My main research motivation is to develop complete autonomous agents that interact with people socially. For an agent to be social with respect to humans, it needs to be able to parse and process the multitude of aspects that comprise the human cultural experience. That in itself gives rise to many fascinating learning problems. I am interested in tackling these fundamental problems from an empirical as well as a theoretical perspective. Music, as a general target domain, serves as an excellent testbed for these research ideas. Musical skills---playing music (alone or in a group), analyzing music or composing it---all involve extremely advanced knowledge representation and problem solving tools. Creating "musical agents"---agents that can interact richly with people in the music domain---is a challenge that holds the potential of advancing social agents research, and contributing important and broadly applicable AI knowledge. This belief is fueled not just by my background in computer science and artificial intelligence, but also by my deep passion for music as well as my extensive musical training. One key aspect of musical intelligence which hasn’t been sufficiently studied is that of sequential decision-making. My thesis strives to answer the following question: How can a sequential decision making perspective guide us in the creation of better music agents, and social agents in general? More specifically, this thesis focuses on two aspects of musical intelligence: music recommendation and multiagent interaction in the context of music.

Existing action detection algorithms usually generate action proposals through an extensive search over the video at multiple temporal scales, which brings about huge computational overhead and deviates from the human perception procedure. We argue that the process of detecting actions should be naturally one of observation and refinement: observe the current window and refine the span of attended window to cover true action regions. In this paper, we propose a Self-Adaptive Proposal (SAP) model that learns to find actions through continuously adjusting the temporal bounds in a self-adaptive way. The whole process can be deemed as an agent, which is firstly placed at the beginning of the video and traverse the whole video by adopting a sequence of transformations on the current attended region to discover actions according to a learned policy. We utilize reinforcement learning, especially the Deep Q-learning algorithm to learn the agent’s decision policy. In addition, we use temporal pooling operation to extract more effective feature representation for the long temporal window, and design a regression network to adjust the position offsets between predicted results and the ground truth. Experiment results on THUMOS’14 validate the effectiveness of SAP, which can achieve competitive performance with current action detection algorithms via much fewer proposals.

Face grouping remains a challenging problem despite the remarkable capability of deep learning approaches in learning face representation. In particular, grouping results can still be egregious given profile faces and a large number of uninteresting faces and noisy detections. Often, a user needs to correct the erroneous grouping manually. In this study, we formulate a novel face grouping framework that learns clustering strategy from ground-truth simulated behavior. This is achieved through imitation learning (a.k.a apprenticeship learning or learning by watching) via inverse reinforcement learning (IRL). In contrast to existing clustering approaches that group instances by similarity, our framework makes sequential decision to dynamically decide when to merge two face instances/groups driven by short- and long-term rewards. Extensive experiments on three benchmark datasets show that our framework outperforms unsupervised and supervised baselines.

The existing image captioning approaches typically train a one-stage sentence decoder, which is difficult to generate rich fine-grained descriptions. On the other hand, multi-stage image caption model is hard to train due to the vanishing gradient problem. In this paper, we propose a coarse-to-fine multi-stage prediction framework for image captioning, composed of multiple decoders each of which operates on the output of the previous stage, producing increasingly refined image descriptions. Our proposed learning approach addresses the difficulty of vanishing gradients during training by providing a learning objective function that enforces intermediate supervisions. Particularly, we optimize our model with a reinforcement learning approach which utilizes the output of each intermediate decoder's test-time inference algorithm as well as the output of its preceding decoder to normalize the rewards, which simultaneously solves the well-known exposure bias problem and the loss-evaluation mismatch problem. We extensively evaluate the proposed approach on MSCOCO and show that our approach can achieve the state-of-the-art performance.

The existing methods for image captioning usually train the language model under the cross entropy loss, which results in the exposure bias and inconsistency of evaluation metric. Recent research has shown these two issues can be well addressed by policy gradient method in reinforcement learning domain attributable to its unique capability of directly optimizing the discrete and non-differentiable evaluation metric. In this paper, we utilize reinforcement learning method to train the image captioning model. Specifically, we train our image captioning model to maximize the overall reward of the sentences by adopting the temporal-difference (TD) learning method, which takes the correlation between temporally successive actions into account. In this way, we assign different values to different words in one sampled sentence by a discounted coefficient when back-propagating the gradient with the REINFORCE algorithm, enabling the correlation between actions to be learned. Besides, instead of estimating a "baseline" to normalize the rewards with another network, we utilize the reward of another Monte-Carlo sample as the "baseline" to avoid high variance. We show that our proposed method can improve the quality of generated captions and outperforms the state-of-the-art methods on the benchmark dataset MS COCO in terms of seven evaluation metrics.

