Deep neural networks achieve stellar generalisation even when they have enough parameters to easily fit all their training data. We study the dynamics and the performance of two-layer neural networks in the teacher-student setup, where one network, the student, is trained on data generated by another network, called the teacher, using stochastic gradient descent (SGD). We show how the dynamics of SGD is captured by a set of differential equations and prove that this description is asymptotically exact in the limit of large inputs. Using this framework, we calculate the final generalisation error of student networks that have more parameters than their teachers. We find that the final generalisation error of the student increases with network size when training only the first layer, but stays constant or even decreases with size when training both layers. We show that these different behaviours have their root in the different solutions SGD finds for different activation functions. Our results indicate that achieving good generalisation in neural networks goes beyond the properties of SGD alone and depends on the interplay of at least the algorithm, the model architecture, and the data set.

The goal of this paper is to design image classification systems that, after an initial multi-task training phase, can automatically adapt to new tasks encountered at test time. We introduce a conditional neural process based approach to the multi-task classification setting for this purpose, and establish connections to the meta-learning and few-shot learning literature. The resulting approach, called CNAPs, comprises a classifier whose parameters are modulated by an adaptation network that takes the current task's dataset as input. We demonstrate that CNAPs achieves state-of-the-art results on the challenging Meta-Dataset benchmark indicating high-quality transfer-learning. We show that the approach is robust, avoiding both over-fitting in low-shot regimes and under-fitting in high-shot regimes. Timing experiments reveal that CNAPs is computationally efficient at test-time as it does not involve gradient based adaptation. Finally, we show that trained models are immediately deployable to continual learning and active learning where they can outperform existing approaches that do not leverage transfer learning.

Consider semi-supervised learning for classification, where both labeled and unlabeled data are available for training. The goal is to exploit both datasets to achieve higher prediction accuracy than just using labeled data alone. We develop a semi-supervised logistic learning method based on exponential tilt mixture models, by extending a statistical equivalence between logistic regression and exponential tilt modeling. We study maximum nonparametric likelihood estimation and derive novel objective functions which are shown to be Fisher consistent. We also propose regularized estimation and construct simple and highly interpretable EM algorithms. Finally, we present numerical results which demonstrate the advantage of the proposed methods compared with existing methods.

In supervised clustering, standard techniques for learning a pairwise dissimilarity function often suffer from a discrepancy between the training and clustering objectives, leading to poor cluster quality. Rectifying this discrepancy necessitates matching the procedure for training the dissimilarity function to the clustering algorithm. In this paper, we introduce a method for training the dissimilarity function in a way that is tightly coupled with hierarchical clustering, in particular single linkage. However, the appropriate clustering algorithm for a given dataset is often unknown. Thus we introduce an approach to supervised hierarchical clustering that smoothly interpolates between single, average, and complete linkage, and we give a training procedure that simultaneously learns a linkage function and a dissimilarity function. We accomplish this with a novel Exponential Linkage function that has a learnable parameter that controls the interpolation. In experiments on four datasets, our joint training procedure consistently matches or outperforms the next best training procedure/linkage function pair and gives up to 8 points improvement in dendrogram purity over discrepant pairs.

Most stochastic optimization methods use gradients once before discarding them. While variance reduction methods have shown that reusing past gradients can be beneficial when there is a finite number of datapoints, they do not easily extend to the online setting. One issue is the staleness due to using past gradients. We propose to correct this staleness using the idea of implicit gradient transport (IGT) which transforms gradients computed at previous iterates into gradients evaluated at the current iterate without using the Hessian explicitly. In addition to reducing the variance and bias of our updates over time, IGT can be used as a drop-in replacement for the gradient estimate in a number of well-understood methods such as heavy ball or Adam. We show experimentally that it achieves state-of-the-art results on a wide range of architectures and benchmarks. Additionally, the IGT gradient estimator yields the optimal asymptotic convergence rate for online stochastic optimization in the restricted setting where the Hessians of all component functions are equal.

