In pattern recognition, samples are directly classified in the feature space or mapped to higher dimensional spaces by kernel functions. In order to improve the classification effect, the following methods are commonly used:1. Solve for better separation surface; 2. More efficient feature space. This paper proposes a Dynamic Hypersphere Algorithm (DHA) for classification that can find a more efficient feature space. DHA uses dynamic feature transformation and optimizes the same type of data into the same hypersphere. This paper proves that DHA can obtain better classification results by finding effective feature space by experimenting on standard data sets. In addition, in order to further prove the validity of the feature space, In this paper, DHA is applied to the MNIST handwriting by reducing the training samples and reducing the original sample 784 to 10 dimensions, a recognition rate of 90.18% is obtained, and good results are also obtained in the unbalanced handwriting.
在模式识别中，样本在特征空间被直接用作来分类或被核函数映射至更高维空间进行分类。为了提高分类效果，一般使用以下方法：1.更好的分隔面；2.更有效的特征空间。本文提出了一种使用神经网络寻找到更有效特征空间的动态超球体算法（Dynamic Hypersphere Algorithm for classification，DHA），DHA采用了动态的特征变换，通过满足本文定义的优化超球体的条件获得更有效的特征空间，最后通过欧氏距离得到分类结果。本文通过在标准数据集上实验证明了DHA能够通过动态的特征变换寻找有效特征空间获得更好的分类效果。此外，为了进一步证明特征空间的有效性，本文将DHA应用到MNIST手写体，通过减少训练样本并且将原样本784维降至10维获得了90.18%的识别率，并且在不平衡手写体中也获得了较好的效果。
论文中手写体形成的特征空间在这里
引入了tf.split，将算法重构为类，并重写了作图机制
增加了 batch_size，增加了算法稳定性和速度，减少了设备压力