In Python, for general Machine Learning I really like scikits.learn (http://scikit-learn.sourceforge.net/). The library is Pythonic and includes many common ML tools I need.
I have heard good things about Shogun (http://www.shogun-toolbox.org/), which is written in C++ and has bindings for many other languages, including Python. Shogun appears to be more focused on kernel algorithms (e.g. SVMs).
There's a GSOC project to bind Vowpal Wabbit in Shogun. That would be sweet. This was mentioned by John Langford on the VW mailing list.
Theano (http://deeplearning.net/software/theano/) is great if you are devising your own ML models, particularly ones that are based upon gradient descent. It's Python and it automatically compiles to C/C++ and then machine code for your CPU or GPU.
This looks really fantastic.
We discussed Python ML packages on MetaOptimize: http://metaoptimize.com/qa/questions/3053/python-machine-lea...
Here is a little more detailed summary about Theano, which I wrote up on that thread:
Theano is a CPU and GPU compiler for mathematical expressions in Python. It combines the convenience of NumPy with the speed of optimized native machine language. For gradient-based machine learning algorithms (like training an MLP or convolutional net), Theano is from 1.6x to 7.5x faster than competitive alternatives (including those in C/C++, NumPy, SciPy, and Matlab) when compiled for the CPU and between 6.5x and 44x faster when compiled for the GPU. The SciPy 2010 paper is a good introduction: http://www.iro.umontreal.ca/~lisa/pointeurs/theano_scipy2010...