Emotional regulation improves deep learning-based image classification
Summary
A novel framework called Emotional Regulation significantly improves deep learning-based image classification by modeling emotion through artificial subjective experience. This method employs pre-training on affective stimuli, balancing non-emotional and emotionally-influenced responses during downstream task optimization. Extensive experiments using ResNet and ViT architectures, pre-trained on four emotional datasets (EMOd, Diffused-EMOd, Abstract, EmoSet), demonstrated consistent improvements on CIFAR-10 and -100 benchmarks. Learnable regulation configurations, such as Learn-E-Reg and Rand-Learn-E-Reg, achieved the best results, outperforming original backbones by up to 1.57% on CIFAR-10 and 3.19% on CIFAR-100. The framework also established a new state-of-the-art in emotion-augmented deep learning for large-scale vision datasets, surpassing previous models by 77.77% and 213.48% accuracy on CIFAR-10 and -100, respectively.
Key takeaway
For AI Scientists and Machine Learning Engineers aiming to enhance image classification generalization, consider integrating the Emotional Regulation framework. This approach, which models artificial subjective experience and employs learnable regulation, has demonstrated state-of-the-art performance on CIFAR datasets. You should explore pre-training emotional encoders on diverse affective stimuli and implement dynamic regulation to balance emotional and non-emotional model outputs, potentially unlocking significant accuracy gains in your computer vision applications.
Key insights
Emotional Regulation, using artificial subjective experience, significantly enhances deep learning image classification performance.
Principles
- Emotion-augmented deep learning improves generalization in neural networks.
- Subjective emotional experience is a critical, often overlooked, factor in AI emotion modeling.
- Balancing emotional and non-emotional predictions through regulation enhances learning.
Method
The Emotional Regulation framework uses three encoders (Non-emotional, Emotionally-influenced, Emotional). The Emotional Encoder is pre-trained on affective stimuli, and a learnable regulation function weights non-emotional and emotionally-influenced predictions based on the elicited affective state.
In practice
- Pre-train emotional encoders on diverse datasets like EMOd, Abstract, or EmoSet.
- Implement learnable regulation for dynamic weighting of emotional and non-emotional model outputs.
- Utilize ResNet-50 or ViT-B/16 as backbone architectures for vision tasks.
Topics
- Emotional Regulation
- Deep Learning
- Image Classification
- Emotion-Augmented AI
- Vision Transformers
- ResNet
- Artificial Subjectivity
Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.