Archive for neural network

sequential neural likelihood estimation as ABC substitute

Posted in Books, Kids, Statistics, University life with tags , , , , , , , , , , , , , , , , , , on May 14, 2020 by xi'an

A JMLR paper by Papamakarios, Sterratt, and Murray (Edinburgh), first presented at the AISTATS 2019 meeting, on a new form of likelihood-free inference, away from non-zero tolerance and from the distance-based versions of ABC, following earlier papers by Iain Murray and co-authors in the same spirit. Which I got pointed to during the ABC workshop in Vancouver. At the time I had no idea as to autoregressive flows meant. We were supposed to hold a reading group in Paris-Dauphine on this paper last week, unfortunately cancelled as a coronaviral precaution… Here are some notes I had prepared for the meeting that did not take place.

A simulator model is a computer program, which takes a vector of parameters θ, makes internal calls to a random number generator, and outputs a data vector x.”

Just the usual generative model then.

“A conditional neural density estimator is a parametric model q(.|φ) (such as a neural network) controlled by a set of parameters φ, which takes a pair of datapoints (u,v) and outputs a conditional probability density q(u|v,φ).”

Less usual, in that the outcome is guaranteed to be a probability density.

“For its neural density estimator, SNPE uses a Mixture Density Network, which is a feed-forward neural network that takes x as input and outputs the parameters of a Gaussian mixture over θ.”

In which theoretical sense would it improve upon classical or Bayesian density estimators? Where are the error evaluation, the optimal rates, the sensitivity to the dimension of the data? of the parameter?

“Our new method, Sequential Neural Likelihood (SNL), avoids the bias introduced by the proposal, by opting to learn a model of the likelihood instead of the posterior.”

I do not get the argument in that the final outcome (of using the approximation within an MCMC scheme) remains biased since the likelihood is not the exact likelihood. Where is the error evaluation? Note that in the associated Algorithm 1, the learning set is enlarged on each round, as in AMIS, rather than set back to the empty set ∅ on each round.

…given enough simulations, a sufficiently flexible conditional neural density estimator will eventually approximate the likelihood in the support of the proposal, regardless of the shape of the proposal. In other words, as long as we do not exclude parts of the parameter space, the way we propose parameters does not bias learning the likelihood asymptotically. Unlike when learning the posterior, no adjustment is necessary to account for our proposing strategy.”

This is a rather vague statement, with the only support being that the Monte Carlo approximation to the Kullback-Leibler divergence does converge to its actual value, i.e. a direct application of the Law of Large Numbers! But an interesting point I informally made a (long) while ago that all that matters is the estimate of the density at x⁰. Or at the value of the statistic at x⁰. The masked auto-encoder density estimator is based on a sequence of bijections with a lower-triangular Jacobian matrix, meaning the conditional density estimate is available in closed form. Which makes it sounds like a form of neurotic variational Bayes solution.

The paper also links with ABC (too costly?), other parametric approximations to the posterior (like Gaussian copulas and variational likelihood-free inference), synthetic likelihood, Gaussian processes, noise contrastive estimation… With experiments involving some of the above. But the experiments involve rather smooth models with relatively few parameters.

“A general question is whether it is preferable to learn the posterior or the likelihood (…) Learning the likelihood can often be easier than learning the posterior, and it does not depend on the choice of proposal, which makes learning easier and more robust (…) On the other hand, methods such as SNPE return a parametric model of the posterior directly, whereas a further inference step (e.g. variational inference or MCMC) is needed on top of SNL to obtain a posterior estimate”

A fair point in the conclusion. Which also mentions the curse of dimensionality (both for parameters and observations) and the possibility to work directly with summaries.

Getting back to the earlier and connected Masked autoregressive flow for density estimation paper, by Papamakarios, Pavlakou and Murray:

“Viewing an autoregressive model as a normalizing flow opens the possibility of increasing its flexibility by stacking multiple models of the same type, by having each model provide the source of randomness for the next model in the stack. The resulting stack of models is a normalizing flow that is more flexible than the original model, and that remains tractable.”

