Exploring the emergence of temperature concept within deep neural networks through next-frame image prediction

Abstract

Deep learning has revolutionized numerous domains by enabling the creation of powerful models capable of learning complex patterns from vast amounts of data. However, the intrinsic opacity of deep neural networks has raised concerns about their decision-making processes, limiting their application in critical domains such as healthcare, finance and justice. The study of explainability in deep learning models aims to shed light on the inner workings of these models, enabling us to understand how they arrive at their predictions and uncovering emergent concepts that are not explicitly present in the training data. A study of the emergence of the concept of temperature was performed by forcing a reconstruction and prediction of the next frame in images of physical systems of balls with movement. The goal was to determine whether the neural network could acquire an understanding of temperature. Experimental observations were performed with architectures such as Convolutional Autoencoders and U-Net, but the neural networks did not accomplish the task of learning the temperature concept due to many needed concepts that were not considered initially. Despite that, a prediction of the temperature value was performed by using the optical flow of the system.