Collective action demands that individuals efficiently coordinate how much, where, and when to cooperate. Laboratory experiments have extensively explored the first part of this process, demonstrating that a variety of social-cognitive mechanisms influence how much individuals choose to invest in group efforts. However, experimental research has been unable to shed light on how social cognitive mechanisms contribute to the where and when of collective action. We leverage multi-agent deep reinforcement learning to model how a social-cognitive mechanism--specifically, the intrinsic motivation to achieve a good reputation--steers group behavior toward specific spatial and temporal strategies for collective action in a social dilemma. We also collect behavioral data from groups of human participants challenged with the same dilemma. The model accurately predicts spatial and temporal patterns of group behavior: in this public goods dilemma, the intrinsic motivation for reputation catalyzes the development of a non-territorial, turn-taking strategy to coordinate collective action.

In this thesis, we develop various techniques for working with sets in machine learning. Each input or output is not an image or a sequence, but a set: an unordered collection of multiple objects, each object described by a feature vector. Their unordered nature makes them suitable for modeling a wide variety of data, ranging from objects in images to point clouds to graphs. Deep learning has recently shown great success on other types of structured data, so we aim to build the necessary structures for sets into deep neural networks. The first focus of this thesis is the learning of better set representations (sets as input). Existing approaches have bottlenecks that prevent them from properly modeling relations between objects within the set. To address this issue, we develop a variety of techniques for different scenarios and show that alleviating the bottleneck leads to consistent improvements across many experiments. The second focus of this thesis is the prediction of sets (sets as output). Current approaches do not take the unordered nature of sets into account properly. We determine that this results in a problem that causes discontinuity issues with many set prediction tasks and prevents them from learning some extremely simple datasets. To avoid this problem, we develop two models that properly take the structure of sets into account. Various experiments show that our set prediction techniques can significantly benefit over existing approaches.

In offline reinforcement learning (RL), we seek to utilize offline data to evaluate (or learn) policies in scenarios where the data are collected from a distribution that substantially differs from that of the target policy to be evaluated. Recent theoretical advances have shown that such sample-efficient offline RL is indeed possible provided certain strong representational conditions hold, else there are lower bounds exhibiting exponential error amplification (in the problem horizon) unless the data collection distribution has only a mild distribution shift relative to the target policy. This work studies these issues from an empirical perspective to gauge how stable offline RL methods are. In particular, our methodology explores these ideas when using features from pre-trained neural networks, in the hope that these representations are powerful enough to permit sample efficient offline RL. Through extensive experiments on a range of tasks, we see that substantial error amplification does occur even when using such pre-trained representations (trained on the same task itself); we find offline RL is stable only under extremely mild distribution shift. The implications of these results, both from a theoretical and an empirical perspective, are that successful offline RL (where we seek to go beyond the low distribution shift regime) requires substantially stronger conditions beyond those which suffice for successful supervised learning.

Despite the rich theoretical foundation of model-based deep reinforcement learning (RL) agents, their effectiveness in real-world robotics-applications is less studied and understood. In this paper, we, therefore, investigate how such agents generalize to real-world autonomous-vehicle control-tasks, where advanced model-free deep RL algorithms fail. In particular, we set up a series of time-lap tasks for an F1TENTH racing robot, equipped with high-dimensional LiDAR sensors, on a set of test tracks with a gradual increase in their complexity. In this continuous-control setting, we show that model-based agents capable of learning in imagination, substantially outperform model-free agents with respect to performance, sample efficiency, successful task completion, and generalization. Moreover, we show that the generalization ability of model-based agents strongly depends on the observation-model choice. Finally, we provide extensive empirical evidence for the effectiveness of model-based agents provided with long enough memory horizons in sim2real tasks.

Limb deficiency severely affects the daily lives of amputees and drives efforts to provide functional robotic prosthetic hands to compensate this deprivation. Convolutional neural network-based computer vision control of the prosthetic hand has received increased attention as a method to replace or complement physiological signals due to its reliability by training visual information to predict the hand gesture. Mounting a camera into the palm of a prosthetic hand is proved to be a promising approach to collect visual data. However, the grasp type labelled from the eye and hand perspective may differ as object shapes are not always symmetric. Thus, to represent this difference in a realistic way, we employed a dataset containing synchronous images from eye- and hand- view, where the hand-perspective images are used for training while the eye-view images are only for manual labelling. Electromyogram (EMG) activity and movement kinematics data from the upper arm are also collected for multi-modal information fusion in future work. Moreover, in order to include human-in-the-loop control and combine the computer vision with physiological signal inputs, instead of making absolute positive or negative predictions, we build a novel probabilistic classifier according to the Plackett-Luce model. To predict the probability distribution over grasps, we exploit the statistical model over label rankings to solve the permutation domain problems via a maximum likelihood estimation, utilizing the manually ranked lists of grasps as a new form of label. We indicate that the proposed model is applicable to the most popular and productive convolutional neural network frameworks.

