Emotion was found to improve memory and learning under certain conditions. In the context of deep learning, many neural models achieved competitive performances by considering the emotional factor in solving tasks of interest. Among them, investigations concerning the introduction of emotion for solving image classification tasks provided significant results. However, to our knowledge, a study on the impact of emotion on solving image classification through trainable encoders has never been conducted, yet. To perform experiments, the present study proposes the Emotional Regulation approach, which mainly consists of selecting non-emotional and emotionally-influenced representations according to a criterion. In particular, emotionally-influenced representations are learned by concatenating original embeddings with the representations obtained from a frozen emotional encoder. Experiments were performed by employing ResNet-50 and ViT-B/16 architectures, assuming CIFAR-10 and -100 as target datasets for training and evaluation. A set of emotional stimuli was employed to provide an emotional history, while the regulation process was conditioned on positive, neutral, and negative semantics. The results show that our approach improved the original backbones in classifying the considered target datasets, providing evidence for the effectiveness of emotion in supporting image classification based on deep learning.
An Investigation of the Impact of Emotion in Image Classification Based on Deep Learning
Chinnici M.;
2024-01-01
Abstract
Emotion was found to improve memory and learning under certain conditions. In the context of deep learning, many neural models achieved competitive performances by considering the emotional factor in solving tasks of interest. Among them, investigations concerning the introduction of emotion for solving image classification tasks provided significant results. However, to our knowledge, a study on the impact of emotion on solving image classification through trainable encoders has never been conducted, yet. To perform experiments, the present study proposes the Emotional Regulation approach, which mainly consists of selecting non-emotional and emotionally-influenced representations according to a criterion. In particular, emotionally-influenced representations are learned by concatenating original embeddings with the representations obtained from a frozen emotional encoder. Experiments were performed by employing ResNet-50 and ViT-B/16 architectures, assuming CIFAR-10 and -100 as target datasets for training and evaluation. A set of emotional stimuli was employed to provide an emotional history, while the regulation process was conditioned on positive, neutral, and negative semantics. The results show that our approach improved the original backbones in classifying the considered target datasets, providing evidence for the effectiveness of emotion in supporting image classification based on deep learning.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
