Nowadays, due to the increasing demand for data processing and its many applications, machine learning methods, especially the branch of deep learning, have received much attention. In addition to increasing the accuracy of existing solutions, reducing the response time on edge devices is one of the issues that is highly regarded. One of the traditional methods of using deep networks is to completely implement them on the end devices. This maintains data security, but the main problem is the low power of the device to perform complex calculations. This suffer a long delay. Therefore, an alternative method called server-based can be used. In this method, due to the high power of servers, the inference speed can be improved. The main problem with this method is the transfer of data from the device to the server. The speed of this process depends on the available bandwidth, which can be very low. The main problem with this method is the transfer of data from the device to the server. The speed of this process depends on the available bandwidth, which can be very low. In recent years, a solution called collaborative intelligence has been introduced to address these challenges. In this method, both the power of the device and the server are used for inference. In most Discriminative networks, as we move through the network, the output volume of the layers decreases, and this property causes the transmission in the middle layers of the network to be less delayed than the original data transmission. In the collaborative intelligence method, according to the available bandwidth and output of each layer, a point for network division is selected. Then the initial part of the network is executed on the device and the resulting output is sent to the server in order to continue the inference operation. The results show that this method helps to reduce the inference time. Researchers in recent years have tried to improve the latency obtained by the collaborative intelligence method by using various strategies. In this thesis, the use of cache along with collaborative intelligence is introduced. Cache allows for some data, only the initial part of the network is done on the device and the final answer is sent to the user at this stage. Therefore, the time spent sending data to the server and inferring on it, is saved and the inference time for all data is reduced on average. To eva‎luate this method, alexnet network and twelve classes of imagenet database have been used and tested under different bandwidths. Tests performed on the proposed method show that, depending on the available bandwidth, up to 53% improvement in inference latency is obtained compared to using the simple collaborative intelligence method. Another advantage of the proposed method is that it solves a wide range of problems, and several of the previously proposed methods are a special case of our method

عبدالرضا ميرزايي

مهران صفاياني

محمدعلي خسروي فرد، مريم ذكري