| Popis: |
There is a large body of work in the behavioural psychology literature studying how an organism’s behavior changes in relation to consequences (reinforcement) from the environment. As all behaviors are an outcome of choice, behavioral research focuses on the study of choice behavior. Machine learning (ML) models may assist behavioral research to further understand the mechanisms of choice behavior. However, behavioral psychology datasets can be small and variable, affecting the ML's ability to generalize with new datasets extracted from different populations and/or behavioral experiments and limiting ML's usefulness in this context. Therefore, in this paper, we tested two transfer learning strategies –feature extraction and fine-tuning– to remove the need to retrain ML models for every new dataset. Our approach allowed our state-of-the-art artificial intelligence model to become adaptable to novel instances. Initially, we trained a single spiking neural network (SNN) to identify an organism’s reinforcement history based on five experimental datasets of pigeon binary decision-making. Then we tested two transfer learning strategies by keeping the underlying patterns of the pre-trained SNN the same (i.e., neuron properties and weights) and adapting only the classifier of the outputs (i.e., firing rates) to suit the new datasets. Lastly, we compared the performance of the transfer learning approaches to our baseline SNN model. Our study demonstrated that knowledge gained from a population (baseline model) could be applied to another population’s dataset without retraining the model each time, regardless of which dataset participated in the training or testing of the SNN model. Currently, there is limited use of transfer learning in SNNs and in animal research. Our results may help develop new approaches in the ‘toolbox’ of psychological research to enhance prediction, independent from the dataset, without consuming significant computational resources. |
