Federated Semi-Supervised Learning (FSSL) leverages both labeled and unlabeled data on clients to collaboratively train a model.In FSSL, the heterogeneous data can introduce prediction bias into the model, causing the model's prediction to skew towards some certain classes. Existing FSSL methods primarily tackle this issue by enhancing consistency in model parameters or outputs. However, as the models themselves are biased, merely constraining their consistency is not sufficient to alleviate prediction bias. In this paper, we explore this bias from a Bayesian perspective and demonstrate that it principally originates from label prior bias within the training data. Building upon this insight, we propose a debiasing method for FSSL named FedDB. FedDB utilizes the Average Prediction Probability of Unlabeled Data (APP-U) to approximate the biased prior.During local training, FedDB employs APP-U to refine pseudo-labeling through Bayes' theorem, thereby significantly reducing the label prior bias. Concurrently, during the model aggregation, FedDB uses APP-U from participating clients to formulate unbiased aggregate weights, thereby effectively diminishing bias in the global model. Experimental results show that FedDB can surpass existing FSSL methods. The code is available at https://github.com/GuogangZhu/FedDB.

We present new results to model and understand the role of encoder-decoder design in machine learning (ML) from an information-theoretic angle. We use two main information concepts, information sufficiency (IS) and mutual information loss (MIL), to represent predictive structures in machine learning. Our first main result provides a functional expression that characterizes the class of probabilistic models consistent with an IS encoder-decoder latent predictive structure. This result formally justifies the encoder-decoder forward stages many modern ML architectures adopt to learn latent (compressed) representations for classification. To illustrate IS as a realistic and relevant model assumption, we revisit some known ML concepts and present some interesting new examples: invariant, robust, sparse, and digital models. Furthermore, our IS characterization allows us to tackle the fundamental question of how much performance (predictive expressiveness) could be lost, using the cross entropy risk, when a given encoder-decoder architecture is adopted in a learning setting. Here, our second main result shows that a mutual information loss quantifies the lack of expressiveness attributed to the choice of a (biased) encoder-decoder ML design. Finally, we address the problem of universal cross-entropy learning with an encoder-decoder design where necessary and sufficiency conditions are established to meet this requirement. In all these results, Shannon's information measures offer new interpretations and explanations for representation learning.

Non-negative matrix factorization with the generalized Kullback-Leibler divergence (NMF) and latent Dirichlet allocation (LDA) are two popular approaches for dimensionality reduction of non-negative data. Here, we show that NMF with $\ell_1$ normalization constraints on the columns of both matrices of the decomposition and a Dirichlet prior on the columns of one matrix is equivalent to LDA. To show this, we demonstrate that explicitly accounting for the scaling ambiguity of NMF by adding $\ell_1$ normalization constraints to the optimization problem allows a joint update of both matrices in the widely used multiplicative updates (MU) algorithm. When both of the matrices are normalized, the joint MU algorithm leads to probabilistic latent semantic analysis (PLSA), which is LDA without a Dirichlet prior. Our approach of deriving joint updates for NMF also reveals that a Lasso penalty on one matrix together with an $\ell_1$ normalization constraint on the other matrix is insufficient to induce any sparsity.

We study reinforcement learning with multinomial logistic (MNL) function approximation where the underlying transition probability kernel of the Markov decision processes (MDPs) is parametrized by an unknown transition core with features of state and action. For the finite horizon episodic setting with inhomogeneous state transitions, we propose provably efficient algorithms with randomized exploration having frequentist regret guarantees. For our first algorithm, $\texttt{RRL-MNL}$, we adapt optimistic sampling to ensure the optimism of the estimated value function with sufficient frequency and establish that $\texttt{RRL-MNL}$ is both statistically and computationally efficient, achieving a $\tilde{O}(\kappa^{-1} d^{\frac{3}{2}} H^{\frac{3}{2}} \sqrt{T})$ frequentist regret bound with constant-time computational cost per episode. Here, $d$ is the dimension of the transition core, $H$ is the horizon length, $T$ is the total number of steps, and $\kappa$ is a problem-dependent constant. Despite the simplicity and practicality of $\texttt{RRL-MNL}$, its regret bound scales with $\kappa^{-1}$, which is potentially large in the worst case. To improve the dependence on $\kappa^{-1}$, we propose $\texttt{ORRL-MNL}$, which estimates the value function using local gradient information of the MNL transition model. We show that its frequentist regret bound is $\tilde{O}(d^{\frac{3}{2}} H^{\frac{3}{2}} \sqrt{T} + \kappa^{-1} d^2 H^2)$. To the best of our knowledge, these are the first randomized RL algorithms for the MNL transition model that achieve both computational and statistical efficiency. Numerical experiments demonstrate the superior performance of the proposed algorithms.

Predictive posterior densities (PPDs) are of interest in approximate Bayesian inference. Typically, these are estimated by simple Monte Carlo (MC) averages using samples from the approximate posterior. We observe that the signal-to-noise ratio (SNR) of such estimators can be extremely low. An analysis for exact inference reveals SNR decays exponentially as there is an increase in (a) the mismatch between training and test data, (b) the dimensionality of the latent space, or (c) the size of the test data relative to the training data. Further analysis extends these results to approximate inference. To remedy the low SNR problem, we propose replacing simple MC sampling with importance sampling using a proposal distribution optimized at test time on a variational proxy for the SNR and demonstrate that this yields greatly improved estimates.

