Statistical inference: learning in artificial neural networks.

Area of Science:

• Computational neuroscience
• Machine learning
• Statistical modeling

Background:

• Artificial neural networks (ANNs) are integral to modeling neural activities and cognitive functions.
• Statistical inference offers an objective framework for developing and assessing ANN learning algorithms.
• Overfitting is a key challenge in ANN training, necessitating robust model selection strategies.

Purpose of the Study:

• To review the application of statistical inference in artificial neural network (ANN) learning.
• To discuss methods for addressing the overfitting problem in ANNs.
• To cover both supervised and unsupervised learning algorithms for ANNs.

Main Methods:

• Review of statistical inference techniques for ANN training and evaluation.
• Discussion of model-selection methods (conventional and Bayesian) for overfitting.
• Examination of supervised learning (nonlinear regression) and unsupervised learning (Hebbian law, global objective functions).

Main Results:

• Statistical inference provides objective learning algorithms for ANNs.
• Model-selection techniques effectively address the overfitting problem.
• Ensemble methods like bagging and arching improve ANN predictor performance.

Conclusions:

• Statistical inference is crucial for robust ANN learning and evaluation.
• Both supervised and unsupervised learning algorithms can be effectively implemented using statistical principles.
• Advanced techniques enhance ANN generalization and predictive power.