In this thesis, we explore and apply methods inspired by the free energy principle to two important areas in machine learning and neuroscience. The free energy principle is a general mathematical theory of the necessary information-theoretic behaviours of systems which maintain a separation from their environment. A core postulate of the theory is that complex systems can be seen as performing variational Bayesian inference and minimizing an information-theoretic quantity called the variational free energy. The free energy principle originated in, and has been extremely influential in theoretical neuroscience, having spawned a number of neurophysiologically realistic process theories, and maintaining close links with Bayesian Brain viewpoints. The thesis is split into three main parts where we apply methods and insights from the free energy principle to understand questions first in perception, then action, and finally learning. Specifically, in the first section, we focus on the theory of predictive coding, a neurobiologically plausible process theory derived from the free energy principle under certain assumptions, which argues that the primary function of the brain is to minimize prediction errors. We focus on scaling up predictive coding architectures and simulate large-scale predictive coding networks for perception on machine learning benchmarks; we investigate predictive coding's relationship to other classical filtering algorithms, and we demonstrate that many biologically implausible aspects of current models of predictive coding can be relaxed without unduly harming the performance of predictive coding models which allows for a potentially more literal translation of predictive coding theory into cortical microcircuits. In the second part of the thesis, we focus on the application of methods deriving from the free energy principle to action. We study the extension of methods of'active inference', a neurobiologically grounded account of action through variational message passing, to utilize deep artificial neural networks, allowing these methods to'scale up' to be competitive with state of the art deep reinforcement learning methods.

Optimization in the latent space of variational autoencoders is a promising approach to generate high-dimensional discrete objects that maximize an expensive black-box property (e.g., drug-likeness in molecular generation, function approximation with arithmetic expressions). However, existing methods lack robustness as they may decide to explore areas of the latent space for which no data was available during training and where the decoder can be unreliable, leading to the generation of unrealistic or invalid objects. We propose to leverage the epistemic uncertainty of the decoder to guide the optimization process. This is not trivial though, as a naive estimation of uncertainty in the high-dimensional and structured settings we consider would result in high estimator variance. To solve this problem, we introduce an importance sampling-based estimator that provides more robust estimates of epistemic uncertainty. Our uncertainty-guided optimization approach does not require modifications of the model architecture nor the training process. It produces samples with a better trade-off between black-box objective and validity of the generated samples, sometimes improving both simultaneously. We illustrate these advantages across several experimental settings in digit generation, arithmetic expression approximation and molecule generation for drug design.

We employ both random forests and LSTM networks (more precisely CuDNNLSTM) as training methodologies to analyze their effectiveness in forecasting out-of-sample directional movements of constituent stocks of the S&P 500 from January 1993 till December 2018 for intraday trading. We introduce a multi-feature setting consisting not only of the returns with respect to the closing prices, but also with respect to the opening prices and intraday returns. As trading strategy, we use Krauss et al. (2017) and Fischer & Krauss (2018) as benchmark. On each trading day, we buy the 10 stocks with the highest probability and sell short the 10 stocks with the lowest probability to outperform the market in terms of intraday returns -- all with equal monetary weight. Our empirical results show that the multi-feature setting provides a daily return, prior to transaction costs, of 0.64% using LSTM networks, and 0.54% using random forests. Hence we outperform the single-feature setting in Fischer & Krauss (2018) and Krauss et al. (2017) consisting only of the daily returns with respect to the closing prices, having corresponding daily returns of 0.41% and of 0.39% with respect to LSTM and random forests, respectively.

Knowledge Distillation (KD) is a popular technique to transfer knowledge from a teacher model or ensemble to a student model. Its success is generally attributed to the privileged information on similarities/consistency between the class distributions or intermediate feature representations of the teacher model and the student model. However, directly pushing the student model to mimic the probabilities/features of the teacher model to a large extent limits the student model in learning undiscovered knowledge/features. In this paper, we propose a novel inheritance and exploration knowledge distillation framework (IE-KD), in which a student model is split into two parts - inheritance and exploration. The inheritance part is learned with a similarity loss to transfer the existing learned knowledge from the teacher model to the student model, while the exploration part is encouraged to learn representations different from the inherited ones with a dis-similarity loss. Our IE-KD framework is generic and can be easily combined with existing distillation or mutual learning methods for training deep neural networks. Extensive experiments demonstrate that these two parts can jointly push the student model to learn more diversified and effective representations, and our IE-KD can be a general technique to improve the student network to achieve SOTA performance. Furthermore, by applying our IE-KD to the training of two networks, the performance of both can be improved w.r.t. deep mutual learning. The code and models of IE-KD will be make publicly available at https://github.com/yellowtownhz/IE-KD.