Identifiability of statistical models is a key notion in unsupervised representation learning. Recent work of nonlinear independent component analysis (ICA) employs auxiliary data and has established identifiable conditions. This paper proposes a statistical model of two latent vectors with single auxiliary data generalizing nonlinear ICA, and establishes various identifiability conditions. Unlike previous work, the two latent vectors in the proposed model can have arbitrary dimensions, and this property enables us to reveal an insightful dimensionality relation among two latent vectors and auxiliary data in identifiability conditions. Furthermore, surprisingly, we prove that the indeterminacies of the proposed model has the same as \emph{linear} ICA under certain conditions: The elements in the latent vector can be recovered up to their permutation and scales. Next, we apply the identifiability theory to a statistical model for graph data. As a result, one of the identifiability conditions includes an appealing implication: Identifiability of the statistical model could depend on the maximum value of link weights in graph data. Then, we propose a practical method for identifiable graph embedding. Finally, we numerically demonstrate that the proposed method well-recovers the latent vectors and model identifiability clearly depends on the maximum value of link weights, which supports the implication of our theoretical results

We propose a novel approach to sequential Bayesian inference based on variational Bayes. The key insight is that, in the online setting, we do not need to add the KL term to regularize to the prior (which comes from the posterior at the previous timestep); instead we can optimize just the expected log-likelihood, performing a single step of natural gradient descent starting at the prior predictive. We prove this method recovers exact Bayesian inference if the model is conjugate, and empirically outperforms other online VB methods in the non-conjugate setting, such as online learning for neural networks, especially when controlling for computational costs.

Despite their benefits in terms of simplicity, low computational cost and data requirement, parametric machine learning algorithms, such as linear discriminant analysis, quadratic discriminant analysis or logistic regression, suffer from serious drawbacks including linearity, poor fit of features to the usually imposed normal distribution and high dimensionality. Batch kernel-based nonparametric classifier, which overcomes the linearity and normality of features constraints, represent an interesting alternative for supervised classification problem. However, it suffers from the ``curse of dimension". The problem can be alleviated by the explosive sample size in the era of big data, while large-scale data size presents some challenges in the storage of data and the calculation of the classifier. These challenges make the classical batch nonparametric classifier no longer applicable. This motivates us to develop a fast algorithm adapted to the real-time calculation of the nonparametric classifier in massive as well as streaming data frameworks. This online classifier includes two steps. First, we consider an online principle components analysis to reduce the dimension of the features with a very low computation cost. Then, a stochastic approximation algorithm is deployed to obtain a real-time calculation of the nonparametric classifier. The proposed methods are evaluated and compared to some commonly used machine learning algorithms for real-time fetal well-being monitoring. The study revealed that, in terms of accuracy, the offline (or Batch), as well as, the online classifiers are good competitors to the random forest algorithm. Moreover, we show that the online classifier gives the best trade-off accuracy/computation cost compared to the offline classifier.

A cornerstone of geometric reconstruction, rotation averaging seeks the set of absolute rotations that optimally explains a set of measured relative orientations between them. In addition to being an integral part of bundle adjustment and structure-from-motion, the problem of synchronizing rotations also finds applications in visual simultaneous localization and mapping, where it is used as an initialization for iterative solvers, and camera network calibration. Nevertheless, this optimization problem is both non-convex and high-dimensional. In this paper, we address it from a maximum likelihood estimation standpoint and make a twofold contribution. Firstly, we set forth a novel primal-dual method, motivated by the widely accepted spectral initialization. Further, we characterize stationary points of rotation averaging in cycle graphs topologies and contextualize this result within spectral graph theory. We benchmark the proposed method in multiple settings and certify our solution via duality theory, achieving a significant gain in precision and performance.

Generative models based on flow matching have attracted significant attention for their simplicity and superior performance in high-resolution image synthesis. By leveraging the instantaneous change-of-variables formula, one can directly compute image likelihoods from a learned flow, making them enticing candidates as priors for downstream tasks such as inverse problems. In particular, a natural approach would be to incorporate such image probabilities in a maximum-a-posteriori (MAP) estimation problem. A major obstacle, however, lies in the slow computation of the log-likelihood, as it requires backpropagating through an ODE solver, which can be prohibitively slow for high-dimensional problems. In this work, we propose an iterative algorithm to approximate the MAP estimator efficiently to solve a variety of linear inverse problems. Our algorithm is mathematically justified by the observation that the MAP objective can be approximated by a sum of $N$ ``local MAP'' objectives, where $N$ is the number of function evaluations. By leveraging Tweedie's formula, we show that we can perform gradient steps to sequentially optimize these objectives. We validate our approach for various linear inverse problems, such as super-resolution, deblurring, inpainting, and compressed sensing, and demonstrate that we can outperform other methods based on flow matching.