Teach machine to learn: hand-drawn multi-symbol sketch recognition in one-shot

Abstract

The ability to sequentially learn from few examples and re-utilize previous knowledge is an important milestone on the path to artificial general intelligence. In this paper, we propose Teach Machine to Learn (TML), a few-shot learning model for hand-drawn multi-symbol sketch recognition. The model decomposes multi-symbol sketch into stroke primitives and then explains the observed sequences in a bayesian criterion. A Bidirectional Long Short Term Memory (BiLSTM) encoder is employed for stroke primitives encoding. Meanwhile, a probabilistic Hidden Markov Model (HMM) is constructed for complete sketch inference and recognition. The challenging task of hand-drawn multi-symbol sketch recognition is implemented on two public datasets. The comparative results indicate that the proposed method outperforms the currently booming image-based deep models in recognition accuracy. Furthermore, our method is capable to continuously learn new concepts even in one-shot. The codes are currently available in https://github.com/chongyupan/Teach-Machine-to-Learn.