Generate pictures with higher quality and more variety.
Editor’s note: This article comes from WeChat public account “qubit” (ID: QbitAI) a > author: fish sheep. p>
Look up, look down, look left, look right. Looking at food from all angles, it really makes people look hungrier. p>
p>
And the fact that I do n’t know if it’s good or bad news: these foods never really existed. p>
Yes, this is a “pseudo gourmet book” generated by DeepMind’s latest LOGAN strong>. p>
This GAN defeated the “strongest” BigGAN in its debut, became a new state-of-the-art, and increased FID and IS by 32% strong> and 17% strong>. p>
What concept? In short, it is that LOGAN can generate higher quality and more diverse pseudo photos. strong> p>
On the left is BigGAN (FID / IS: 5.04 / 126.8), and on the right is LOGAN (FID / IS: 5.09 / 217). p>
With the same low FID, LOGAN is more reliable than BigGAN. p>
p>
△ BigGAN on the left and LOGAN on the right p>
Regardless of FID, under similar high IS conditions, although the foods produced are all the same and the heat explodes, obviously LOGAN’s posture level will be more abundant. p>
p>
And, DeepMind said: No need to introduce any architectural changes or other parameters. p>
Potential optimization h3>
The method adopted by DeepMind is to introduce a latent optimisation inspired by CSGAN. p>
p>
First, let the latent variable z be propagated forward through the generator and discriminator. p>
The gradient of the generator loss (red dotted arrow) is then used to calculate the improved z ‘. p>
In the second forward propagation, the optimized z ‘is used. Thereafter, the gradient of the latent optimization calculation discriminator is introduced. p>
Finally, use these gradients to update the model. p>
p>
The core of this method, In fact, it is to strengthen the interaction between the discriminator and the generator to improve the adversity. strong> p>
An important problem with gradient-based optimization in GAN is that the vector field generated by the discriminator and generator loss is not a gradient vector field. Therefore, there is no guarantee that the gradient descent will find a local optimal solution and be recyclable, which will slow down the convergence rate or cause mode collapse and mode jump phenomena. p>
Symplectic gradient adjustment algorithm (SGA) can find stable fixed points in ordinary games, and can improve the dynamics of gradient-based methods in confrontation. However, because the second derivative of all parameters needs to be calculated, SGA’sExpansion costs are high. p>
p>
The potential optimization can only use the second derivative for the latent variable z and the discriminator and generator parameters, respectively, to achieve the effect of approximate SGA. p>
This eliminates the need to use computationally expensive second-order terms involving discriminator and generator parameters. p>
In short, potential optimizations most effectively couple the gradients of discriminators and generators, and are more scalable. p>
And, LOGAN benefits from a powerful optimizer. When researchers used natural gradient descent (NGD) for potential optimization, they found that this approximate second-order optimization method performed better than the exact second-order method. p>
While NGD is also costly in high-dimensional parameter spaces, it is effective for potential optimizations even in very large models. p>
From the experimental results, the potential optimization significantly improves the training effect of GAN. p>
p>
Using the same architecture and number of parameters as the BigGAN-deep baseline, LOGAN performs better on FID and IS. p>
p>
However, during training, LOGAN is 2 to 3 times slower than BigGAN due to the extra forward and backward propagation. p>
Leading Chinese leader h3>
The first work is Yan Wu, a research scientist at DeepMind. p>
He won Cambridge University in 2019He has a PhD degree in computational neuroscience, and entered DeepMind in 16 years. p>
p>
The other authors of the paper are Dr. Jeff Donahue who graduated from UC Berkeley. p>
Dr. David Balduzzi, a graduate in mathematics from the University of Chicago. p>
Karen Simonyan, founder of Vision Factory. p>
And Timothy Lillicrap, a visiting professor at London College University and a PhD in system neuroscience from Queens University. p>
Portal: strong> p>
Paper address: https://arxiv.org/abs/1912.00953 p>
Related papers: p>
SGA: https://arxiv.org/abs/1802.05642CSGAN: https://arxiv.org/abs/1901.03554 p>
Cover image from pexels p>