Refine the Data Used

Thanks to our unique feature of global attributions, it is possible to detect and eliminate superfluous input factors, or reduce lookback length for times series, to build more efficient and less complex models, reducing computational demands during inference.

In a use case for an industrial partner, we trained an initial model and then retained only 10% (50/512) of the most important variables to train a new, reduced model. The performance of the new model was 99.55% (compared to 99.97% for the original one) and inferences ended up being 35 times faster.

Improve Specific Areas of the Model

Identify the most robust decision within your AI model with pinpoint precision to help data scientists and stakeholders understand model behavior across diverse scenarios. Improve specific decisions that lead to errors or need robustness due to possible biases, security reasons, or other business constraints. This leads to improved accuracy, reduced risks, facilitated compliance and increased trust in decision-making.

Run Root Cause Analysis

From model design, drill down on problem areas to uncover the root causes of under-performing segments. This enables targeted improvements and enhances overall model performance. At inference time, explain past predictions for auditing and investigating critical scenarios and incidents, simplifying regulatory compliance, and minimizing financial and legal risks.

Tune Model Complexity

Manage underfitting or overfitting in predictive regions to ensure the model is neither too simple nor too complex, while optimally serving the objectives and requirements of internal customers. Achieve the right balance for optimal performance, balancing training time, inference time, computational demand, and model performance and adaptability to new scenarios in live use.