Neural Processes (NPs) are deep probabilistic models that represent stochastic processes by conditioning their prior distributions on a set of context points. Despite their advantages in uncertainty estimation for complex distributions, NPs enforce parameterization coupling between the conditional prior model and the posterior model. We show that this coupling amounts to prior misspecification and revisit the NP objective to address this issue. More specifically, we propose Rényi Neural Processes (RNP), a method that replaces the standard KL divergence with the Rényi divergence, dampening the effects of the misspecified prior during posterior updates. We validate our approach across multiple benchmarks including regression and image inpainting tasks, and show significant performance improvements of RNPs in real-world problems. Our extensive experiments show consistently better log-likelihoods over state-of-the-art NP models.

The ability to evaluate uncertainties in evolving data streams has become equally, if not more, crucial than building a static predictor. For instance, during the pandemic, a model should consider possible uncertainties such as governmental policies, meteorological features, and vaccination schedules. Neural process families (NPFs) have recently shone a light on predicting such uncertainties by bridging Gaussian processes (GPs) and neural networks (NNs). Their abilities to output average predictions and the acceptable variances, i.e., uncertainties, made them suitable for predictions with insufficient data, such as meta-learning or few-shot learning. However, existing models have not addressed continual learning which imposes a stricter constraint on the data access. Regarding this, we introduce a member meta-continual learning with neural process (MCLNP) for uncertainty estimation. We enable two levels of uncertainty estimations: the local uncertainty on certain points and the global uncertainty p(z) that represents the function evolution in dynamic environments. To facilitate continual learning, we hypothesize that the previous knowledge can be applied to the current task, hence adopt a coreset as a memory buffer to alleviate catastrophic forgetting. The relationships between the degree of global uncertainties with the intratask diversity and model complexity are discussed. We have estimated prediction uncertainties with multiple evolving types including abrupt/gradual/recurrent shifts. The applications encompass meta-continual learning in the 1-D, 2-D datasets, and a novel spatial-temporal COVID dataset. The results show that our method outperforms the baselines on the likelihood and can rebound quickly even for heavily evolved data streams.

The ability to deal with uncertainty in machine learning models has become equally, if not more, crucial to their predictive ability itself. For instance, during the pandemic, governmental policies and personal decisions are constantly made around uncertainties. Targeting this, Neural Process Families (NPFs) have recently shone a light on prediction with uncertainties by bridging Gaussian processes and neural networks. Latent neural process, a member of NPF, is believed to be capable of modelling the uncertainty on certain points (local uncertainty) as well as the general function priors (global uncertainties). Nonetheless, some critical questions remain unresolved, such as a formal definition of global uncertainties, the causality behind global uncertainties, and the manipulation of global uncertainties for generative models. Regarding this, we build a member GloBal Convolutional Neural Process(GBCoNP) that achieves the SOTA log-likelihood in latent NPFs. It designs a global uncertainty representation p(z), which is an aggregation on a discretized input space. The causal effect between the degree of global uncertainty and the intra-task diversity is discussed. The learnt prior is analyzed on a variety of scenarios, including 1D, 2D, and a newly proposed spatial-temporal COVID dataset. Our manipulation of the global uncertainty not only achieves generating the desired samples to tackle few-shot learning, but also enables the probability evaluation on the functional priors.

A model can be easily influenced by unseen factors in nonstationary environments and fail to fit dynamic data distribution. In a classification scenario, this is known as a concept drift. For instance, the shopping preference of customers may change after they move from one city to another. Therefore, a shopping website or application should alter recommendations based on its poorer predictions of such user patterns. In this article, we propose a novel approach called the multiscale drift detection test (MDDT) that efficiently localizes abrupt drift points when feature values fluctuate, meaning that the current model needs immediate adaption. MDDT is based on a resampling scheme and a paired student t -test. It applies a detection procedure on two different scales. Initially, the detection is performed on a broad scale to check if recently gathered drift indicators remain stationary. If a drift is claimed, a narrow scale detection is performed to trace the refined change time. This multiscale structure reduces the massive time of constantly checking and filters noises in drift indicators. Experiments are performed to compare the proposed method with several algorithms via synthetic and real-world datasets. The results indicate that it outperforms others when abrupt shift datasets are handled, and achieves the highest recall score in localizing drift points.

A support vector machine (SVM) plays a prominent role in classic machine learning, especially classification and regression. Through its structural risk minimization, it has enjoyed a good reputation in effectively reducing overfitting, avoiding dimensional disaster, and not falling into local minima. Nevertheless, existing SVMs do not perform well when facing class imbalance and large-scale samples. Undersampling is a plausible alternative to solve imbalanced problems in some way, but suffers from soaring computational complexity and reduced accuracy because of its enormous iterations and random sampling process. To improve their classification performance in dealing with data imbalance problems, this work proposes a weighted undersampling (WU) scheme for SVM based on space geometry distance, and thus produces an improved algorithm named WU-SVM. In WU-SVM, majority samples are grouped into some subregions (SRs) and assigned different weights according to their Euclidean distance to the hyper plane. The samples in an SR with higher weight have more chance to be sampled and put to use in each learning iteration, so as to retain the data distribution information of original data sets as much as possible. Comprehensive experiments are performed to test WU-SVM via 21 binary-class and six multiclass publically available data sets. The results show that it well outperforms the state-of-the-art methods in terms of three popular metrics for imbalanced classification, i.e., area under the curve, F-Measure, and G-Mean.