利用遷移和元學習提升弱監督搜索效能

Abstract

弱監督搜尋在原理上具有以下兩個優點：既能夠在實驗數據上進行訓練，又能夠學習到獨特的信號特性。然而，由於在弱監督下成功訓練神經網絡可能需要大量的信號，因此這種搜尋策略的實際應用性受到嚴重限制。在本研究中，我們嘗試開發更高效和更智能的神經網絡，通過利用遷移學習和元學習來從較少的實驗數據信號中學習。其基本思路是首先在模擬數據上訓練神經網絡，學習關鍵概念並成為更高效的學習者。隨後，神經網絡再在實際數據上進行訓練，通過利用從模擬中獲得的知識和概念，期望能夠在學習中需要較少的信號。我們發現，遷移學習和元學習可以顯著提高弱監督搜尋的性能。

Weak supervision searches have in principle the advantages of both being able to train on experimental data and being able to learn distinctive signal properties. However, because successfully training a neural network under weak supervision can require a large amount of signal, the practical applicability of this search strategy is seriously limited. In this study, we try to develop more efficient and smarter neural networks that can learn from less signal in the experimental data by utilizing transfer learning and meta-learning. The general idea is to first train a neural network on simulations, learning critical concepts and becoming a more efficient learner. Subsequently, the neural network is trained on real data and, by exploiting the knowledge and concepts acquired from simulations, should hopefully require less signals to learn. We find that transfer and meta-learning can substantially improve the performance of weak supervision searches.