The Influence of Symmetry of Synaptic Connections in MLP on the Values of Object Recognition Metrics
Year: 2026
Type: article
Abstract: The paper considers the problem of recognizing objects (anti-tank mines, anti-personnel mines, M14, booby traps) located in soils with different structures (dry sandy, dry calcareous, dry humus, moist sandy, humid and humus, moist calcareous) using MLP with symmetric synaptic connections between neurons of different layers (SMLP) based on data from FLC100 magnetic field sensors with a sensitivity of 10-10 – 10-4 Tesla. To train the SMLP network, a dataset of fluxgate magnetometer measurements in six soil types was used, divided into training (80%) and test (20%) subsets. The input data is encoded as a three-dimensional vector (voltage, height of the sensor above the ground with a mine, one-hot encoding of 6 soil types). The main difference between SMLP and MLP is the imposition of symmetric constraints on the weight matrix of the second hidden layer, which almost halves the number of synaptic connections between neurons in the hidden layers without significantly degrading recognition quality. The proposed symmetrization mechanism mitigates over-parameterization and overfitting in small-scale magnetometric datasets by introducing structural inductive bias and implicit regularization through symmetry constraints in parameter space. In contrast to convolutional architectures developed for high-dimensional spatial representations, the proposed approach is tailored to structured low-dimensional magnetometric measurements. The MLP and SMLP models were trained for 100 epochs with Adam, RMSprop, and SGD optimizers with a learning rate reduction using ReduceLROnPlateau callback. Experimental results show that SMLP achieves an average accuracy of 99.20% and AUC = 0.9996, which is only 0.09% lower than the traditional MLP (99.29%, AUC = 0.9997), but reduces the training time by 20–30% and behaves more stably under different optimizers. Thus, the SMLP model is productive for use on embedded devices with limited resources.
Institution: Lviv Polytechnic National University
Language: English
Subfield: Geophysics
Field: Earth and Planetary Sciences
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7150242597