Current domain-independent, classical planners require symbolic models of the problem domain and instance as input, resulting in a knowledge acquisition bottleneck. Meanwhile, although deep learning has achieved significant success in many fields, the knowledge is encoded in a subsymbolic representation which is incompatible with symbolic systems such as planners. We propose Latplan, an unsupervised architecture combining deep learning and classical planning. Given only an unlabeled set of image pairs showing a subset of transitions allowed in the environment (training inputs), Latplan learns a complete propositional PDDL action model of the environment. Later, when a pair of images representing the initial and the goal states (planning inputs) is given, Latplan finds a plan to the goal state in a symbolic latent space and returns a visualized plan execution. We evaluate Latplan using image-based versions of 6 planning domains: 8-puzzle, 15-Puzzle, Blocksworld, Sokoban and Two variations of LightsOut.

Cybersecurity is a very challenging topic of research nowadays, as digitalization increases the interaction of people, software and services on the Internet by means of technology devices and networks connected to it. The field is broad and has a lot of unexplored ground under numerous disciplines such as management, psychology, and data science. Its large disciplinary spectrum and many significant research topics generate a considerable amount of information, making it hard for us to find what we are looking for when researching a particular subject. This work proposes a new search engine for scientific publications which combines both information retrieval and reading comprehension algorithms to extract answers from a collection of domain-specific documents. The proposed solution although being applied to the context of cybersecurity exhibited great generalization capabilities and can be easily adapted to perform under other distinct knowledge domains.

Model-agnostic meta-learning (MAML) is arguably the most popular meta-learning algorithm nowadays, given its flexibility to incorporate various model architectures and to be applied to different problems. Nevertheless, its performance on few-shot classification is far behind many recent algorithms dedicated to the problem. In this paper, we point out several key facets of how to train MAML to excel in few-shot classification. First, we find that a large number of gradient steps are needed for the inner loop update, which contradicts the common usage of MAML for few-shot classification. Second, we find that MAML is sensitive to the permutation of class assignments in meta-testing: for a few-shot task of $N$ classes, there are exponentially many ways to assign the learned initialization of the $N$-way classifier to the $N$ classes, leading to an unavoidably huge variance. Third, we investigate several ways for permutation invariance and find that learning a shared classifier initialization for all the classes performs the best. On benchmark datasets such as MiniImageNet and TieredImageNet, our approach, which we name UNICORN-MAML, performs on a par with or even outperforms state-of-the-art algorithms, while keeping the simplicity of MAML without adding any extra sub-networks.

Images can convey rich semantics and induce various emotions in viewers. Recently, with the rapid advancement of emotional intelligence and the explosive growth of visual data, extensive research efforts have been dedicated to affective image content analysis (AICA). In this survey, we will comprehensively review the development of AICA in the recent two decades, especially focusing on the state-of-the-art methods with respect to three main challenges -- the affective gap, perception subjectivity, and label noise and absence. We begin with an introduction to the key emotion representation models that have been widely employed in AICA and description of available datasets for performing evaluation with quantitative comparison of label noise and dataset bias. We then summarize and compare the representative approaches on (1) emotion feature extraction, including both handcrafted and deep features, (2) learning methods on dominant emotion recognition, personalized emotion prediction, emotion distribution learning, and learning from noisy data or few labels, and (3) AICA based applications. Finally, we discuss some challenges and promising research directions in the future, such as image content and context understanding, group emotion clustering, and viewer-image interaction.

In the big data and AI era, context is widely exploited as extra information which makes it easier to learn a more complex pattern in machine learning systems. However, most of the existing related studies seldom take context into account. The difficulty lies in the unknown generalization ability of both context and its modeling techniques across different scenarios. To fill the above gaps, we conduct a large-scale analytical and empirical study on the spatiotemporal crowd prediction (STCFP) problem that is a widely-studied and hot research topic. We mainly make three efforts:(i) we develop new taxonomy about both context features and context modeling techniques based on extensive investigations in prevailing STCFP research; (ii) we conduct extensive experiments on seven datasets with hundreds of millions of records to quantitatively evaluate the generalization ability of both distinct context features and context modeling techniques; (iii) we summarize some guidelines for researchers to conveniently utilize context in diverse applications.

Studying neural network loss landscapes provides insights into the nature of the underlying optimization problems. Unfortunately, loss landscapes are notoriously difficult to visualize in a human-comprehensible fashion. One common way to address this problem is to plot linear slices of the landscape, for example from the initial state of the network to the final state after optimization. On the basis of this analysis, prior work has drawn broader conclusions about the difficulty of the optimization problem. In this paper, we put inferences of this kind to the test, systematically evaluating how linear interpolation and final performance vary when altering the data, choice of initialization, and other optimizer and architecture design choices. Further, we use linear interpolation to study the role played by individual layers and substructures of the network. We find that certain layers are more sensitive to the choice of initialization and optimizer hyperparameter settings, and we exploit these observations to design custom optimization schemes. However, our results cast doubt on the broader intuition that the presence or absence of barriers when interpolating necessarily relates to the success of optimization.