As regards the recent advancements in the field of technology, libraries, and data science, it has been predicted that in a near future, powerful algorithms will be developed using machine learning techniques. The application of these algorithms is very likely to be used in the prediction areas considering the existing data. In 1991, a study in which a set of dynamic actions was applied to a monkey was conducted. The results of the research indicated that there were neurons located in the F5 area of the monkey’s brain which became activated during an operation such as grasping an object or observing a similar process performed by another person. Any problem in the function of these neurons can lead to several diseases such as autism, schizophrenia, Down syndrome, MS, smoking, and obesity. Besides the mirror neurons, canonical neurons have been also discovered. Canonical neurons are activated only during an activity that is implemented by a person themselves while the neurons are no longer active during another person's activities. Some models have been introduced in order to explain the function of the mirror neurons using the concept of canonical neurons. In these models, it was claimed that in addition to observing the action, the tools associated with that action are helpful in diagnosing that action as well. This way, any information related to the action can affect the action detection process. The main aim of this research is to develop an artificial neural network which has been inspired by the function of mirror neurons for data classification purposes. In this particular research, the combination of the mirror and canonical neurons with the help of the canonical neurons are expected to exhibit better performance than the mirror neurons. To accomplish the final goal of the study, a dataset whose auxiliary features were Separately extractable was chosen and applied to the network as the canonical neurons. Therefore, the MNIST dataset was utilized and the form of writing was considered as the auxiliary feature. In this thesis, the recurrent and convolutional neural networks were combined in two states and according to the results and some eva‎luation criteria, good accuracy (91 and 89.5 percent) was achieved. In the other words, the function of the recurrent network of mirror neurons was improved by 2.5 and 1 percent in the combination with the convolutional network of the canonical neurons.

وحيد سالاري، فرزانه شايق

فرهاد شهبازي

فرهاد فضيله، مجتبي اعلائي