A model is worth tens of thousands of examples for estimation and thousands for classification

Thomas Dagès, Laurent D. Cohen, Alfred M. Bruckstein

Abstract: Highlights•Neural networks systematically outperform traditional expert methods when trained on sufficiently large datasets, yet the necessary amount of training data has been largely overlooked.•We study three well-posed learning problems covering both estimation and classification: deconvolving one-dimensional Gaussian signals, estimating the position and radius of a disk in an image, and classifying a random dot image to possess a significant line and if so locate it.•We train various shallow custom and deep famous convolution architectures, either with transfer-learning, finetuning, or from scratch on various amounts of training data and compare them to expert (close to) optimal approaches.•We find that networks require tens of thousands of training samples for estimation problems and only thousands for classification to compete or outperform expert data-agnostic methods if they can.