Upper limb and hand functionality is critical to many activities of daily living and the amputation of one can lead to significant functionality loss for individuals. From this perspective, advanced prosthetic hands of the future are anticipated to benefit from improved shared control between a robotic hand and its human user, but more importantly from the improved capability to infer human intent from multimodal sensor data to provide the robotic hand perception abilities regarding the operational context. Such multimodal sensor data may include various environment sensors including vision, as well as human physiology and behavior sensors including electromyography and inertial measurement units. A fusion methodology for environmental state and human intent estimation can combine these sources of evidence in order to help prosthetic hand motion planning and control. In this paper, we present a dataset of this type that was gathered with the anticipation of cameras being built into prosthetic hands, and computer vision methods will need to assess this hand-view visual evidence in order to estimate human intent. Specifically, paired images from human eye-view and hand-view of various objects placed at different orientations have been captured at the initial state of grasping trials, followed by paired video, EMG and IMU from the arm of the human during a grasp, lift, put-down, and retract style trial structure. For each trial, based on eye-view images of the scene showing the hand and object on a table, multiple humans were asked to sort in decreasing order of preference, five grasp types appropriate for the object in its given configuration relative to the hand. The potential utility of paired eye-view and hand-view images was illustrated by training a convolutional neural network to process hand-view images in order to predict eye-view labels assigned by humans.

Echo State Networks (ESNs) are a special type of recurrent neural networks (RNNs), in which the input and recurrent connections are traditionally generated randomly, and only the output weights are trained. Despite the recent success of ESNs in various tasks of audio, image and radar recognition, we postulate that a purely random initialization is not the ideal way of initializing ESNs. The aim of this work is to propose an unsupervised initialization of the input connections using the K-Means algorithm on the training data. We show that this initialization performs equivalently or superior than a randomly initialized ESN whilst needing significantly less reservoir neurons (2000 vs. 4000 for spoken digit recognition, and 300 vs. 8000 neurons for f0 extraction) and thus reducing the amount of training time. Furthermore, we discuss that this approach provides the opportunity to estimate the suitable size of the reservoir based on the prior knowledge about the data.

We present a slot-wise, object-based transition model that decomposes a scene into objects, aligns them (with respect to a slot-wise object memory) to maintain a consistent order across time, and predicts how those objects evolve over successive frames. The model is trained end-to-end without supervision using losses at the level of the object-structured representation rather than pixels. Thanks to its alignment module, the model deals properly with two issues that are not handled satisfactorily by other transition models, namely object persistence and object identity. We show that the combination of an object-level loss and correct object alignment over time enables the model to outperform a state-of-the-art baseline, and allows it to deal well with object occlusion and re-appearance in partially observable environments.

Continuous engineering of autonomous driving functions commonly requires deploying vehicles in road testing to obtain inputs that cause problematic decisions. Although the discovery leads to producing an improved system, it also challenges the foundation of testing using equivalence classes and the associated relative test coverage criterion. In this paper, we propose believed equivalence, where the establishment of an equivalence class is initially based on expert belief and is subject to a set of available test cases having a consistent valuation. Upon a newly encountered test case that breaks the consistency, one may need to refine the established categorization in order to split the originally believed equivalence into two. Finally, we focus on modules implemented using deep neural networks where every category partitions an input over the real domain. We establish new equivalence classes by guiding the new test cases following directions suggested by its k-nearest neighbors, complemented by local robustness testing. The concept is demonstrated in a lane-keeping assist module indicating the potential of our proposed approach.

Sentiment analysis is the basis of intelligent human-computer interaction. As one of the frontier research directions of artificial intelligence, it can help computers better identify human intentions and emotional states so that provide more personalized services. However, as human present sentiments by spoken words, gestures, facial expressions and others which involve variable forms of data including text, audio, video, etc., it poses many challenges to this study. Due to the limitations of unimodal sentiment analysis, recent research has focused on the sentiment analysis of videos containing time series data of multiple modalities. When analyzing videos with multimodal data, the key problem is how to fuse these heterogeneous data. In consideration that the contribution of each modality is different, current fusion methods tend to extract the important information of single modality prior to fusion, which ignores the consistency and complementarity of bimodal interaction and has influences on the final decision. To solve this problem, a video sentiment analysis method using multi-head attention with bimodal information augmented is proposed. Based on bimodal interaction, more important bimodal features are assigned larger weights. In this way, different feature representations are adaptively assigned corresponding attention for effective multimodal fusion. Extensive experiments were conducted on both Chinese and English public datasets. The results show that our approach outperforms the existing methods and can give an insight into the contributions of bimodal interaction among three modalities.