Introducing an Auxiliary Information Module into ANN for Distributional Change Adaptation

Training data has a significant impact on the performance of artificial neural network (ANN) models. This becomes evident when the model is used in a variable environment with input data featuring a changing distribution, leading to a low reliable prediction. To address this problem, we propose to move away from the deterministic nature of the ANN and introduce a non-deterministic information module. We aim to enable the model to adapt to distributional changes in the data during the deployment phase based on auxiliary sensory information. We parameterize the weights of the model with an information module that updates their values automatically based on auxiliary sensory information to compensate for the changes in distribution. The update of the weights is made online by a regulator. This means that retraining is not necessary. Furthermore, we show that the proposed information module can be easily integrated into commonly used ANN models. Finally, the results of case studies show the effectiveness and robustness of the proposed approach that yields more reliable predictions.