Conditional Human Activity Signal Generation and Generative Classification with a GPT-2 Model
Abstract
In this work, the results of an exploratory study into the use of the publicly available GPT-2 model for simultaneous conditional human activity signal synthesis and classification are presented. To accomplish this, the small variant of a pre-trained GPT-2 model is fine-tuned on quantized and windowed human activity signal sequences. The conditional generation is achieved by appending and prepending class labels to each window to introduce class information. During the generation phase, the model is prompted either with a class label for synthesizing signals, or a signal sequence for classification by synthesizing a class label. The study is conducted on the publicly available WISDM and UCI-HAR datasets, and the generated signals are evaluated using an LSTM-CNN model. As a classifier, the fine-tuned GPT-2 model achieved an overall accuracy of 90.47% on the WISDM, and 82.29% on the UCI-HAR dataset, which is 4% and 6% lower than the LSTM-CNN model evaluated on the same test subsets. When used as a multi-class generator, on average, 86.33% of the generated data are classified as valid samples by the LSTM-CNN model. While there is room for improvement, these results demonstrate that GPT family architectures can be used for simultaneous conditional signal synthesis and classification, opening up new opportunities for human activity recognition.