Self-Orgnizing Error-Driven Learning (SOEDL)

Most AI systems are trained to maximize prediction accuracy. Businesses, however, make decisions based on expected value, where both the estimated probability and the context of the decision determine the optimal action.

Whether deciding to execute a trade, purchase an asset, approve a loan, accept an insurance policy, or bid on a contract, the best decision depends on more than simply predicting the most likely outcome.

Self-Organizing Deep Learning (SOEDL) is an active research program at DeepDecisionLab focused on developing AI systems that learn representations optimized for decision-making rather than prediction alone.

SOEDL builds upon our earlier research, Self-Organizing and Error-Driven (SOED) Artificial Neural Network, which demonstrated that maximizing prediction quality alone does not necessarily maximize decision quality. The research showed that different populations may require different decision treatments, even when their predicted probabilities appear similar, highlighting the need for learning richer decision-oriented representations.

SOEDL extends this research by investigating how self-organization can learn latent probability representations that better support high-value decisions under uncertainty.

Use cases:

• Algorithmic trading and portfolio allocation

• Investment and acquisition decisions

• Credit approval and risk pricing

• Insurance underwriting

• Bid/no-bid and offer acceptance decisions

At DeepDecisionLab, our goal is not simply to build models that predict better, but to develop AI systems that enable organizations to make better decisions.