turkmath.org

Mathematical programming provides a structured framework for formulating and solving optimization problems, which is essential in the training of neural networks, where objective functions—such as the minimization of loss functions—must be effectively minimized. Techniques like gradient descent and optimization methods derived from mathematical programming are crucial in adjusting the weights of neural networks to improve accuracy. For instance, techniques like linear programming and mixed-integer programming can be employed to refine the search for model parameters, ensuring they align closely with the data. In this talk, ensembles of neural network architectures modeled by second-order conic programming will be discussed with applications in image recognition. As a neural network architecture, Convolutional Neural Networks (CNNs) are considered to train image data sets in the ensemble. The best models are selected based on their accuracy and diversities within the ensemble via second-order conic model. The experimental results show promising performance on test data in terms of prediction accuracy and statistical significance.