A Selective Overview of Deep Learning

Livro digital

Título:
A Selective Overview of Deep Learning

Autor:
Fan, Ma, Zhong

Categoria:
Tecnologia > IA

Doador:
Raffaello D. N.

Sinopse:
Deep learning can look like a black box, but this survey starts by unpacking the core question from its own table of contents: what makes deep learning different from classical methods, and why does depth matter at all? The paper opens with an introduction, then moves into feed-forward neural networks and back-propagation in computational graphs, building a statistical view of the field instead of a purely engineering one. From there, it walks through the main model families and training ideas that shaped modern practice, including convolutional neural networks, recurrent neural networks, modules, autoencoders, generative adversarial networks, stochastic gradient descent, regularization, and numerical stabilization. It also examines representation power through approximation theory, then shifts to the harder questions of training and generalization, with separate discussion of algorithm-independent and algorithm-dependent controls. The result is a compact but unusually broad overview of deep learning’s structure, strengths, and open theoretical problems. With three authors listed alphabetically, the paper reads as a collaborative academic survey that connects practical success in vision and language with the deeper statistical and theoretical questions still driving research.