Enhancing Neural Network Performance through Model-Generated Training Signals

Neural networks have been widely applied to tasks such as natural language processing, time series analysis, and multimodal understanding. Their optimization typically requires large volumes of high-quality data, but human annotation is expensive and does not scale effectively. As an alternative, this thesis explores methods that leverage model-generated training signals to either replace or supplement manual labels, enabling robust neural network optimization while reducing reliance on human annotation. This contributes to the development of more auto mated and efficient artificial general intelligence. This thesis demonstrates that leveraging model-generated signals can lead to superior performance across various domains. We validate this methodology through successful applications in three major areas, which form the core components of this thesis: time series forecasting and change-point detection, text classification with limited or no labeled data, and large language model alignment. In Part I:We improve text classification by incorporating model-augmented document content and label descriptions in few-shot and zero-shot learning settings. In Part II: We utilize generated correlation graphs as augmented signals to enhance change-point detection in time series analysis. In Part III: We leverage self-enhancement techniques, such as reinforcement learning, to align large language models for developing more robust chatbots, improving instruction following in video-language models, and enhancing reasoning in vision-language models. Collectively, this work advances neural network optimization through model generated signals across multiple domains, contributing to the development of intelligent AI systems with minimal human supervision