Auto Byte



Science AI





  • <NIPS 2016 Tutorial:Generative Adversarial Networks>

1.Generative adversarial networks are an example of generative models. The term “generative model” is used in many different ways. The term refers to any model that takes a training set, consisting of samples drawn from a distribution p_{data} , and learns to represent an estimate of that distribution somehow. The result is a probability distribution p_{model} . In some cases, the model estimates p_{model} explicitly. In other cases, the model is only able to generate samples from p_{model} . Some models are able to do both. GANs focus primarily on sample generation, though it is possible to design GANs that can do both.

2.Why study generative modeling?

There are several reasons to study generative models, including:

  1. Training and sampling from generative models is an excellent test of our ability to represent and manipulate high-dimensional probability distributions.
  2. Generative models can be incorporated into reinforcement learning in several ways. Reinforcement learning algorithms can be divided into two categories; model-based and model-free. GAN is related to model-based methods. Generative models(such as GAN) can be used to simulate possible futures. So it can be used for planning and reinforcement learning. (1)A generative model used for planning can learn a conditional distribution over future states of the world, given the current state of the world and hypothetical actions an agent might take as input. (2) Another way that generative models might be used for reinforcement learning is to enable learning in an imaginary environment, where mistaken actions do not cause real damage to the agent.
  3. Generative models can be trained with missing data and can provide predictions on inputs that are missing data. The learning algorithm can improve its generalization by studying a large number of unlabeled examples which, which are usually easier to obtain. Generative models, and GANs in particular, are able to perform semi-supervised learning reasonably well. ( Generative models, and GANs in particular, enable machine learning to work with multi-modal outputs. For many tasks, a single input may correspond to many different correct answers, each of which is acceptable. )

Examples of some of these tasks that intrinsically require the generation of good samples include:

  • Single image super-resolution: In this task, the goal is to take a lowresolution image and synthesize a high-resolution equivalent.
  • Tasks where the goal is to create art: GANs, can be used to create interactive programs that assist the user in creating realistic images that correspond to rough scenes in the user’s imagination.
  • Image-to-image translation: Convert aerial photos into maps or convert sketches to images.


入门GANNIPS 2016



强化学习是一种试错方法,其目标是让软件智能体在特定环境中能够采取回报最大化的行为。强化学习在马尔可夫决策过程环境中主要使用的技术是动态规划(Dynamic Programming)。流行的强化学习方法包括自适应动态规划(ADP)、时间差分(TD)学习、状态-动作-回报-状态-动作(SARSA)算法、Q 学习、深度强化学习(DQN);其应用包括下棋类游戏、机器人控制和工作调度等。