Training Neural Networks from Scratch with Parallel Low-Rank Adapters