tfp.vi.mutual_information.lower_bound_jensen_shannon

Lower bound on Jensen-Shannon (JS) divergence.

This lower bound on JS divergence is proposed in [Goodfellow et al. (2014)][1] and [Nowozin et al. (2016)][2]. When estimating lower bounds on mutual information, one can also use different approaches for training the critic w.r.t. estimating mutual information [(Poole et al., 2018)][3]. The JS lower bound is used to train the critic with the standard lower bound on the Jensen-Shannon divergence as used in GANs, and then evaluates the critic using the NWJ lower bound on KL divergence, i.e. mutual information. As Eq.7 and Eq.8 of [Nowozin et al. (2016)][2], the bound is given by

I_JS = E_p(x,y)[log( D(x,y) )] + E_p(x)p(y)[log( 1 - D(x,y) )]

where the first term is the expectation over the samples from joint distribution (positive samples), and the second is for the samples from marginal distributions (negative samples), with

D(x, y) = sigmoid(f(x, y)),
log(D(x, y)) = softplus(-f(x, y)).

f(x, y) is a critic function that scores all pairs of samples.

Example:

X, Y are samples from a joint Gaussian distribution, with correlation 0.8 and both of dimension 1.

batch_size, rho, dim = 10000, 0.8, 1
y, eps = tf.split(
    value=tf.random.normal(shape=(2 * batch_size, dim), seed=7),
    num_or_size_splits=2, axis=0)
mean, conditional_stddev = rho * y, tf.sqrt(1. - tf.square(rho))
x = mean + conditional_stddev * eps

# Scores/unnormalized likelihood of pairs of samples `x[i], y[j]`
# (For JS lower bound, the optimal critic is of the form `f(x, y) = 1 +
# log(p(x | y) / p(x))` [(Poole et al., 2018)][3].)
conditional_dist = tfd.MultivariateNormalDiag(
    mean, scale_diag=conditional_stddev * tf.ones((batch_size, dim)))
conditional_scores = conditional_dist.log_prob(y[:, tf.newaxis, :])
marginal_dist = tfd.MultivariateNormalDiag(tf.zeros(dim), tf.ones(dim))
marginal_scores = marginal_dist.log_prob(y)[:, tf.newaxis]
scores = 1 + conditional_scores - marginal_scores

# Mask for joint samples in the score tensor
# (The `scores` has its shape [x_batch_size, y_batch_size], i.e.
# `scores[i, j] = f(x[i], y[j]) = log p(x[i] | y[j])`.)
joint_sample_mask = tf.eye(batch_size, dtype=bool)

# Lower bound on Jensen Shannon divergence
lower_bound_jensen_shannon(logu=scores, joint_sample_mask=joint_sample_mask)

logu float-like Tensor of size [batch_size_1, batch_size_2] representing critic scores (scores) for pairs of points (x, y) with logu[i, j] = f(x[i], y[j]).
joint_sample_mask bool-like Tensor of the same size as logu masking the positive samples by True, i.e. samples from joint distribution p(x, y). Default value: None. By default, an identity matrix is constructed as the mask.
validate_args Python bool, default False. Whether to validate input with asserts. If validate_args is False, and the inputs are invalid, correct behavior is not guaranteed.
name Python str name prefixed to Ops created by this function. Default value: None (i.e., 'lower_bound_jensen_shannon').

lower_bound float-like scalar for lower bound on JS divergence.

References:

[1]: Ian J. Goodfellow, et al. Generative Adversarial Nets. In Conference on Neural Information Processing Systems, 2014. https://arxiv.org/abs/1406.2661 [2]: Sebastian Nowozin, Botond Cseke, Ryota Tomioka. f-GAN: Training Generative Neural Samplers using Variational Divergence Minimization. In Conference on Neural Information Processing Systems, 2016. https://arxiv.org/abs/1606.00709 [3]: Ben Poole, Sherjil Ozair, Aaron van den Oord, Alexander A. Alemi, George Tucker. On Variational Bounds of Mutual Information. In International Conference on Machine Learning, 2019. https://arxiv.org/abs/1905.06922