoc use, milliseconds are not important and worth of additional effort. The clean and simple code is more important. Even two times faster program in not worth - just waiting 2 vs 4 hours for results.... But for professional use, and for scientific program used repeatedly, it is important (7 days vs 14 or even 70! And another 14 or 70 if you found error or you or your boss wanted to check another version of model).

Sometimes I was implementing the same educational models (CA, ABM, GA) both in Java and in C++. And between Java and C++ I had offten 10 times or more difference, even without optimization and using just MSVC++... Of course, when I switched off the most time consuming things such as continuous visualization and disk writing.

So for real models that need lot of calculations I always use C++ - sometimes simplified to almost only C syntax (but with strict type checking, lack of it is one of the first source of strange errors in plain C)

I know VB-dot net and java both. But when I want to go for the implementation part, I feel dot net better option. This utility is far advanced than Java, and sometimes you can get some ready made codes for practicing for any new research.

If performance is not so important Java may be better than C++, but, even JIT Java code have to be outperformed with C++ code because of some immanent properties of Java language - e.g. all object allocated on heap, no pointer arthmetic...

https://en.wikipedia.org/wiki/Comparison_of_Java_and_C%2B%2B#Performance

Sinece the questions was asked in 2013 here are how things look now

https://www.c-sharpcorner.com/article/best-programming-language-for-machine-learning/

So of course this days Python is number 1 and will continue to be for some long ,long period !