Which makes it sound like a sort of a neural network in the density space. Optimised by Kullback-Leibler minimisation to get asymptotically close to the likelihood. But a form of Bayesian indirect inference in the end, namely an MLE on a pseudo-model, using the estimated model as a proxy in Bayesian inference…

Nature tidbits [the Bayesian brain]

Posted in Statistics with tags , , , , , , , , , , , , , , on March 8, 2020 by xi'an

In the latest Nature issue, a long cover of Asimov’s contributions to science and rationality. And a five page article on the dopamine reward in the brain seen as a probability distribution, seen as distributional reinforcement learning by researchers from DeepMind, UCL, and Harvard. Going as far as “testing” for this theory with a p-value of 0.008..! Which could be as well a signal of variability between neurons to dopamine rewards (with a p-value of 10⁻¹⁴, whatever that means). Another article about deep learning about protein (3D) structure prediction. And another one about learning neural networks via specially designed devices called memristors. And yet another one on West Africa population genetics based on four individuals from the Stone to Metal age (8000 and 3000 years ago), SNPs, PCA, and admixtures. With no ABC mentioned (I no longer have access to the journal, having missed renewal time for my subscription!). And the literal plague of a locust invasion in Eastern Africa. Making me wonder anew as to why proteins could not be recovered from the swarms of locust to partly compensate for the damages. (Locusts eat their bodyweight in food every day.) And the latest news from NeurIPS about diversity and inclusion. And ethics, as in checking for responsibility and societal consequences of research papers. Reviewing the maths of a submitted paper or the reproducibility of an experiment is already challenging at times, but evaluating the biases in massive proprietary datasets or the long-term societal impact of a classification algorithm may prove beyond the realistic.

Julyan’s talk on priors in Bayesian neural networks [cancelled!]

Posted in pictures, Statistics, Travel, University life with tags , , , , , , , , , , , , , on March 5, 2020 by xi'an

Next Friday, 13 March at 1:30p.m., Julyan Arbel, researcher at Inria Grenoble will give a All about that Bayes talk at CMLA, ENS Paris-Saclay (building D’Alembert, room Condorcet, Cachan, RER stop Bagneux) on

Understanding Priors in Bayesian Neural Networks at the Unit Level

We investigate deep Bayesian neural networks with Gaussian weight priors and a class of ReLU-like nonlinearities. Bayesian neural networks with Gaussian priors are well known to induce an L², “weight decay”, regularization. Our results characterize a more intricate regularization effect at the level of the unit activations. Our main result establishes that the induced prior distribution on the units before and after activation becomes increasingly heavy-tailed with the depth of the layer. We show that first layer units are Gaussian, second layer units are sub-exponential, and units in deeper layers are characterized by sub-Weibull distributions. Our results provide new theoretical insight on deep Bayesian neural networks, which we corroborate with simulation experiments.

 

séminaire P de S

Posted in Books, pictures, Statistics, University life with tags , , , , , , , , , on February 18, 2020 by xi'an

As I was in Paris and free for the occasion (!), I attended the Paris Statistics seminar this afternoon, in the Latin Quarter. With a first talk by Kweku Abraham on Bayesian inverse problems set a prior on the quantity of interest, γ, rather than its transform G(γ), observed with noise. Always perturbed by the juggling of different distances, like L² versus Kullback-Leibler, in non-parametric frameworks. Reminding me of probabilistic numerics, at least in the framework, since the crux of the talk was 100% about convergence. And a second talk by Leanaïc Chizat on convex neural networks corresponding to an infinite number of neurons, with surprising properties, including implicit bias. And a third talk by Anne Sabourin on PCA for extremes. Which assumed very little on the model but more on the geometry of the distribution, like extremes being concentrated on a subspace. As I was rather tired from an intense week at Warwick, and after a weekend of reading grant applications and Biometrika submissions (!), my foggy brain kept switching to these proposals, trying to make connections with the talks, not completely inappropriately in two cases out of three. (I am afraid the same may happen tomorrow at our probability seminar on computer-based proofs!)

double descent

Posted in Books, Statistics, University life with tags , , , , , , , , , , , on November 7, 2019 by xi'an

Last Friday, I [and a few hundred others!] went to the SMILE (Statistical Machine Learning in Paris) seminar where Francis Bach was giving a talk. (With a pleasant ride from Dauphine along the Seine river.) Fancis was talking about the double descent phenomenon observed in recent papers by Belkin & al. (2018, 2019), and Mei & Montanari (2019). (As the seminar room at INRIA was quite crowded and as I was sitting X-legged on the floor close to the screen, I took a few slides from below!) The phenomenon is that the usual U curve warning about over-fitting and reproduced in most statistics and machine-learning courses can under the right circumstances be followed by a second decrease in the testing error when the number of features goes beyond the number of observations. This is rather puzzling and counter-intuitive, so I briefkly checked the 2019 [8 pages] article by Belkin & al., who are studying two examples, including a standard “large p small n” Gaussian regression. where the authors state that

“However, as p grows beyond n, the test risk again decreases, provided that the model is fit using a suitable inductive bias (e.g., least norm solution). “

