We introduce and study the neighbourhood lattice decomposition of a distribution, which is a compact, non-graphical representation of conditional independence that is valid in the absence of a faithful graphical representation. The idea is to view the set of neighbourhoods of a variable as a subset lattice, and partition this lattice into convex sublattices, each of which directly encodes a collection of conditional independence relations. We show that this decomposition exists in any compositional graphoid and can be computed efficiently and consistently in high-dimensions. {In particular, this gives a way to encode all of independence relations implied by a distribution that satisfies the composition axiom, which is strictly weaker than the faithfulness assumption that is typically assumed by graphical approaches.} We also discuss various special cases such as graphical models and projection lattices, each of which has intuitive interpretations. Along the way, we see how this problem is closely related to neighbourhood regression, which has been extensively studied in the context of graphical models and structural equations.
Providing expert trajectories in the context of Imitation Learning is often expensive and time-consuming. The goal must therefore be to create algorithms which require as little expert data as possible. In this paper we present an algorithm that imitates the higher-level strategy of the expert rather than just imitating the expert on action level, which we hypothesize requires less expert data and makes training more stable. As a prior, we assume that the higher-level strategy is to reach an unknown target state area, which we hypothesize is a valid prior for many domains in Reinforcement Learning. The target state area is unknown, but since the expert has demonstrated how to reach it, the agent tries to reach states similar to the expert. Building on the idea of Temporal Coherence, our algorithm trains a neural network to predict whether two states are similar, in the sense that they may occur close in time. During inference, the agent compares its current state with expert states from a Case Base for similarity. The results show that our approach can still learn a near-optimal policy in settings with very little expert data, where algorithms that try to imitate the expert at the action level can no longer do so.
Equality of opportunity has emerged as an important ideal of distributive justice. Empirically, Inequality of Opportunity (IOp) is measured in two steps: first, an outcome (e.g., income) is predicted given individual circumstances; and second, an inequality index (e.g., Gini) of the predictions is computed. Machine Learning (ML) methods are tremendously useful in the first step. However, they can cause sizable biases in IOp since the bias-variance trade-off allows the bias to creep in the second step. We propose a simple debiased IOp estimator robust to such ML biases and provide the first valid inferential theory for IOp. We demonstrate improved performance in simulations and report the first unbiased measures of income IOp in Europe. Mother's education and father's occupation are the circumstances that explain the most. Plug-in estimators are very sensitive to the ML algorithm, while debiased IOp estimators are robust. These results are extended to a general U-statistics setting.
This paper presents a novel, fast and privacy preserving implementation of deep autoencoders. DAEF (Deep Autoencoder for Federated learning), unlike traditional neural networks, trains a deep autoencoder network in a non-iterative way, which drastically reduces its training time. Its training can be carried out in a distributed way (several partitions of the dataset in parallel) and incrementally (aggregation of partial models), and due to its mathematical formulation, the data that is exchanged does not endanger the privacy of the users. This makes DAEF a valid method for edge computing and federated learning scenarios. The method has been evaluated and compared to traditional (iterative) deep autoencoders using seven real anomaly detection datasets, and their performance have been shown to be similar despite DAEF's faster training.
Output reachability and adversarial robustness are among the most relevant safety properties of neural networks. We show that in the context of Message Passing Neural Networks (MPNN), a common Graph Neural Network (GNN) model, formal verification is impossible. In particular, we show that output reachability of graph-classifier MPNN, working over graphs of unbounded size, non-trivial degree and sufficiently expressive node labels, cannot be verified formally: there is no algorithm that answers correctly (with yes or no), given an MPNN, whether there exists some valid input to the MPNN such that the corresponding output satisfies a given specification. However, we also show that output reachability and adversarial robustness of node-classifier MPNN can be verified formally when a limit on the degree of input graphs is given a priori. We discuss the implications of these results, for the purpose of obtaining a complete picture of the principle possibility to formally verify GNN, depending on the expressiveness of the involved GNN models and input-output specifications.
