An increasing number of data science and machine learning problems rely on computation with tensors, which better capture the multi-way relationships and interactions of data than matrices. When tapping into this critical advantage, a key challenge is to develop computationally efficient and provably correct algorithms for extracting useful information from tensor data that are simultaneously robust to corruptions and ill-conditioning. This paper tackles tensor robust principal component analysis (RPCA), which aims to recover a low-rank tensor from its observations contaminated by sparse corruptions, under the Tucker decomposition. To minimize the computation and memory footprints, we propose to directly recover the low-dimensional tensor factors -- starting from a tailored spectral initialization -- via scaled gradient descent (ScaledGD), coupled with an iteration-varying thresholding operation to adaptively remove the impact of corruptions. Theoretically, we establish that the proposed algorithm converges linearly to the true low-rank tensor at a constant rate that is independent with its condition number, as long as the level of corruptions is not too large. Empirically, we demonstrate that the proposed algorithm achieves better and more scalable performance than state-of-the-art matrix and tensor RPCA algorithms through synthetic experiments and real-world applications.

Transfer learning has recently become the dominant paradigm of machine learning. Pre-trained models fine-tuned for downstream tasks achieve better performance with fewer labelled examples. Nonetheless, it remains unclear how to develop models that specialise towards multiple tasks without incurring negative interference and that generalise systematically to non-identically distributed tasks. Modular deep learning has emerged as a promising solution to these challenges. In this framework, units of computation are often implemented as autonomous parameter-efficient modules. Information is conditionally routed to a subset of modules and subsequently aggregated. These properties enable positive transfer and systematic generalisation by separating computation from routing and updating modules locally. We offer a survey of modular architectures, providing a unified view over several threads of research that evolved independently in the scientific literature. Moreover, we explore various additional purposes of modularity, including scaling language models, causal inference, programme induction, and planning in reinforcement learning. Finally, we report various concrete applications where modularity has been successfully deployed such as cross-lingual and cross-modal knowledge transfer. Related talks and projects to this survey, are available at https://www.modulardeeplearning.com/.

Spectral clustering became a popular choice for data clustering for its ability of uncovering clusters of different shapes. However, it is not always preferable over other clustering methods due to its computational demands. One of the effective ways to bypass these computational demands is to perform spectral clustering on a subset of points (data representatives) then generalize the clustering outcome, this is known as approximate spectral clustering (ASC). ASC uses sampling or quantization to select data representatives. This makes it vulnerable to 1) performance inconsistency (since these methods have a random step either in initialization or training), 2) local statistics loss (because the pairwise similarities are extracted from data representatives instead of data points). We proposed a refined version of $k$-nearest neighbor graph, in which we keep data points and aggressively reduce number of edges for computational efficiency. Local statistics were exploited to keep the edges that do not violate the intra-cluster distances and nullify all other edges in the $k$-nearest neighbor graph. We also introduced an optional step to automatically select the number of clusters $C$. The proposed method was tested on synthetic and real datasets. Compared to ASC methods, the proposed method delivered a consistent performance despite significant reduction of edges.

Usually standard algorithms employ a loss where each error is the mere absolute difference between the true value and the prediction, in case of a regression task. In the present, we introduce several error weighting schemes that are a generalization of the consolidated routine. We study both a binary classification model for Support Vector Classification and a regression net for Multi-layer Perceptron. Results proves that the error is never worse than the standard procedure and several times it is better.

Accurate traffic prediction benefits urban management and improves transportation efficiency. Recently, data-driven methods have been widely applied in traffic prediction and outperformed traditional methods. However, data-driven methods normally require massive data for training, while data scarcity is ubiquitous in low-developmental or newly constructed regions. To tackle this problem, we can extract meta knowledge from data-rich cities to data-scarce cities via transfer learning. Besides, relations among urban regions can be organized into various semantic graphs, e.g. proximity and POI similarity, which is barely considered in previous studies. In this paper, we propose Semantic-Fused Hierarchical Graph Transfer Learning (SF-HGTL) model to achieve knowledge transfer across cities with fused semantics. In detail, we employ hierarchical graph transformation followed by meta-knowledge retrieval to achieve knowledge transfer in various granularity. In addition, we introduce meta semantic nodes to reduce the number of parameters as well as share information across semantics. Afterwards, the parameters of the base model are generated by fused semantic embeddings to predict traffic status in terms of task heterogeneity. We implement experiments on five real-world datasets and verify the effectiveness of our SF-HGTL model by comparing it with other baselines.

