Nowadays, there is a great interest in artificial intelligence. One of its popular branches is deep learning, based on the neural networks consisting of artificial neurons, which are adders with a nonlinear activation function. This approach requires large amounts of training data. For example, computer vision contains different tasks of this nature, where one of the well-known problems is recognizing images of handwritten numbers. Various approaches to training convolutional neural networks on the MNIST database show less than 0.5% error. Another area of artificial intelligence is neuromorphic computing, whose goal is to build biologically plausible models of neurons in the human brain. The motivation for the approach of neuromorphic computation is that accurate modeling of biological neurons will allow solving many problems, in particular, control problems, with the same accuracy the human does. Moreover, it opens up the possibility of creating electrical circuits that mimic the behavior of biological neurons, which will reduce energy consumption. In particular, the physical realization of neurons can be coupled oscillators that are ubiquitous in nature. The phase or frequency of the oscillators contains information that allows development of such computational schemes. In this paper, a hybrid antiferromagnetic-heavy metal structure is used to solve the problem of image recognition. During the model training the coupling coefficients of the oscillators are calculated so that pairs of oscillators corresponding to different pixel values oscillate asynchronously, and the oscillators reacting to the same value set the oscillation mode in phase. Pattern recognition error is estimated depending on the number of neurons, patterns, and signal-to-noise ratio (SNR).