Plasmodium falciparum malaria still poses one of the greatest threats to human life with over 200 million cases globally leading to half-million deaths annually. Of these, 90% of cases and of the mortality occurs in sub-Saharan Africa, mostly among children. Although malaria prediction systems are central to the 2016-2030 malaria Global Technical Strategy, currently these are inadequate at capturing and estimating the burden of disease in highly endemic countries. We developed and validated a computational system that exploits the predictive power of current Machine Learning approaches on 22-years of prospective data from the high-transmission holoendemic malaria urban-densely-populated sub-Saharan West-Africa metropolis of Ibadan. Our dataset of >9x104 screened study participants attending our clinical and community services from 1996 to 2017 contains monthly prevalence, temporal, environmental and host features. Our Locality-specific Elastic-Net based Malaria Prediction System (LEMPS) achieves good generalization performance, both in magnitude and direction of the prediction, when tasked to predict monthly prevalence on previously unseen validation data (MAE<=6x10-2, MSE<=7x10-3) within a range of (+0.1 to -0.05) error-tolerance which is relevant and usable for aiding decision-support in a holoendemic setting. LEMPS is well-suited for malaria prediction, where there are multiple features which are correlated with one another, and trading-off between regularization-strength L1-norm and L2-norm allows the system to retain stability. Data-driven systems are critical for regionally-adaptable surveillance, management of control strategies and resource allocation across stretched healthcare systems.

The K-nearest neighbor (KNN) classifier is one of the simplest and most common classifiers, yet its performance competes with the most complex classifiers in the literature. The core of this classifier depends mainly on measuring the distance or similarity between the tested example and the training examples. This raises a major question about which distance measures to be used for the KNN classifier among a large number of distance and similarity measures? This review attempts to answer the previous question through evaluating the performance (measured by accuracy, precision and recall) of the KNN using a large number of distance measures, tested on a number of real world datasets, with and without adding different levels of noise. The experimental results show that the performance of KNN classifier depends significantly on the distance used, the results showed large gaps between the performances of different distances. We found that a recently proposed non-convex distance performed the best when applied on most datasets comparing to the other tested distances. In addition, the performance of the KNN degraded only about $20\%$ while the noise level reaches $90\%$, this is true for all the distances used. This means that the KNN classifier using any of the top $10$ distances tolerate noise to a certain degree. Moreover, the results show that some distances are less affected by the added noise comparing to other distances.

The choice of constellations largely affects the performance of communication systems. When designing constellations, both the locations and probability of occurrence of the points can be optimized. These approaches are referred to as geometric and probabilistic shaping, respectively. Usually, the geometry of the constellation is fixed, e.g., quadrature amplitude modulation (QAM) is used. In such cases, the achievable information rate can still be improved by probabilistic shaping. In this work, we show how autoencoders can be leveraged to perform probabilistic shaping of constellations. We devise an information-theoretical description of autoencoders, which allows learning of capacity-achieving symbol distributions and constellations. Recently, machine learning techniques to perform geometric shaping were proposed. However, probabilistic shaping is more challenging as it requires the optimization of discrete distributions. Furthermore, the proposed method enables joint probabilistic and geometric shaping of constellations over any channel model. Simulation results show that the learned constellations achieve information rates very close to capacity on an additive white Gaussian noise (AWGN) channel and outperform existing approaches on both AWGN and fading channels.

Graph-based semi-supervised learning is the problem of propagating labels from a small number of labelled data points to a larger set of unlabelled data. This paper is concerned with the consistency of optimization-based techniques for such problems, in the limit where the labels have small noise and the underlying unlabelled data is well clustered. We study graph-based probit for binary classification, and a natural generalization of this method to multi-class classification using one-hot encoding. The resulting objective function to be optimized comprises the sum of a quadratic form defined through a rational function of the graph Laplacian, involving only the unlabelled data, and a fidelity term involving only the labelled data.

City Logistics is characterized by multiple stakeholders that often have different views of such a complex system. From a public policy perspective, identifying stakeholders, issues and trends is a daunting challenge, only partially addressed by traditional observation systems. Nowadays, social media is one of the biggest channels of public expression and is often used to communicate opinions and content related to City Logistics. The idea of this research is that analysing social media content could help in understanding the public perception of City logistics. This paper offers a methodology for collecting content from Twitter and implementing Machine Learning techniques (Unsupervised Learning and Natural Language Processing), to perform content and sentiment analysis. The proposed methodology is applied to more than 110 000 tweets containing City Logistics key-terms. Results allowed the building of an Interest Map of concepts and a Sentiment Analysis to determine if City Logistics entries are positive, negative or neutral.