1: \begin{abstract}

2: We present the first public release of our generic neural network

3: training algorithm, called {\sc SkyNet}. This efficient and robust

4: machine learning tool is able to train large and deep feed-forward

5: neural networks, including autoencoders, for use in a wide range of

6: supervised and unsupervised learning applications, such as regression,

7: classification, density estimation, clustering and dimensionality

8: reduction. {\sc SkyNet} uses a `pre-training' method to

9: obtain a set of network parameters that has empirically been shown to be close to a good solution, 

10: followed by further optimisation

11: using a regularised variant of Newton's method, where the level of regularisation is determined

12: and adjusted automatically; the

13: latter uses second-order derivative information to improve

14: convergence, but without the need to evaluate or store the full

15: Hessian matrix, by using a fast approximate method to calculate

16: Hessian-vector products. This combination of methods allows for the

17: training of complicated networks that are difficult to optimise using

18: standard backpropagation techniques. {\sc SkyNet} employs convergence

19: criteria that naturally prevent overfitting, and also includes a fast

20: algorithm for estimating the accuracy of network outputs. The utility

21: and flexibility of {\sc SkyNet} are demonstrated by application to a

22: number of toy problems, and to astronomical problems focusing on the

23: recovery of structure from blurred and noisy images, the

24: identification of gamma-ray bursters, and the compression and

25: denoising of galaxy images. The {\sc SkyNet} software, which is

26: implemented in standard ANSI C and fully parallelised using MPI, is

27: available at {\tt http://www.mrao.cam.ac.uk/software/skynet/}.

28: \end{abstract}

29: