【AI头条&优质资源】深度学习近期和未来有哪些突破？Yann LeCun的Quora解答

优秀论文精选』

• 《A Tour of TensorFlow》链接：https://arxiv.org/pdf/1610.01178.pdf 『AI头条』
• 《麦肯锡AI研究报告》链接: https://pan.baidu.com/s/1gePqW0r 密码: k4xe
• 《（麻省理工科技评论）本年度全球50大最聪明公司》链接：https://www.technologyreview.com/lists/companies/2017/intro/#nvidia

『深度学习tips』

《What are some recent and potentially upcoming breakthroughs in deep learning?》

• From Quora by Yann LeCun, Director of AI Research at Facebook and Professor at NYU

There are many interesting recent development in deep learning, probably too many for me to describe them all here. But there are a few ideas that caught my attention enough for me to get personally involved in research projects.

The most important one, in my opinion, is adversarial training (also called GAN for Generative Adversarial Networks). This is an idea that was originally proposed by Ian Goodfellow when he was a student with Yoshua Bengio at the University of Montreal (he since moved to Google Brain and recently to OpenAI).

This, and the variations that are now being proposed is the most interesting idea in the last 10 years in ML, in my opinion.

The idea is to simultaneously train two neural nets. The first one, called the Discriminator — let’s denote it D(Y) — takes an input (e.g. an image) and outputs a scalar that indicates whether the image Y looks “natural” or not. In one instance of adversarial training, D(Y) can be seem as some sort of energy function that takes a low value (e.g. close to 0) when Y is a real sample (e.g. an image from a database) and a positive value when it is not (e.g. if it’s a noisy or strange looking image). The second network is called the generator, denoted G(Z), where Z is generally a vector randomly sampled in a simple distribution (e.g. Gaussian). The role of the generator is to produce images so as to train the D(Y) function to take the right shape (low values for real images, higher values for everything else). During training D is shown a real image, and adjusts its parameter to make its output lower. Then D is shown an image produced from G and adjusts its parameters to make its output D(G(Z)) larger (following the gradient of some objective predefined function). But G(Z) will train itself to produce images so as to fool D into thinking they are real. It does this by getting the gradient of D with respect to Y for each sample it produces. In other words, it’s trying to minimize the output of D while D is trying to maximize it. Hence the name adversarial training.

The original formulation uses a considerably more complicated probabilistic framework, but that’s the gist of it.

Why is that so interesting? It allows us to train a discriminator as an unsupervised “density estimator”, i.e. a contrast function that gives us a low value for data and higher output for everything else. This discriminator has to develop a good internal representation of the data to solve this problem properly. It can then be used as a feature extractor for a classifier, for example.

But perhaps more interestingly, the generator can be seen as parameterizing the complicated surface of real data: give it a vector Z, and it maps it to a point on the data manifold. There are papers where people do amazing things with this, like generating pictures of bedrooms, doing arithmetic on faces in the Z vector space: [man with glasses] - [man without glasses] + [woman without glasses] = [woman with glasses].

There has been a series of interesting papers from FAIR on the topic:

• Denton et al. “Deep Generative Image Models using a Laplacian Pyramid of Adversarial Networks” (NIPS 2015) : https://scholar.google.com/citat...
• Radford et al. “Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks” (ICLR 2015): https://scholar.google.com/citat...
• Mathieu et al. “Deep multi-scale video prediction beyond mean square error” : https://scholar.google.com/citat...

This last one is on video prediction with adversarial training. It solves a really important issue, which is that when you train a neural net (or any other model) to predict the future, and when there are several possible futures, a network trained the traditional way (e.g. with least square) will predict the average of all the possible futures. In the case of video, it will produce a blurry mess. Adversarial training lets the system produce whatever it wants, as long as it’s within the set that the discriminator likes. This solves the “blurriness” problem when predicting under. uncertainty.

It seems like a rather technical issue, but I really think it opens the door to an entire world of possibilities.