An unified framework for learning machines
Most learning machines in Nieme are expressed through the unifying framework of energy-based machines. This framework allows a very clean decomposition of learning methods into four components. The central one is a parameterized architecture, which computes outputs given an input vector. The two next components, the loss and the regularizers define an energy quantifying how good the parameters are, with respect to a given training set. The last component, the learner, searches for parameters that minimize the learning energy.
Many classical learning machines can be cast in the energy-based machines framework. As an example, a Perceptron uses a linear architecture, a Perceptron loss, no regularizers and a stochastic gradient descent learner. Here are some examples of energy-based machines supported by Nieme:
Classification and Ranking
- Large-Margin Perceptrons
- Pegasos Linear SVM
- Batch LogIt Classifiers
- Exponential Loss Ranking Machines
Multiclass Classification
- Maximum Entropy classifiers
- L1-regularized Maxent classifiers
- Multiclass Pegasos Linear SVM
Regression
- Least-square Linear Regression
- L1-regularized regressions