The article discusses a new method for weight optimization in fully-connected feed-forward neural networks, which uses least squares (LS) methodology instead of the commonly used back-propagation (BP). This method allows for the optimization of weights in a single iteration, making it faster and more efficient than existing methods. Furthermore, it is adaptable and works well even when the input-to-output mapping is not injective. The proposed method is also ideal for parallel implementation as computations for each neuron in a layer are independent from each other.
Publication date: 15 Jan 2024
Project Page: https://arxiv.org/abs/2401.06699
Paper: https://arxiv.org/pdf/2401.06699
