Gaussian process state-space models (GPSSMs) are a versatile and principled family of nonlinear dynamical system models. However, existing variational learning and inference methods for GPSSMs often necessitate optimizing a substantial number of variational parameters, leading to inadequate performance and efficiency. To overcome this issue, we propose incorporating the ensemble Kalman filter (EnKF), a well-established model-based filtering technique, into the variational inference framework to approximate the posterior distribution of latent states. This utilization of EnKF can effectively exploit the dependencies between latent states and GP dynamics, while eliminating the need for parameterizing the variational distribution, thereby significantly reducing the number of variational parameters. Moreover, we show that our proposed algorithm allows straightforward evaluation of an approximated evidence lower bound (ELBO) in variational inference via simply summating multiple terms with readily available closed-form solutions. Leveraging automatic differentiation tools, we hence can maximize the ELBO and train the GPSSM efficiently. We also extend the proposed algorithm to accommodate an online setting and provide detailed algorithmic analyses and insights. Extensive evaluation on diverse real and synthetic datasets demonstrates the superiority of our EnKF-aided variational inference algorithms in terms of learning and inference performance compared to existing methods.

The likelihood ratio is a crucial quantity for statistical inference in science that enables hypothesis testing, construction of confidence intervals, reweighting of distributions, and more. Many modern scientific applications, however, make use of data- or simulation-driven models for which computing the likelihood ratio can be very difficult or even impossible. By applying the so-called ``likelihood ratio trick,'' approximations of the likelihood ratio may be computed using clever parametrizations of neural network-based classifiers. A number of different neural network setups can be defined to satisfy this procedure, each with varying performance in approximating the likelihood ratio when using finite training data. We present a series of empirical studies detailing the performance of several common loss functionals and parametrizations of the classifier output in approximating the likelihood ratio of two univariate and multivariate Gaussian distributions as well as simulated high-energy particle physics datasets.

Due to the expanding scope of machine learning (ML) to the fields of sensor networking, cooperative robotics and many other multi-agent systems, distributed deployment of inference algorithms has received a lot of attention. These algorithms involve collaboratively learning unknown parameters from dispersed data collected by multiple agents. There are two competing aspects in such algorithms, namely, intra-agent computation and inter-agent communication. Traditionally, algorithms are designed to perform both synchronously. However, certain circumstances need frugal use of communication channels as they are either unreliable, time-consuming, or resource-expensive. In this paper, we propose gossip-based asynchronous communication to leverage fast computations and reduce communication overhead simultaneously. We analyze the effects of multiple (local) intra-agent computations by the active agents between successive inter-agent communications. For local computations, Bayesian sampling via unadjusted Langevin algorithm (ULA) MCMC is utilized. The communication is assumed to be over a connected graph (e.g., as in decentralized learning), however, the results can be extended to coordinated communication where there is a central server (e.g., federated learning). We theoretically quantify the convergence rates in the process. To demonstrate the efficacy of the proposed algorithm, we present simulations on a toy problem as well as on real world data sets to train ML models to perform classification tasks. We observe faster initial convergence and improved performance accuracy, especially in the low data range. We achieve on average 78% and over 90% classification accuracy respectively on the Gamma Telescope and mHealth data sets from the UCI ML repository.

This manuscript is technical memoir about my work at Stanford and Microsoft Research. Included are fundamental concepts central to machine learning and artificial intelligence, applications of these concepts, and stories behind their creation.

Abstract: Recent strides in the field of neural computation has seen the adoption of Winner-Take-All (WTA) circuits to facilitate the unification of hierarchical Bayesian inference and spiking neural networks as a neurobiologically plausible model of information processing. However, researchers have not yet reached consensus about how best to translate the stochastic responses from these networks into discrete decisions, a process known as population decoding. Despite being an often underexamined part of SNNs, in this work we show that population decoding has a significanct impact on the classification performance of WTA networks. For this purpose, we apply a WTA network to the problem of cancer subtype diagnosis from multi-omic data, using datasets from The Cancer Genome Atlas (TCGA). In doing so we utilise a novel implementation of gene similarity networks, a feature encoding technique based on Kohoen's self-organising map algorithm. We further show that the impact of selecting certain population decoding methods is amplified when facing imbalanced datasets. Multi-omics data integration in cancer diagnosis Alternatively, some research focuses on the timedependent refers to the integration of information from various relationship of spiking neurons, for biological "omics" e.g., genomics, transcriptomics, instance by weighting neuron responses more highly metabolomics, to provide a more comprehensive based on how quickly they fire (Grün & Rotter, 2010; understanding of the molecular landscape of cancer. In order to extract information from SNNs, we neurobiologically inspired method of information examine the spikes generated by a population of processing which aim to solve tasks using plausible neurons in response to a stimulus.

