Horizontal nystagmus is a common symptom accompanied by benign paroxysmal positional vertigo. Effective recognition of horizontal nystagmus can assist the doctors in clinical diagnosis of related diseases. Based on this, a new method of horizontal nystagmus recognition is proposed based on machine learning in the paper. This network structure consists of five blocks. The attention mechanism is added before the convolution operation in the module, which is helpful for the feature expression and recognition of nystagmus. The activation function after each convolution calculation is the mish activation function, which avoids the problem that the gradient rapidly increases or disappears in the process of gradient processing. In order to prevent local information loss caused by multi-layer dilated convolution superposition, residual structure is introduced to preserve the detailed original data features. In block 4, a new kind of pooling layer is used to achieve better and faster convergence of the network. The recognition accuracy of the proposed method for horizontal nystagmus is 1.94% higher than other methods. From a series of experimental results in this paper, the designed method can effectively recognize horizontal nystagmus. Compared with other processing methods, the designed method has better discrimination effect of nystagmus.