Dynamic Downscaling of Feature Inputs in Convolutional Neural Networks for Spatio-Temporal Models

Music City Center 

Standard CNNs inherently build hierarchical representations, yet their fixed receptive fields and pooling operations can limit effective multi-resolution trend capture, often missing large scale trends in favor of overfitting and finding false signal from local features. We propose an agent-driven, progressive training method that overcomes these limitations by dynamically adapting the model to explicitly integrate multi-resolution inputs. Starting with a low-resolution base model that captures coarse global features, an adaptive agent monitors performance metrics to decide when to incorporate higher-resolution inputs and add additional layers. Applied to Sea Surface Temperature data for El Niño prediction, our method effectively balances global context with local detail, leading to improved predictive performance. We further evaluate our approach on a traditional image detection dataset and compare it with existing multi-resolution techniques in CNNs, such as dilated convolutions, skip connections, and feature pyramid networks.

Spatio-Temporal Forecasting

Convolutional Neural Networks

Multi-Resolution Models

Agent-based Modeling 

Section on Statistics and the Environment