eversvd,+7(2)

Low-cost microcontrollers with constrained processing and memory capabilities are valuable for developing embedded systems. Current market devices with dedicated 32-bit architectures are ideal for Internet of Things (IoT) applications, yet traditional convolutional networks are too complex for these devices. This paper explores an alternative approach using Multilayer Neural Networks (MNNs), which due to their shallower structure, are more suitable for embedded systems. MNNs can model complex decision surfaces for high-dimensional classification tasks, such as handwritten character recognition. This study addresses the absence of a clear architectural strategy for MNNs aligned with specific problems. A performance analysis was conducted focusing on the networks' depth to establish criteria for their optimal size in particular applications. The evaluation utilized the MNIST public database alongside classical performance metrics for classification models. The findings indicate that the depth of the network significantly influences its performance, which aids in the proper selection of network architecture for effective handwritten character recognition.