基於最大化邊界的稀疏編碼

Maximum Margin Sparse Coding

趙紘陞(Hung-Shung Chao)

指導教授 ： 李嘉晃 ； 劉建良

國立交通大學/資訊學院/資訊科學與工程研究所/碩士(2014年)

https://doi.org/10.6842/NCTU.2014.00848

摘要

在本論文中，我們提出了一個基於最大化邊界的稀疏編碼演算法，同時考慮了重構損失與hinge損失，以達到增加分類準確率。基於最大化邊界的稀疏表示法可以類比於支持向量機器的核函數技巧與最大化邊界性質，也解釋了為什麼我們的演算法可以在分類問題中表現出色。
在這次論文中我們將會使用座標下降法去學習出模型中所需要的元素並且證明出我們演算法的收斂性。實驗方面，我們做了有關人臉辨識與物品分類的實驗，並且與其他方法比較準確率。實驗的結果顯示我們的實驗比其他的演算法效果還要好，此外，大部分的研究都表示字典的學習都需要過完備的字典數量才能達到較好的效果。然而，當處理資料集的特徵很大時，過完備的字典數量是非常耗時的。我們的實驗結果顯示我們的演算法在字典數量非過完備的時候表現也是相當不錯的，是非常好的優勢在處理高量特徵的資料集。

關鍵字

最大化邊界 ； 稀疏編碼 ； 坐標下降法

並列摘要

In this work, a novel sparse coding algorithm with max margin, simultaneously considering reconstruction loss and hinge loss, is proposed to enhance classification accuracy. The sparse representation with max margin is analogous to SVM kernel trick and maximum margin properties, explaining why the proposed algorithm perform well in classification task. In this work we use coordinate descent to learn all the elements of the proposed model, and prove the convergence for the proposed algorithm. Furthermore, this work conducts some experiments with face recognition and object recognition images, and compares the proposed algorithm with several algorithms. The experimental results show that the proposed algorithm leads to state-of-the-art performance. Besides, most previous studies about dictionary learning prefer using overcomplete dictionary to improve performance. However, overcomplete dictionary learning is computationally intensive when the dimension of input data is huge. The experimental results indicate that the proposed algorithm still work well without overcomplete dictionary, providing a competitiveness in speed to deal with high-dimensional data.

並列關鍵字

Max-margin ； Sparse Coding ； coordinate descent

參考文獻

in scene images with unsupervised feature learning. In Proceedings of the 2011 In-

ternational Conference on Document Analysis and Recognition, ICDAR '11, pages

440{445, Washington, DC, USA, 2011. IEEE Computer Society.

[2] Meng Yang, Lei Zhang, Jian Yang, and David Zhang. Robust sparse coding for face

[3] Mark Gales and Steve Young. The application of hidden markov models in speech