Francois Lanusse @EiffL
Do you know this person?
Probably not, this is a randomly generated person: thispersondoesntexist.com
True
samples
Model
Distance between pairs of points drawn from a Gaussian distribution
Hint: Deep Learning is involved...
Fake images from a PixelCNN
Real SDSS images
They are data-driven models, can complement physical models
VAE model of galaxy morphology
Simulation of Dark Matter Maps
Observations
Model convolved with PSF
Model
Residuals
Observed data
Imagined solutions
Grathwohl et al. 2019
Problem: In the data, I only have access to the output , but how can I train if I never see the input ????
Why do we expect this to work? We are saying that the data can actually be represented on the low dimensionality manifold in latent space.
The encoder tries to guess the latent variable that generates the image
Encoder
Decoder
Auto-Encoded MNIST digits in 2d
Examples in Physics:
Reconstruction Error
Code Regularization
In this case, this is equivalent to the AE loss if
A distance between distributions: the Kullback-Leibler Divergence
The ELBO is maximal when the input x is close to the output, and the code is close to a Gaussian
Reconstruction Error
Code Regularization
import tensorflow as tf
import tensorflow_probability as tfp
tfd = tfp.distributions
# Build model.
model = tf.keras.Sequential([
tf.keras.layers.Dense(1+1),
tfp.layers.IndependentNormal(1),
])
# Define the loss function:
negloglik = lambda x, q: - q.log_prob(x)
# Do inference.
model.compile(optimizer='adam', loss=negloglik)
model.fit(x, y, epochs=500)
# Make predictions.
yhat = model(x_tst)
Traditional GAN (Goodfellow 2014)
BigGAN
VQ-VAE
Arjovsky et al. 2017
128x128 images, state of the art in 2017
WGAN-GP
1024x1024, state of the art circa end of 2019
This is extremely compute expensive and extremely technical