The study of algorithmic fairness received growing attention recently. This stems from the awareness that bias in the input data for machine learning systems may result in discriminatory outputs. For clustering tasks, one of the most central notions of fairness is the formalization by Chierichetti, Kumar, Lattanzi, and Vassilvitskii [NeurIPS 2017]. A clustering is said to be fair, if each cluster has the same distribution of manifestations of a sensitive attribute as the whole input set. This is motivated by various applications where the objects to be clustered have sensitive attributes that should not be over- or underrepresented. We discuss the applicability of this fairness notion to Correlation Clustering. The existing literature on the resulting Fair Correlation Clustering problem either presents approximation algorithms with poor approximation guarantees or severely limits the possible distributions of the sensitive attribute (often only two manifestations with a 1:1 ratio are considered). Our goal is to understand if there is hope for better results in between these two extremes. To this end, we consider restricted graph classes which allow us to characterize the distributions of sensitive attributes for which this form of fairness is tractable from a complexity point of view. While existing work on Fair Correlation Clustering gives approximation algorithms, we focus on exact solutions and investigate whether there are efficiently solvable instances. The unfair version of Correlation Clustering is trivial on forests, but adding fairness creates a surprisingly rich picture of complexities. We give an overview of the distributions and types of forests where Fair Correlation Clustering turns from tractable to intractable. The most surprising insight to us is the fact that the cause of the hardness of Fair Correlation Clustering is not the strictness of the fairness condition.

Understanding how subsets of items are chosen from offered sets is critical to assortment planning, wireless network planning, and many other applications. There are two seemingly unrelated subset choice models that capture dependencies between items: intuitive and interpretable random utility models; and tractable determinantal point processes (DPPs). This paper connects the two. First, all DPPs are shown to be random utility models. Next, a determinantal choice model that enjoys the best of both worlds is specified; the model is shown to subsume logistic regression when dependence is minimal, and MNL when dependence is maximally negative. This makes the model interpretable, while retaining the tractability of DPPs. A simulation study verifies that the model can learn a continuum of negative dependencies from data, and an applied study using original experimental data produces novel insights on wireless interference in LoRa networks.

Some approaches [14], [15] aim to infer global feature importance for a given data Transparency of machine learning (ML) based system, projection. Another work [16] estimates feature importance applied in the real world, is nowadays a widely accepted locally for a vicinity around a projected data point, using requirement - the importance of transparency was also recognized locally linear models. A recent paper [17] proposes to use by the policy makers and therefore made its way local feature importance explanations by computing a local into legal regulations like the EU's GDPR [1]. A popular linear approximation for each reduced dimension, extracting way of achieving transparency is by means of explanations [2] feature importances from the weight vectors. Further, saliency which then gave rise to the field of eXplainable AI (XAI) [3], map approaches such as the layer-wise relevance propagation [4]. Although a lot of different explanation methodologies (LRP) [18] could in principle be applied to a parametric for ML based systems have been developed [2], [4], it is dimensionality reduction mapping in order to obtain locally important to realize that it is still somewhat unclear what relevant features. However, these approaches do not provide exactly makes up a good explanation [5], [6]. Therefore contrasting explanations, in which we are interested here.

The emerging availability of trained machine learning models has put forward the novel concept of Machine Learning Model Market in which one can harness the collective intelligence of multiple well-trained models to improve the performance of the resultant model through one-shot federated learning and ensemble learning in a data-free manner. However, picking the models available in the market for ensemble learning is time-consuming, as using all the models is not always the best approach. It is thus crucial to have an effective ensemble selection strategy that can find a good subset of the base models for the ensemble. Conventional ensemble selection techniques are not applicable, as we do not have access to the local datasets of the parties in the federated learning setting. In this paper, we present a novel Data-Free Diversity-Based method called DeDES to address the ensemble selection problem for models generated by one-shot federated learning in practical applications such as model markets. Experiments showed that our method can achieve both better performance and higher efficiency over 5 datasets and 4 different model structures under the different data-partition strategies.

Bernstein's condition is a key assumption that guarantees fast rates in machine learning. For example, the Gibbs algorithm with prior $\pi$ has an excess risk in $O(d_{\pi}/n)$, as opposed to the standard $O(\sqrt{d_{\pi}/n})$, where $n$ denotes the number of observations and $d_{\pi}$ is a complexity parameter which depends on the prior $\pi$. In this paper, we examine the Gibbs algorithm in the context of meta-learning, i.e., when learning the prior $\pi$ from $T$ tasks (with $n$ observations each) generated by a meta distribution. Our main result is that Bernstein's condition always holds at the meta level, regardless of its validity at the observation level. This implies that the additional cost to learn the Gibbs prior $\pi$, which will reduce the term $d_\pi$ across tasks, is in $O(1/T)$, instead of the expected $O(1/\sqrt{T})$. We further illustrate how this result improves on standard rates in three different settings: discrete priors, Gaussian priors and mixture of Gaussians priors.