Replace the sigmoid function with the softplus function, which has a smoother derivative that avoids zero points and reduces vanishing gradients during backpropagation. This change comes at the cost of slightly increased complexity in the model’s output layer.