Stress in Artificial Neural Networks : An explorative stress implementation through emotional weights

Sammanfattning

The exploration of human neurology has been integral to the development of machine learning with neural networks. Initially inspired by the firing of neurons, neural network have seen significant improvements by incorporating cognitive aspects like memory with LSTM, and attention with transformer methods. Drawing from human cognition to enhance neural networks has led to significant advancements in language processing and generative tasks. We explore the integration of stress into neural networks, motivated by its significance for attention and task management. Similar to memory and attention, we propose to directly integrate into neural networks and explore implementation by creating a custom self-attention module for Transformers. The aim is to enhance neural networks' adaptability to environmental challenges by identifying threats and developing strong responses to these. By incorporating stress, we seek to improve relational reasoning abilities in neural networks, crucial for effective resource management and learning. Incorporating more cognitive concepts also aims to make neural networks utilize human neurological strengths as emotions. Two models were developed—one without stress implementation and the other with stress—to assess the architectural changes' effects. Both models were trained and tested in the StarCraft II environment, known for its dynamic and challenging gameplay. Stress was implemented in the self-attention module of a Transformer layer. An appraisal method for threats and resources was implemented to initiate three different physiological states. Threat weights were learnt through sampling attentional outputs of the self-attention module and modifying it by the game outcome. Queries were linearly transformed with these weights to represent a Threat value for different elements. Resources was determined by the key-query calculation in the dot-product where each pair was seen as confidence in resources. Evaluation then either lowered the Attention if threat-resources were different to simulate apathetic or distressed states, while similar evaluations caused a challenging state increasing focus toward the element. During training both models showed negligible differences in fitting to the dataset and processing speed. Neither model fitted the entire dataset showing that further training could show further differences in progress. When evaluating the models both models showed relatively poor performance only being able to beat bots of the lowest difficulty. The Baseline model won more games and lost more games, while the stress model resulted in timeouts. The stress model showed some effects of stress in regards to changing its strategy to be more cautious and conservative. Stress as a phenomenon in neural networks show possible use cases of changing model strategy to be cautious. Our implementation requires explicit goals limiting it's use cases in domains as NLP. Some aspects of stress remains unexplored such as layer based effects. Overall stress shows promise of affecting agent behaviour with additional testing and fine-tuning required