CURRENNT

Features

• Uni- and bidirectional Long Short-Term Memory (LSTM) layers with forget gates and peepholes
• Feedforward layers with tanh, logistic sigmoid and softmax activation functions
• Deep neural network architectures supported
• Cached on-line learning from large data sets (training data does not need to fit in main memory)
• Reads training data from NetCDF files
• Gradient descent with momentum
• Supports on-line, batch and hybrid on-line/batch learning
• Minimization of cross-entropy and squared error objectives
• Supports regression and binary/multiclass classification tasks
• Training with input activation noise for improved generalization
• Autosave after each training epoch

Follow CURRENNT

CURRENNT Web Site