As humans, we learn new emotional intelligence and social behaviour skills throughout our lives by observing, imitating and interacting with others. Furthermore, we continually refine and enhance this stored learning through everyday use and by learning complementary skills and behaviours. Machines capable of displaying these essential anthropomorphic traits remain a core goal in the field of automatic emotion recognition (AER). At the same time, contemporary deep learning techniques are transforming intelligent signal analysis, endowing machines with near-human capabilities in a range of analysis, classification and regression tasks. These advances are raising the general public expectations of modern computing systems including enabling emotional intelligence and bi-directional affective human-computer interactions.

Deep learning is not new to the field of AER, and the benefits of techniques/architectures such as feedforward, recurrent and convolutional networks within the field of AER are already well-established. However, there are many emerging techniques whose applications to AER, regarding improved accuracy and increased anthropomorphism, have yet to be fully realised. For example, zero- and few-shot learning, as well as meta-learning techniques could be used to ensure an AER system can robustly learn from a small number of samples. Alternatively, reinforcement learning techniques can endow an AER system with adaptive behaviours. Similarly, continual lifelong learning techniques could empower an AER system with the ability to learn new skills perpetually. Adversarial training could be used to improve performance against confounding factors limited emotional expressions and unknown environmental conditions during system training. Explainable artificial intelligence techniques can be used to improve the interpretability and understandability of the generated models offering previously unavailable insights into the learning process of AER systems. Finally, research can also reveal the limitations of these techniques and underscore the need for novel computational approaches that both look beyond and complement deep learning.

Therefore, this Research Topic aims to solicit high-quality papers that report recent and emerging research developments in the field of deep learning for automatic emotion recognition. Potential authors are invited to submit original contributions. Topics include, but are not limited to:

- Adaptive approaches based on reward modelling or reinforcement learning,
- Generative adversarial training paradigms for data augmentation,
- Adversarial training for enhanced robustness or latent space analysis,
- (Deep) learning paradigms for zero or low resource settings,
- Fusion strategies based on multimodal attention modules,
- Low-resource networks for mobile applications,
- Multitarget, multitask and transfer learning paradigms,
- Novel end-to-end learning approaches,
- Explainable AI frameworks for AER,
- Multi and cross-modal representation learning strategies.

Submit your idea

Impact

Frontiers in Computer Science

Participating Journals

Impact