One explanation [I found after checking the paper] is that the variates (features) in the regression are selected at random rather than in an optimal sequential order. Double descent is missing with interpolating and deterministic estimators. Hence requiring on principle all candidate variates to be included to achieve minimal averaged error. The infinite spike is when the number p of variate is near the number n of observations. (The expectation accounts as well for the randomisation in T. Randomisation that remains an unclear feature in this framework…)

likelihood-free approximate Gibbs sampling

Posted in Books, Statistics with tags , , , , , , , , on June 19, 2019 by xi'an

“Low-dimensional regression-based models are constructed for each of these conditional distributions using synthetic (simulated) parameter value and summary statistic pairs, which then permit approximate Gibbs update steps (…) synthetic datasets are not generated during each sampler iteration, thereby providing efficiencies for expensive simulator models, and only require sufficient synthetic datasets to adequately construct the full conditional models (…) Construction of the approximate conditional distributions can exploit known structures of the high-dimensional posterior, where available, to considerably reduce computational overheads”

Guilherme Souza Rodrigues, David Nott, and Scott Sisson have just arXived a paper on approximate Gibbs sampling. Since this comes a few days after we posted our own version, here are some of the differences I could spot in the paper:

  1. Further references to earlier occurrences of Gibbs versions of ABC, esp. in cases when the likelihood function factorises into components and allows for summaries with lower dimensions. And even to ESP.
  2. More an ABC version of Gibbs sampling that a Gibbs version of ABC in that approximations to the conditionals are first constructed and then used with no further corrections.
  3. Inherently related to regression post-processing à la Beaumont et al.  (2002) in that the regression model is the start to designing an approximate full conditional, conditional on the “other” parameters and on the overall summary statistic. The construction of the approximation is far from automated. And may involve neural networks or other machine learning estimates.
  4. As a consequence of the above, a preliminary ABC step to design the collection of approximate full conditionals using a single and all-purpose multidimensional summary statistic.
  5. Once the approximations constructed, no further pseudo-data is generated.
  6. Drawing from the approximate full conditionals is done exactly, possibly via a bootstrapped version.
  7. Handling a highly complex g-and-k dynamic model with 13,140 unknown parameters, requiring a ten days simulation.

“In certain circumstances it can be seen that the likelihood-free approximate Gibbs sampler will exactly target the true partial posterior (…) In this case, then Algorithms 2 and 3 will be exact.”

Convergence and coherence are handled in the paper by setting the algorithm(s) as noisy Monte Carlo versions, à la Alquier et al., although the issue of incompatibility between the full conditionals is acknowledged, with the main reference being the finite state space analysis of Chen and Ip (2015). It thus remains unclear whether or not the Gibbs samplers that are implemented there do converge and if they do what is the significance of the stationary distribution.

selecting summary statistics [a tale of two distances]

Posted in Books, Statistics with tags , , , , , , , , , , , , , , on May 23, 2019 by xi'an

As Jonathan Harrison came to give a seminar in Warwick [which I could not attend], it made me aware of his paper with Ruth Baker on the selection of summaries in ABC. The setting is an ABC-SMC algorithm and it relates with Fearnhead and Prangle (2012), Barnes et al. (2012), our own random forest approach, the neural network version of Papamakarios and Murray (2016), and others. The notion here is to seek the optimal weights of different summary statistics in the tolerance distance, towards a maximization of a distance (Hellinger) between prior and ABC posterior (Wasserstein also comes to mind!). A sort of dual of the least informative prior. Estimated by a k-nearest neighbour version [based on samples from the prior and from the ABC posterior] I had never seen before. I first did not get how this k-nearest neighbour distance could be optimised in the weights since the posterior sample was already generated and (SMC) weighted, but the ABC sample can be modified by changing the [tolerance] distance weights and the resulting Hellinger distance optimised this way. (There are two distances involved, in case the above description is too murky!)

“We successfully obtain an informative unbiased posterior.”

The paper spends a significant while in demonstrating that the k-nearest neighbour estimator converges and much less on the optimisation procedure itself, which seems like a real challenge to me when facing a large number of particles and a high enough dimension (in the number of statistics). (In the examples, the size of the summary is 1 (where does the weight matter?), 32, 96, 64, with 5 10⁴, 5 10⁴, 5 10³ and…10 particles, respectively.) The authors address the issue, though, albeit briefly, by mentioning that, for the same overall computation time, the adaptive weight ABC is indeed further from the prior than a regular ABC with uniform weights [rather than weighted by the precisions]. They also argue that down-weighting some components is akin to selecting a subset of summaries, but I beg to disagree with this statement as the weights are never exactly zero, as far as I can see, hence failing to fight the curse of dimensionality. Some LASSO version could implement this feature.