Samir Brahim Belhaouari

Abstract. This paper explores advanced methods for pruning deep neural networks, leveraging curve fitting and the evolution of weights to enhance model efficiency without significantly sacrificing accuracy. The proposed methodology employs polynomial curve fitting to approximate the weight distribution, followed by an evolutionary algorithm to iteratively refine the pruned model. Formally, given a weight matrix \(W\), we approximate \(W \approx \sum_{i=0}^{n} a_i x^i\), where \(a_i\) are the polynomial coefficients. The evolutionary pruning process iteratively adjusts \(W\) to minimize the loss function \(L(W)\), subject to a sparsity constraint \(\|W\|_0 \leq k\). Experimental results demonstrate that this approach achieves a substantial reduction in model size while maintaining performance metrics comparable to the unpruned network.
    

Illustration of the proposed experimental design.