After a surge in popularity of supervised Deep Learning, the desire to reduce the dependence on curated, labelled data sets and to leverage the vast quantities of unlabelled data available recently triggered renewed interest in unsupervised learning algorithms. Despite a significantly improved performance due to approaches such as the identification of disentangled latent representations, contrastive learning, and clustering optimisations, the performance of unsupervised machine learning still falls short of its hypothesised potential. Machine learning has previously taken inspiration from neuroscience and cognitive science with great success. However, this has mostly been based on adult learners with access to labels and a vast amount of prior knowledge. In order to push unsupervised machine learning forward, we argue that developmental science of infant cognition might hold the key to unlocking the next generation of unsupervised learning approaches. Conceptually, human infant learning is the closest biological parallel to artificial unsupervised learning, as infants too must learn useful representations from unlabelled data. In contrast to machine learning, these new representations are learned rapidly and from relatively few examples. Moreover, infants learn robust representations that can be used flexibly and efficiently in a number of different tasks and contexts. We identify five crucial factors enabling infants' quality and speed of learning, assess the extent to which these have already been exploited in machine learning, and propose how further adoption of these factors can give rise to previously unseen performance levels in unsupervised learning.

The advent of deep learning (DL) has fundamentally changed the landscape of modern software. Generally, a DL system is comprised of several interconnected computational units that form "layers" which perform mathematical transformations, according to sets of learnable parameters, on data passing through them. These architectures can be "trained" for specific tasks by updating the parameters according to a model's performance on a labeled set of training data. DL represents a fundamental shift in the manner by which machines learn patterns from data by automatically extracting salient features for a given computational task, as opposed to relying upon human intuition. These DL systems can be viewed as an inflection point for software development, as they enable new capabilities that cannot be realized cost-effectively through "traditional" software wherein the behavior of a program must be specified analytically.

In this thesis we look into programming by example (PBE), which is about finding a program mapping given inputs to given outputs. PBE has traditionally seen a split between formal versus neural approaches, where formal approaches typically involve deductive techniques such as SAT solvers and types, while the neural approaches involve training on sample input-outputs with their corresponding program, typically using sequence-based machine learning techniques such as LSTMs [41]. As a result of this split, programming types had yet to be used in neural program synthesis techniques. We propose a way to incorporate programming types into a neural program synthesis approach for PBE. We introduce the Typed Neuro-Symbolic Program Synthesis (TNSPS) method based on this idea, and test it in the functional programming context to empirically verify type information may help improve generalization in neural synthesizers on limited-size datasets. Our TNSPS model builds upon the existing Neuro-Symbolic Program Synthesis (NSPS), a tree-based neural synthesizer combining info from input-output examples plus the current program, by further exposing information on types of those input-output examples, of the grammar production rules, as well as of the hole that we wish to expand in the program. We further explain how we generated a dataset within our domain, which uses a limited subset of Haskell as the synthesis language. Finally we discuss several topics of interest that may help take these ideas further. For reproducibility, we release our code publicly.

Mitigating the risk arising from extreme events is a fundamental goal with many applications, such as the modelling of natural disasters, financial crashes, epidemics, and many others. To manage this risk, a vital step is to be able to understand or generate a wide range of extreme scenarios. Existing approaches based on Generative Adversarial Networks (GANs) excel at generating realistic samples, but seek to generate typical samples, rather than extreme samples. Hence, in this work, we propose ExGAN, a GAN-based approach to generate realistic and extreme samples. To model the extremes of the training distribution in a principled way, our work draws from Extreme Value Theory (EVT), a probabilistic approach for modelling the extreme tails of distributions. For practical utility, our framework allows the user to specify both the desired extremeness measure, as well as the desired extremeness probability they wish to sample at. Experiments on real US Precipitation data show that our method generates realistic samples, based on visual inspection and quantitative measures, in an efficient manner. Moreover, generating increasingly extreme examples using ExGAN can be done in constant time (with respect to the extremeness probability), as opposed to the exponential time required by the baseline approach.

Inducing a meaningful structural representation from one or a set of dialogues is a crucial but challenging task in computational linguistics. Advancement made in this area is critical for dialogue system design and discourse analysis. It can also be extended to solve grammatical inference. In this work, we propose to incorporate structured attention layers into a Variational Recurrent Neural Network (VRNN) model with discrete latent states to learn dialogue structure in an unsupervised fashion. Compared to a vanilla VRNN, structured attention enables a model to focus on different parts of the source sentence embeddings while enforcing a structural inductive bias. Experiments show that on two-party dialogue datasets, VRNN with structured attention learns semantic structures that are similar to templates used to generate this dialogue corpus. While on multi-party dialogue datasets, our model learns an interactive structure demonstrating its capability of distinguishing speakers or addresses, automatically disentangling dialogues without explicit human annotation.

