Convolutional neural networks (CNNs) have powerful representation learning capabilities by automatically learning and extracting features directly from inputs. In classification applications, CNN models are typically composed of: convolutional layers, pooling layers, and fully connected (FC) layer(s). In a chain-based deep neural network, the FC layers contain most of the parameters of the network, which affects memory occupancy and computational complexity. For many real-world problems, speeding up inference time is an important matter because of the hardware design implications. To deal with this problem, we propose the replacement of the FC layers with a Hopfield neural network (HNN). The proposed architecture combines both a CNN and an HNN: A pretrained CNN model is used for feature extraction, followed by an HNN, which is considered as an associative memory that saves all features created by the CNN. Then, to deal with the limitation of the storage capacity of the HNN, the proposed work uses multiple HNNs. To optimize this step, the knapsack problem formulation is proposed, and a genetic algorithm (GA) is used solve it. According to the results obtained on the Noisy MNIST Dataset, our work outperformed the state-of-the-art algorithms.