Effects of different activation functions for unsupervised convolutional LSTM spatiotemporal learning

Abstract

Convolutional LSTMs are widely used for spatiotemporal prediction. We study the effect of using different activation functions for two types of units within convolutional LSTM modules, namely gate units and non-gate units. The research provides guidance for choosing the best activation function to use in convolutional LSTMs for video prediction. Moreover, this paper studies the behavior of the gate activation and unit activation functions for spatiotemporal training. Our methodology studies the different non-linear activation functions used deep learning APIs (such as Keras and Tensorflow) using the moving MNIST dataset which is a baseline for video prediction problems. Our new results indicate that: 1) the convolutional LSTM gate activations are responsible for learning the movement trajectory within a video frame sequence; and, 2) the non-gate units are responsible for learning the precise shape and characteristics of each object within the video sequence.