The Slutsky equation, central in consumer choice theory, is derived from the usual hypotheses underlying most standard models in Economics, such as full rationality, homogeneity, and absence of interactions. We present a statistical physics framework that allows us to relax such assumptions. We first derive a general fluctuation-response formula that relates the Slutsky matrix to spontaneous fluctuations of consumption rather than to response to changing prices and budget. We then show that, within our hypotheses, the symmetry of the Slutsky matrix remains valid even when agents are only boundedly rational but non-interacting. We then propose a model where agents are influenced by the choice of others, leading to a phase transition beyond which consumption is dominated by herding (or `"fashion") effects. In this case, the individual Slutsky matrix is no longer symmetric, even for fully rational agents. The vicinity of the transition features a peak in asymmetry.
Kernel Stein discrepancy (KSD) is a widely used kernel-based measure of discrepancy between probability measures. It is often employed in the scenario where a user has a collection of samples from a candidate probability measure and wishes to compare them against a specified target probability measure. KSD has been employed in a range of settings including goodness-of-fit testing, parametric inference, MCMC output assessment and generative modelling. However, so far the method has been restricted to finite-dimensional data. We provide the first analysis of KSD in the generality of data lying in a separable Hilbert space, for example functional data. The main result is a novel Fourier representation of KSD obtained by combining the theory of measure equations with kernel methods. This allows us to prove that KSD can separate measures and thus is valid to use in practice. Additionally, our results improve the interpretability of KSD by decoupling the effect of the kernel and Stein operator. We demonstrate the efficacy of the proposed methodology by performing goodness-of-fit tests for various Gaussian and non-Gaussian functional models in a number of synthetic data experiments.
One of the fundamental challenges in the deep learning community is to theoretically understand how well a deep neural network generalizes to unseen data. However, current approaches often yield generalization bounds that are either too loose to be informative of the true generalization error or only valid to the compressed nets. In this study, we present a simple yet non-vacuous generalization bound from the optimization perspective. We achieve this goal by leveraging that the hypothesis set accessed by stochastic gradient algorithms is essentially fractal-like and thus can derive a tighter bound over the algorithm-dependent Rademacher complexity. The main argument rests on modeling the discrete-time recursion process via a continuous-time stochastic differential equation driven by fractional Brownian motion. Numerical studies demonstrate that our approach is able to yield plausible generalization guarantees for modern neural networks such as ResNet and Vision Transformer, even when they are trained on a large-scale dataset (e.g. ImageNet-1K).
Kernel methods are learning algorithms that enjoy solid theoretical foundations while suffering from important computational limitations. Sketching, which consists in looking for solutions among a subspace of reduced dimension, is a well studied approach to alleviate these computational burdens. However, statistically-accurate sketches, such as the Gaussian one, usually contain few null entries, such that their application to kernel methods and their non-sparse Gram matrices remains slow in practice. In this paper, we show that sparsified Gaussian (and Rademacher) sketches still produce theoretically-valid approximations while allowing for important time and space savings thanks to an efficient \emph{decomposition trick}. To support our method, we derive excess risk bounds for both single and multiple output kernel problems, with generic Lipschitz losses, hereby providing new guarantees for a wide range of applications, from robust regression to multiple quantile regression. Our theoretical results are complemented with experiments showing the empirical superiority of our approach over SOTA sketching methods.
We provide quantitative bounds measuring the $L^2$ difference in function space between the trajectory of a finite-width network trained on finitely many samples from the idealized kernel dynamics of infinite width and infinite data. An implication of the bounds is that the network is biased to learn the top eigenfunctions of the Neural Tangent Kernel not just on the training set but over the entire input space. This bias depends on the model architecture and input distribution alone and thus does not depend on the target function which does not need to be in the RKHS of the kernel. The result is valid for deep architectures with fully connected, convolutional, and residual layers. Furthermore the width does not need to grow polynomially with the number of samples in order to obtain high probability bounds up to a stopping time. The proof exploits the low-effective-rank property of the Fisher Information Matrix at initialization, which implies a low effective dimension of the model (far smaller than the number of parameters). We conclude that local capacity control from the low effective rank of the Fisher Information Matrix is still underexplored theoretically.
