QQ交谈 +852 2655 6100 enquiry@clustertech.com


ClusterTech has developed a highly optimized software library CNNL (ClusterTech Neural Network Library) which consists of Bayesian Multilayer Perceptron (BMLP) and Hierarchical Mixture of Expert (HME).

BMLP is excellent for learning non-linear relationships in multivariate data sets, such as financial time series predictions and TV audience sizes as a function of its input factors. BMLP is more sophisticated than the ordinary multilayer perceptron (MLP) commonly provided in other data mining software suites and uses a probability distribution to quantify the uncertainty of the model with respect to the observed data allowing better understanding of the model and data. In particular, the Bayesian approach has the following major advantages:

-Whereas a MLP estimates the expected value of an input, since Bayesian techniques model the probability distribution, one can also obtain the variance. This enables error bars from the model to be obtained.
-Using the Bayesian approach allows one to favour simpler models as opposed to complex models. Since a distribution rather than a single expected value is available, parameter uncertainty is quantified and overfitting can be addressed without the need for a cross validation set. This allows all of the data to be used for training.
-Automatic relevance determination (ARD) can be used to infer salient inputs and automatically disable the irrelevant inputs.
-In our experience, the BNN gives better results than the MLP in all non-trivial cases.
-HME is suitable for combining the output of several models, for example, a BMLP and a SVM regressor. Very often, different models give better predictions for different input factor patterns. For example, in financial time series modeling, a model performs better when the market is ranging while another model performs better when the market is trending. HME uses an Expectation-Maximization (EM) algorithm to learn which model performs better in which range of input parameter values and assigns a greater weight to a more accurate model in the combination of model output.

CNNL has been used in several data modeling projects and deployed to clients for production use.

Improve your experience on our website

We notice that your current browser is out of date. In order to improve your experience on our website, we suggest that you upgrade your browser. Upgrading should only take a few minutes. If you are already using the most updated version, please turn off the Compatibility Mode of your browser.





<a target="_blank" href="http://www.getfirefox.com/">Mozilla Firefox</a><a target="_blank" href="http://www.google.com/chrome/">Google Chrome</a>