Neural networks are widely believed to succeed because they learn useful representations from training data. But proving that this learned representation gives a computational edge over fixed, hand-crafted features has been tricky. A new paper on arXiv takes up this question and claims a separation that holds without distributional assumptions.
According to the abstract, known results that demonstrate a computational advantage of learned features over fixed ones are limited in some way (the abstract cuts off). The new work appears to close that gap by offering a distribution-free separation, meaning the result does not depend on specific probability distributions over the data.
While the abstract is short and the full technical details are not yet available in the summary, the contribution is significant for theoretical machine learning: it strengthens the case that representation learning is not just practically useful but also computationally necessary in a well-defined sense.