For joint inference over multiple variables, a variety of structured prediction techniques have been developed to model correlations among variables and thereby improve predictions. However, many classical approaches suffer from one of two primary drawbacks: they either lack the ability to model high-order correlations among variables while maintaining computationally tractable inference, or they do not allow to explicitly model known correlations. To address this shortcoming, we introduce `Graph Structured Prediction Energy Networks,' for which we develop inference techniques that allow to both model explicit local and implicit higher-order correlations while maintaining tractability of inference. We apply the proposed method to tasks from the natural language processing and computer vision domain and demonstrate its general utility.

Despite the exploding interest in graph neural networks there has been little effort to verify and improve their robustness. This is even more alarming given recent findings showing that they are extremely vulnerable to adversarial attacks on both the graph structure and the node attributes. We propose the first method for verifying certifiable (non-)robustness to graph perturbations for a general class of models that includes graph neural networks and label/feature propagation. By exploiting connections to PageRank and Markov decision processes our certificates can be efficiently (and under many threat models exactly) computed. Furthermore, we investigate robust training procedures that increase the number of certifiably robust nodes while maintaining or improving the clean predictive accuracy.

Energy-based models (EBMs) are powerful probabilistic models, but suffer from intractable sampling and density evaluation due to the partition function. As a result, inference in EBMs relies on approximate sampling algorithms, leading to a mismatch between the model and inference. Motivated by this, we consider the sampler-induced distribution as the model of interest and maximize the likelihood of this model. This yields a class of energy-inspired models (EIMs) that incorporate learned energy functions while still providing exact samples and tractable log-likelihood lower bounds. We describe and evaluate three instantiations of such models based on truncated rejection sampling, self-normalized importance sampling, and Hamiltonian importance sampling. These models outperform or perform comparably to the recently proposed Learned Accept/Reject Sampling algorithm and provide new insights on ranking Noise Contrastive Estimation and Contrastive Predictive Coding. Moreover, EIMs allow us to generalize a recent connection between multi-sample variational lower bounds and auxiliary variable variational inference. We show how recent variational bounds can be unified with EIMs as the variational family.

Black Box Machine Learning models leak information about the proprietary model parameters and architecture, both through side channels and output predictions. An adversary can thus, exploit this leakage to reconstruct a substitute architecture similar to the target model, violating the model privacy and Intellectual Property. However, all such attacks, infer a subset of the target model attributes and identifying the rest of the architecture and parameters (optimally) is a search problem. Extracting the exact target model is not possible owing to the uncertainty in the inference attack outputs and stochastic nature of the training process. In this work, we propose a probabilistic framework, Airavata, to estimate the leakage in such model extraction attacks. Specifically, we use Bayesian Networks to capture the uncertainty, under the subjective notion of probability, in estimating the target model attributes using various model extraction attacks. We experimentally validate the model under different adversary assumptions commonly adopted by various model extraction attacks to reason about the attack efficacy. Further, this provides a practical approach of inferring actionable knowledge about extracting black box models and identify the best combination of attacks which maximise the knowledge extracted (information leaked) from the target model.

We consider a special type of Restricted Boltzmann machine (RBM), namely a Gaussian-spherical RBM where the visible units have Gaussian priors while the vector of hidden variables is constrained to stay on an ${\mathbbm L}_2$ sphere. The spherical constraint having the advantage to admit exact asymptotic treatments, various scaling regimes are explicitly identified based solely on the spectral properties of the coupling matrix (also called weight matrix of the RBM). Incidentally these happen to be formally related to similar scaling behaviours obtained in a different context dealing with spatial condensation of zero range processes. More specifically, when the spectrum of the coupling matrix is doubly degenerated an exact treatment can be proposed to deal with finite size effects. Interestingly the known parallel between the ferromagnetic transition of the spherical model and the Bose-Einstein condensation can be made explicit in that case. More importantly this gives us the ability to extract all needed response functions with arbitrary precision for the training algorithm of the RBM. This allows us then to numerically integrate the dynamics of the spectrum of the weight matrix during learning in a precise way. This dynamics reveals in particular a sequential emergence of modes from the Marchenko-Pastur bulk of singular vectors of the coupling matrix.

