This page points to useful snippets of code (gists in github-speak) that implement features related to scikit-learn but that might not be suitable for inclusion (yet):

• Multi-Layer-Perceptron (neural network) classifier, trained with SGD. Could serve as a basis for a Cython-Version for later inclusion.
• Generating data with non-parametric Gaussian mixture models. Useful if you need "random" data that should have non-trivial structure.
• Adaptive Lasso (should be added to linear_model)
• Non-Negative Garotte (should be added to linear_model)
• Kernel SGD
• Fuzzy K-means and K-medians