The key of Human-Object Interaction(HOI) recognition is to infer the relationship between human and objects. Recently, the image's Human-Object Interaction(HOI) detection has made significant progress. However, there is still room for improvement in video HOI detection performance. Existing one-stage methods use well-designed end-to-end networks to detect a video segment and directly predict an interaction. It makes the model learning and further optimization of the network more complex. This paper introduces the Spatial Parsing and Dynamic Temporal Pooling (SPDTP) network, which takes the entire video as a spatio-temporal graph with human and object nodes as input. Unlike existing methods, our proposed network predicts the difference between interactive and non-interactive pairs through explicit spatial parsing, and then performs interaction recognition. Moreover, we propose a learnable and differentiable Dynamic Temporal Module(DTM) to emphasize the keyframes of the video and suppress the redundant frame. Furthermore, the experimental results show that SPDTP can pay more attention to active human-object pairs and valid keyframes. Overall, we achieve state-of-the-art performance on CAD-120 dataset and Something-Else dataset.
Exact inference of the most probable explanation (MPE) in Bayesian networks is known to be NP-complete. In this paper, we propose an algorithm for approximate MPE inference that is based on the incremental build-infer-approximate (IBIA) framework. We use this framework to obtain an ordered set of partitions of the Bayesian network and the corresponding max-calibrated clique trees. We show that the maximum belief in the last partition gives an estimate of the probability of the MPE assignment. We propose an iterative algorithm for decoding, in which the subset of variables for which an assignment is obtained is guaranteed to increase in every iteration. There are no issues of convergence, and we do not perform a search for solutions. Even though it is a single shot algorithm, we obtain valid assignments in 100 out of the 117 benchmarks used for testing. The accuracy of our solution is comparable to a branch and bound search in majority of the benchmarks, with competitive run times.
Instrumental variable (IV) is a powerful approach to inferring the causal effect of a treatment on an outcome of interest from observational data even when there exist latent confounders between the treatment and the outcome. However, existing IV methods require that an IV is selected and justified with domain knowledge. An invalid IV may lead to biased estimates. Hence, discovering a valid IV is critical to the applications of IV methods. In this paper, we study and design a data-driven algorithm to discover valid IVs from data under mild assumptions. We develop the theory based on partial ancestral graphs (PAGs) to support the search for a set of candidate Ancestral IVs (AIVs), and for each possible AIV, the identification of its conditioning set. Based on the theory, we propose a data-driven algorithm to discover a pair of IVs from data. The experiments on synthetic and real-world datasets show that the developed IV discovery algorithm estimates accurate estimates of causal effects in comparison with the state-of-the-art IV based causal effect estimators.
Learning for control of dynamical systems with formal guarantees remains a challenging task. This paper proposes a learning framework to simultaneously stabilize an unknown nonlinear system with a neural controller and learn a neural Lyapunov function to certify a region of attraction (ROA) for the closed-loop system. The algorithmic structure consists of two neural networks and a satisfiability modulo theories (SMT) solver. The first neural network is responsible for learning the unknown dynamics. The second neural network aims to identify a valid Lyapunov function and a provably stabilizing nonlinear controller. The SMT solver then verifies that the candidate Lyapunov function indeed satisfies the Lyapunov conditions. We provide theoretical guarantees of the proposed learning framework in terms of the closed-loop stability for the unknown nonlinear system. We illustrate the effectiveness of the approach with a set of numerical experiments.
This work explores the physics-driven machine learning technique Operator Inference (OpInf) for predicting the state of chaotic dynamical systems. OpInf provides a non-intrusive approach to infer approximations of polynomial operators in reduced space without having access to the full order operators appearing in discretized models. Datasets for the physics systems are generated using conventional numerical solvers and then projected to a low-dimensional space via Principal Component Analysis (PCA). In latent space, a least-squares problem is set to fit a quadratic polynomial operator, which is subsequently employed in a time-integration scheme in order to produce extrapolations in the same space. Once solved, the inverse PCA operation is applied to reconstruct the extrapolations in the original space. The quality of the OpInf predictions is assessed via the Normalized Root Mean Squared Error (NRMSE) metric from which the Valid Prediction Time (VPT) is computed. Numerical experiments considering the chaotic systems Lorenz 96 and the Kuramoto-Sivashinsky equation show promising forecasting capabilities of the OpInf reduced order models with VPT ranges that outperform state-of-the-art machine learning methods such as backpropagation and reservoir computing recurrent neural networks [1], as well as Markov neural operators [2].
