ReLU, Sparseness, and the Encoding of Optic Flow in Neural Networks.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Accurate self-motion estimation is crucial for mobile robot navigation, especially in GPS-denied areas.
• Optic flow, derived from camera sensors, is a key method for estimating self-motion.

Purpose of the Study:

• To investigate how different activation functions (ReLU, leaky ReLU, GELU, Mish) affect neural network performance for self-motion estimation from optic flow.
• To analyze the impact on accuracy, robustness, and internal data encoding properties.

Main Methods:

• Trained Convolutional Neural Networks (CNNs) and Multi-layer Perceptrons (MLPs) using optic flow data.
• Evaluated network performance with various activation functions under different conditions.

Main Results:

• Networks utilizing ReLU and leaky ReLU demonstrated superior accuracy in estimating self-motion from novel optic flow patterns.
• These networks also showed enhanced robustness in challenging environmental conditions.
• ReLU and leaky ReLU induced sparser data representations compared to GELU and Mish.

Conclusions:

• ReLU and leaky ReLU activation functions are optimal for enhancing self-motion estimation accuracy and robustness in neural networks.
• The induced sparsity by these functions plays a significant role in improving performance for optic flow-based robotic navigation.