In neuroscience, the detection of cognitive load is important for delving into the neuroscience behind cognitive processes. The human brain is thought of as a complex system that can be in a number of states at once. It is a well-established fact that mental effort demands more cognitive ability. In this work a publicly available multi-arithmetic task electroencephalogram (MAT-EEG) dataset was utilized. There are 36 people in total, both at rest and performing an arithmetic activity. Pre-trained convolution neural network (CNN) architectures detect cognitive load in the proposed research. EEG data pre-train CNN on network designs. This article included the loss function to decrease error in pre-trained networks with new versions of DenseNet201, InceptionNetV4, and EfficientNetB7. EfficientNetB7 achieved an accuracy (ACC%) of 99.08 with a minimum error loss of 0.0297. The results of the proposed strategy are better than the other approaches currently used to detect cognitive load.