We consider the problem of inference in shift-share research designs. The choice between existing approaches that allow for unrestricted spatial correlation involves tradeoffs, varying in terms of their validity when there are relatively few or concentrated shocks, and in terms of the assumptions on the shock assignment process and treatment effects heterogeneity. We propose alternative randomization inference methods that combine the advantages of different approaches. These methods are valid in finite samples under relatively stronger assumptions, while asymptotically valid under weaker assumptions.
Neural Radiance Fields (NeRF) has been wildly applied to various tasks for its high-quality representation of 3D scenes. It takes long per-scene training time and per-image testing time. In this paper, we present EfficientNeRF as an efficient NeRF-based method to represent 3D scene and synthesize novel-view images. Although several ways exist to accelerate the training or testing process, it is still difficult to much reduce time for both phases simultaneously. We analyze the density and weight distribution of the sampled points then propose valid and pivotal sampling at the coarse and fine stage, respectively, to significantly improve sampling efficiency. In addition, we design a novel data structure to cache the whole scene during testing to accelerate the rendering speed. Overall, our method can reduce over 88\% of training time, reach rendering speed of over 200 FPS, while still achieving competitive accuracy. Experiments prove that our method promotes the practicality of NeRF in the real world and enables many applications.
Forecasting tasks surrounding the dynamics of low-level human behavior are of significance to multiple research domains. In such settings, methods for explaining specific forecasts can enable domain experts to gain insights into the predictive relationships between behaviors. In this work, we introduce and address the following question: given a probabilistic forecasting model how can we identify observed windows that the model considers salient when making its forecasts? We build upon a general definition of information-theoretic saliency grounded in human perception and extend it to forecasting settings by leveraging a crucial attribute of the domain: a single observation can result in multiple valid futures. We propose to express the saliency of an observed window in terms of the differential entropy of the resulting predicted future distribution. In contrast to existing methods that either require explicit training of the saliency mechanism or access to the internal states of the forecasting model, we obtain a closed-form solution for the saliency map for commonly used density functions in probabilistic forecasting. We empirically demonstrate how our framework can recover salient observed windows from head pose features for the sample task of speaking-turn forecasting using a synthesized conversation dataset.
We develop a theory of monotone comparative statics for models with adjustment costs. We show that comparative-statics conclusions may be drawn under the usual ordinal complementarity assumptions on the objective function, assuming very little about costs: only a mild monotonicity condition is required. We use this insight to prove a general Le Chatelier principle: under the ordinal complementarity assumptions, if short-run adjustment is subject to a monotone cost, then the long-run response to a shock is greater than the short-run response. We extend these results to a fully dynamic model of adjustment over time: the Le Chatelier principle remains valid, and under slightly stronger assumptions, optimal adjustment follows a monotone path. We apply our results to models of saving, production, pricing, labor supply and investment.
Multi-head attention empowers the recent success of transformers, the state-of-the-art models that have achieved remarkable success in sequence modeling and beyond. These attention mechanisms compute the pairwise dot products between the queries and keys, which results from the use of unnormalized Gaussian kernels with the assumption that the queries follow a mixture of Gaussian distribution. There is no guarantee that this assumption is valid in practice. In response, we first interpret attention in transformers as a nonparametric kernel regression. We then propose the FourierFormer, a new class of transformers in which the dot-product kernels are replaced by the novel generalized Fourier integral kernels. Different from the dot-product kernels, where we need to choose a good covariance matrix to capture the dependency of the features of data, the generalized Fourier integral kernels can automatically capture such dependency and remove the need to tune the covariance matrix. We theoretically prove that our proposed Fourier integral kernels can efficiently approximate any key and query distributions. Compared to the conventional transformers with dot-product attention, FourierFormers attain better accuracy and reduce the redundancy between attention heads. We empirically corroborate the advantages of FourierFormers over the baseline transformers in a variety of practical applications including language modeling and image classification.