Regularization methods, specifically those which directly alter weights like $L_1$ and $L_2$, are an integral part of many learning algorithms. Both the regularizers mentioned above are formulated by assuming certain priors in the parameter space and these assumptions, in some cases, induce sparsity in the parameter space. Regularizers help in transferring beliefs one has on the dataset or the parameter space by introducing adequate terms in the loss function. Any kind of formulation represents a specific set of beliefs: $L_1$ regularization conveys that the parameter space should be sparse whereas $L_2$ regularization conveys that the parameter space should be bounded and continuous. These regularizers in turn leverage certain priors to express these inherent beliefs. A better understanding of how the prior affects the behavior of the parameters and how the priors can be updated based on the dataset can contribute greatly in improving the generalization capabilities of a function estimator. In this work, we introduce a weakly informative prior and then further extend it to an informative prior in order to formulate a regularization penalty, which shows better results in terms of inducing sparsity experimentally, when compared to regularizers based only on Gaussian and Laplacian priors. Experimentally, we verify that a regularizer based on an adapted prior improves the generalization capabilities of any network. We illustrate the performance of the proposed method on the MNIST and CIFAR-10 datasets.

Bayesian inference in state-space models is challenging due to high-dimensional state trajectories. A viable approach is particle Markov chain Monte Carlo, combining MCMC and sequential Monte Carlo to form "exact approximations" to otherwise intractable MCMC methods. The performance of the approximation is limited to that of the exact method. We focus on particle Gibbs and particle Gibbs with ancestor sampling, improving their performance beyond that of the underlying Gibbs sampler (which they approximate) by marginalizing out one or more parameters. This is possible when the parameter prior is conjugate to the complete data likelihood. Marginalization yields a non-Markovian model for inference, but we show that, in contrast to the general case, this method still scales linearly in time. While marginalization can be cumbersome to implement, recent advances in probabilistic programming have enabled its automation. We demonstrate how the marginalized methods are viable as efficient inference backends in probabilistic programming, and demonstrate with examples in ecology and epidemiology.

Markov networks are widely studied and used throughout multivariate statistics and computer science. In particular, the problem of learning the structure of Markov networks from data without invoking chordality assumptions in order to retain expressiveness of the model class has been given a considerable attention in the recent literature, where numerous constraint-based or score-based methods have been introduced. Here we develop a new search algorithm for the network score-optimization that has several computational advantages and scales well to high-dimensional data sets. The key observation behind the algorithm is that the neighborhood of a variable can be efficiently captured using local penalized likelihood ratio (PLR) tests by exploiting an exponential decay of correlations across the neighborhood with an increasing graph-theoretic distance from the focus node. The candidate neighborhoods are then processed by a two-stage hill-climbing (HC) algorithm. Our approach, termed fully as PLRHC-BIC$_{0.5}$, compares favorably against the state-of-the-art methods in all our experiments spanning both low- and high-dimensional networks and a wide range of sample sizes. An efficient implementation of PLRHC-BIC$_{0.5}$ is freely available from the URL: https://github.com/jurikuronen/plrhc.

This work introduces two strategies for training network classifiers with heterogeneous agents. One strategy promotes global smoothing over the graph and a second strategy promotes local smoothing over neighbourhoods. It is assumed that the feature sizes can vary from one agent to another, with some agents observing insufficient attributes to be able to make reliable decisions on their own. As a result, cooperation with neighbours is necessary. However, due to the fact that the feature dimensions are different across the agents, their classifier dimensions will also be different. This means that cooperation cannot rely on combining the classifier parameters. We instead propose smoothing the outputs of the classifiers, which are the predicted labels. By doing so, the dynamics that describes the evolution of the network classifier becomes more challenging than usual because the classifier parameters end up appearing as part of the regularization term as well. We illustrate performance by means of computer simulations.

Causal inference can be formalized as Bayesian inference that combines a prior distribution over causal models and likelihoods that account for both observations and interventions. We show that it is possible to implement this approach using a sufficiently expressive probabilistic programming language. Priors are represented using probabilistic programs that generate source code in a domain specific language. Interventions are represented using probabilistic programs that edit this source code to modify the original generative process. This approach makes it straightforward to incorporate data from atomic interventions, as well as shift interventions, variance-scaling interventions, and other interventions that modify causal structure. This approach also enables the use of general-purpose inference machinery for probabilistic programs to infer probable causal structures and parameters from data. This abstract describes a prototype of this approach in the Gen probabilistic programming language.