The paper presents a novel approach to machine learning using orbital angular momentum (OAM) states in optical neural networks. The authors demonstrate an all-optical convolutional neural network (CNN) based on Laguerre-Gaussian (LG) beam modes, which can extract and encode information features in the OAM domain. The CNN architecture includes a trainable OAM mode-dispersion impulse as a convolutional kernel for feature extraction and deep-learning diffractive layers for classification. This method achieves high accuracy in encoding MNIST database images (97.2%) and resistance to eavesdropping in free-space transmission. Additionally, the OAM mode-dispersion selectivity is extended to multiplexed OAM states for all-optical dimension reduction in anomaly detection with an accuracy of 85%. The work provides insights into the mechanism of machine learning with spatial modes and has potential applications in various machine-vision tasks.The paper presents a novel approach to machine learning using orbital angular momentum (OAM) states in optical neural networks. The authors demonstrate an all-optical convolutional neural network (CNN) based on Laguerre-Gaussian (LG) beam modes, which can extract and encode information features in the OAM domain. The CNN architecture includes a trainable OAM mode-dispersion impulse as a convolutional kernel for feature extraction and deep-learning diffractive layers for classification. This method achieves high accuracy in encoding MNIST database images (97.2%) and resistance to eavesdropping in free-space transmission. Additionally, the OAM mode-dispersion selectivity is extended to multiplexed OAM states for all-optical dimension reduction in anomaly detection with an accuracy of 85%. The work provides insights into the mechanism of machine learning with spatial modes and has potential applications in various machine-vision tasks.