It is still common to use Q-learning and temporal difference (TD) learning-even though they have divergence issues and sound Gradient TD alternatives exist-because divergence seems rare and they typically perform well. However, recent work with large neural network learning systems reveals that instability is more common than previously thought. Practitioners face a difficult dilemma: choose an easy to use and performant TD method, or a more complex algorithm that is more sound but harder to tune and all but unexplored with non-linear function approximation or control. In this paper, we introduce a new method called TD with Regularized Corrections (TDRC), that attempts to balance ease of use, soundness, and performance. It behaves as well as TD, when TD performs well, but is sound in cases where TD diverges. We empirically investigate TDRC across a range of problems, for both prediction and control, and for both linear and non-linear function approximation, and show, potentially for the first time, that gradient TD methods could be a better alternative to TD and Q-learning.

Artificial Intelligence (AI) applications based on Deep Neural Networks (DNN) or Deep Learning (DL) have become popular due to their success in solving problems likeimage analysis and speech recognition. Training a DNN is computationally intensive and High Performance Computing(HPC) has been a key driver in AI growth. Virtualisation and container technology have led to the convergence of cloud and HPC infrastructure. These infrastructures with diverse hardware increase the complexity of deploying and optimising AI training workloads. AI training deployments in HPC or cloud can be optimised with target-specific libraries, graph compilers, andby improving data movement or IO. Graph compilers aim to optimise the execution of a DNN graph by generating an optimised code for a target hardware/backend. As part of SODALITE (a Horizon 2020 project), MODAK tool is developed to optimise application deployment in software defined infrastructures. Using input from the data scientist and performance modelling, MODAK maps optimal application parameters to a target infrastructure and builds an optimised container. In this paper, we introduce MODAK and review container technologies and graph compilers for AI. We illustrate optimisation of AI training deployments using graph compilers and Singularity containers. Evaluation using MNIST-CNN and ResNet50 training workloads shows that custom built optimised containers outperform the official images from DockerHub. We also found that the performance of graph compilers depends on the target hardware and the complexity of the neural network.

Deep neural networks have been shown to be vulnerable to membership inference attacks wherein the attacker aims to detect whether specific input data were used to train the model. These attacks can potentially leak private or proprietary data. We present a new extension of Fano's inequality and employ it to theoretically establish that the probability of success for a membership inference attack on a deep neural network can be bounded using the mutual information between its inputs and its activations. This enables the use of mutual information to measure the susceptibility of a DNN model to membership inference attacks. In our empirical evaluation, we show that the correlation between the mutual information and the susceptibility of the DNN model to membership inference attacks is 0.966, 0.996, and 0.955 for CIFAR-10, SVHN and GTSRB models, respectively.

Memes are graphics and text overlapped so that together they present concepts that become dubious if one of them is absent. It is spread mostly on social media platforms, in the form of jokes, sarcasm, motivating, etc. After the success of BERT in Natural Language Processing (NLP), researchers inclined to Visual-Linguistic (VL) multimodal problems like memes classification, image captioning, Visual Question Answering (VQA), and many more. Unfortunately, many memes get uploaded each day on social media platforms that need automatic censoring to curb misinformation and hate. Recently, this issue has attracted the attention of researchers and practitioners. State-of-the-art methods that performed significantly on other VL dataset, tends to fail on memes classification. In this context, this work aims to conduct a comprehensive study on memes classification, generally on the VL multimodal problems and cutting edge solutions. We propose a generalized framework for VL problems. We cover the early and next-generation works on VL problems. Finally, we identify and articulate several open research issues and challenges. This is the first study that presents the generalized view of the advanced classification techniques concerning memes classification to the best of our knowledge. We believe this study presents a clear road-map for the Machine Learning (ML) research community to implement and enhance memes classification techniques.

The heterogeneous network is a robust data abstraction that can model entities of different types interacting in various ways. Such heterogeneity brings rich semantic information but presents nontrivial challenges in aggregating the heterogeneous relationships between objects - especially those of higher-order indirect relations. Recent graph neural network approaches for representation learning on heterogeneous networks typically employ the attention mechanism, which is often only optimized for predictions based on direct links. Furthermore, even though most deep learning methods can aggregate higher-order information by building deeper models, such a scheme can diminish the degree of interpretability. To overcome these challenges, we explore an architecture - Layer-stacked ATTention Embedding (LATTE) - that automatically decomposes higher-order meta relations at each layer to extract the relevant heterogeneous neighborhood structures for each node. Additionally, by successively stacking layer representations, the learned node embedding offers a more interpretable aggregation scheme for nodes of different types at different neighborhood ranges. We conducted experiments on several benchmark heterogeneous network datasets. In both transductive and inductive node classification tasks, LATTE can achieve state-of-the-art performance compared to existing approaches, all while offering a lightweight model. With extensive experimental analyses and visualizations, the framework can demonstrate the ability to extract informative insights on heterogeneous networks.