適應性機器學習架構於電腦視覺與機器人學習之應用

Abstract

雖然 AI 模型已在各種應用中展現出卓越的效能，但在未知的真實場景中使用這些模型仍然是一項重大挑戰。本論文致力於開發一個機器學習架構，透過針對資料、模型與損失函數三個關鍵項目進行改進，以提升模型在這類環境中的適應能力。

一個典型的機器學習架構包含資料、神經網路模型，以及引導模型使用資料進行優化的損失函數。然而，在真實世界中，每一個項目都可能和理想的使用條件不同，因而需要進行適應才能有效運用這些模型。

在資料方面，預先收集的訓練樣本通常與實際使用時觀察到的資料分佈不同，因此需要使用一些技術來彌合這一差距。在模型方面，由於模型的訓練通常需要大量時間與計算資源，將預訓練模型適應於新任務可以顯著擴展其在不同領域中的應用能力。在損失函數方面，針對特定的應用學習損失函數，可以使模型更有效地利用該函數的優勢。

本論文聚焦於電腦視覺與機器人應用，通過探討特徵解耦、元學習、模型微調以及模仿學習等技術，提出針對上述挑戰的適應性解決方案。

While AI models have demonstrated remarkable effectiveness across various applications, deploying them in unstructured real-world scenarios remains a significant challenge. This thesis focuses on developing a machine learning pipeline designed to enhance adaptability in such environments by addressing three key dimensions: data, models, and learning objectives.

A typical machine learning pipeline consists of a dataset, a neural network model, and a learning objective that guides the model's optimization using the data. However, in real-world scenarios, each of these components may deviate from ideal conditions, necessitating adaptation for effective application.

For data, the distribution of pre-collected training samples often differs from the distribution encountered during inference, requiring strategies to bridge this gap. For models, since the training of a model typically requires substantial time and computational resources, adapting a pretrained model to new tasks significantly expands its applicability across diverse domains. For learning objectives, adapting the objective function to a particular application allows the model to leverage the advantages of the chosen objective more effectively.

This thesis focuses on computer vision and robotic applications and proposes adaptive solutions for the above challenges by exploring techniques such as feature disentanglement, meta-learning, model fine-tuning, and imitation learning.