Low-dimensional Functions are Efficiently Learnable under Randomly Biased Distributions