Flipover outperforms dropout in deep learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Artificial neural networks (ANNs) are susceptible to overfitting and noise.
• Standard dropout techniques randomly deactivate neurons to improve generalization.
• There is a need for more effective regularization methods to enhance ANN robustness.

Purpose of the Study:

• Introduce Flipover, an enhanced dropout technique for ANNs.
• Evaluate Flipover's effectiveness in improving model robustness.
• Compare Flipover's performance against conventional dropout.

Main Methods:

• Flipover randomly selects neurons and inverts their outputs with a negative multiplier during training.
• This method provides stronger regularization compared to standard dropout.
• Experiments were conducted across various neural network architectures.

Main Results:

• Flipover effectively mitigates overfitting, achieving performance comparable to or better than dropout.
• The technique significantly amplifies robustness to noisy data.
• Flipover enhances resilience against adversarial attacks on neural networks.

Conclusions:

• Flipover is a highly effective regularization technique for deep learning.
• It offers superior robustness against overfitting, noise, and adversarial attacks.
• Flipover represents a promising advancement in improving ANN reliability.