Missing data is a widespread problem in many domains, creating challenges in data analysis and decision making. Traditional techniques for dealing with missing data, such as excluding incomplete records or imputing simple estimates (e.g., mean), are computationally efficient but may introduce bias and disrupt variable relationships, leading to inaccurate analyses. Model-based imputation techniques offer a more robust solution that preserves the variability and relationships in the data, but they demand significantly more computation time, limiting their applicability to small datasets. This work enables efficient, high-quality, and scalable data imputation within a database system using the widely used MICE method. We adapt this method to exploit computation sharing and a ring abstraction for faster model training. To impute both continuous and categorical values, we develop techniques for in-database learning of stochastic linear regression and Gaussian discriminant analysis models. Our MICE implementations in PostgreSQL and DuckDB outperform alternative MICE implementations and model-based imputation techniques by up to two orders of magnitude in terms of computation time, while maintaining high imputation quality.

Handwritten signature verification poses a formidable challenge in biometrics and document authenticity. The objective is to ascertain the authenticity of a provided handwritten signature, distinguishing between genuine and forged ones. This issue has many applications in sectors such as finance, legal documentation, and security. Currently, the field of computer vision and machine learning has made significant progress in the domain of handwritten signature verification. The outcomes, however, may be enhanced depending on the acquired findings, the structure of the datasets, and the used models. Four stages make up our suggested strategy. First, we collected a large dataset of 12600 images from 420 distinct individuals, and each individual has 30 signatures of a certain kind (All authors signatures are genuine). In the subsequent stage, the best features from each image were extracted using a deep learning model named MobileNetV2. During the feature selection step, three selectors neighborhood component analysis (NCA), Chi2, and mutual info (MI) were used to pull out 200, 300, 400, and 500 features, giving a total of 12 feature vectors. Finally, 12 results have been obtained by applying machine learning techniques such as SVM with kernels (rbf, poly, and linear), KNN, DT, Linear Discriminant Analysis, and Naive Bayes. Without employing feature selection techniques, our suggested offline signature verification achieved a classification accuracy of 91.3%, whereas using the NCA feature selection approach with just 300 features it achieved a classification accuracy of 97.7%. High classification accuracy was achieved using the designed and suggested model, which also has the benefit of being a self-organized framework. Consequently, using the optimum minimally chosen features, the proposed method could identify the best model performance and result validation prediction vectors.

Meta reinforcement learning (meta-RL) methods such as RL$^2$ have emerged as promising approaches for learning data-efficient RL algorithms tailored to a given task distribution. However, these RL algorithms struggle with long-horizon tasks and out-of-distribution tasks since they rely on recurrent neural networks to process the sequence of experiences instead of summarizing them into general RL components such as value functions. Moreover, even transformers have a practical limit to the length of histories they can efficiently reason about before training and inference costs become prohibitive. In contrast, traditional RL algorithms are data-inefficient since they do not leverage domain knowledge, but they do converge to an optimal policy as more data becomes available. In this paper, we propose RL$^3$, a principled hybrid approach that combines traditional RL and meta-RL by incorporating task-specific action-values learned through traditional RL as an input to the meta-RL neural network. We show that RL$^3$ earns greater cumulative reward on long-horizon and out-of-distribution tasks compared to RL$^2$, while maintaining the efficiency of the latter in the short term. Experiments are conducted on both custom and benchmark discrete domains from the meta-RL literature that exhibit a range of short-term, long-term, and complex dependencies.

Spatial perception is a key task in several machine intelligence applications such as robotics and computer vision. In general, it involves the nonlinear estimation of hidden variables that represent the system's state. However, in the presence of measurement outliers, the standard nonlinear least squared formulation results in poor estimates. Several methods have been considered in the literature to improve the reliability of the estimation process. Most methods are based on heuristics since guaranteed global robust estimation is not generally practical due to high computational costs. Recently general purpose robust estimation heuristics have been proposed that leverage existing non-minimal solvers available for the outlier-free formulations without the need for an initial guess. In this work, we propose three Bayesian heuristics that have similar structures. We evaluate these heuristics in practical scenarios to demonstrate their merits in different applications including 3D point cloud registration, mesh registration and pose graph optimization. The general computational advantages our proposals offer make them attractive candidates for spatial perception tasks.

A core challenge in survival analysis is to model the distribution of censored time-to-event data, where the event of interest may be a death, failure, or occurrence of a specific event. Previous studies have showed that ranking and maximum likelihood estimation (MLE)loss functions are widely-used for survival analysis. However, ranking loss only focus on the ranking of survival time and does not consider potential effect of samples for exact survival time values. Furthermore, the MLE is unbounded and easily subject to outliers (e.g., censored data), which may cause poor performance of modeling. To handle the complexities of learning process and exploit valuable survival time values, we propose a time-adaptive coordinate loss function, TripleSurv, to achieve adaptive adjustments by introducing the differences in the survival time between sample pairs into the ranking, which can encourage the model to quantitatively rank relative risk of pairs, ultimately enhancing the accuracy of predictions. Most importantly, the TripleSurv is proficient in quantifying the relative risk between samples by ranking ordering of pairs, and consider the time interval as a trade-off to calibrate the robustness of model over sample distribution. Our TripleSurv is evaluated on three real-world survival datasets and a public synthetic dataset. The results show that our method outperforms the state-of-the-art methods and exhibits good model performance and robustness on modeling various sophisticated data distributions with different censor rates. Our code will be available upon acceptance.