Designing neural networks for modeling biological data: a statistical perspective.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computational Statistics

Background:

• Selecting the optimal hidden layer size is crucial for feedforward neural network performance.
• Existing methods may be susceptible to data snooping, leading to unreliable model selection.
• Parsimonious models with high predictive accuracy are desirable in many applications.

Purpose of the Study:

• To propose a novel strategy for determining the hidden layer size in feedforward neural networks.
• To address the issue of data snooping in model selection.
• To enhance the out-of-sample predictive ability of neural network models.

Main Methods:

• The proposed strategy involves comparing different feedforward neural network models based on their out-of-sample predictive performance.
• A modified reality check scheme is employed to mitigate data snooping concerns, particularly for nested models.
• The procedure is validated using both simulated and real-world datasets.

Main Results:

• The proposed method effectively selects parsimonious neural network models.
• Models selected using this strategy demonstrate superior predictive accuracy.
• The modified reality check successfully overcomes data snooping issues.

Conclusions:

• The presented strategy provides a robust framework for selecting hidden layer sizes in neural networks.
• This approach leads to more reliable and accurate predictive models.
• The method is applicable to diverse datasets, improving neural network design.