In many reinforcement learning problems, parameters of the model may vary with its phase while the agent attempts to learn through its interaction with the environment. For example, an autonomous car's reward on selecting a path may depend on traffic conditions at the time of the day or the transition dynamics of a drone may depend on the current wind direction. Many such processes exhibit a cyclic phase-structure and could be represented with a control policy parameterized over a circular or cyclic phase space. Attempting to model such phase variations with a standard data-driven approach (e.g. deep networks) without explicitly modeling the phase of the model can be challenging. Ambiguities may arise as the optimal action for a given state can vary depending on the phase. To better model cyclic environments, we propose phase-parameterized policies and value function approximators that explicitly enforce a cyclic structure to the policy or value space. We apply our phase-parameterized reinforcement learning approach to both feed-forward and recurrent deep networks in the context of trajectory optimization and locomotion problems. Our experiments show that our proposed approach has superior modeling performance than traditional function approximators in cyclic environments.

Formal verification provides a high degree of confidence in safe system operation, but only if reality matches the verified model. Although a good model will be accurate most of the time, even the best models are incomplete. This is especially true in Cyber-Physical Systems because high-fidelity physical models of systems are expensive to develop and often intractable to verify. Conversely, reinforcement learning-based controllers are lauded for their flexibility in unmodeled environments, but do not provide guarantees of safe operation. This paper presents an approach for provably safe learning that provides the best of both worlds: the exploration and optimization capabilities of learning along with the safety guarantees of formal verification. Our main insight is that formal verification combined with verified runtime monitoring can ensure the safety of a learning agent. Verification results are preserved whenever learning agents limit exploration within the confounds of verified control choices as long as observed reality comports with the model used for off-line verification. When a model violation is detected, the agent abandons efficiency and instead attempts to learn a control strategy that guides the agent to a modeled portion of the state space. We prove that our approach toward incorporating knowledge about safe control into learning systems preserves safety guarantees, and demonstrate that we retain the empirical performance benefits provided by reinforcement learning. We also explore various points in the design space for these justified speculative controllers in a simple model of adaptive cruise control model for autonomous cars.

Robust reinforcement learning aims to produce policies that have strong guarantees even in the face of environments/transition models whose parameters have strong uncertainty. Existing work uses value-based methods and the usual primitive action setting. In this paper, we propose robust methods for learning temporally abstract actions, in the framework of options. We present a Robust Options Policy Iteration (ROPI) algorithm with convergence guarantees, which learns options that are robust to model uncertainty. We utilize ROPI to learn robust options with the Robust Options Deep Q Network (RO-DQN) that solves multiple tasks and mitigates model misspecification due to model uncertainty. We present experimental results which suggest that policy iteration with linear features may have an inherent form of robustness when using coarse feature representations. In addition, we present experimental results which demonstrate that robustness helps policy iteration implemented on top of deep neural networks to generalize over a much broader range of dynamics than non-robust policy iteration.

We consider sequential decision making problems under uncertainty, in which a user has a general idea of the task to achieve, and gives advice to an agent in charge of computing an optimal policy. Many different notions of advice have been proposed in somewhat different settings, especially in the field of inverse reinforcement learning and for resolution of Markov Decision Problems with Imprecise Rewards. Two key questions are whether the advice required by a specific method is natural for the user to give, and how much advice is needed for the agent to compute a good policy, as evaluated by the user. We give a unified view of a number of proposals made in the literature, and propose a new notion of advice, which corresponds to a user telling why she would take a given action in a given state. For all these notions, we discuss their naturalness for a user and the integration of advice. We then report on an experimental study of the amount of advice needed for the agent to compute a good policy. Our study shows in particular that continual interaction between the user and the agent is worthwhile, and sheds light on the pros and cons of each type of advice.

In this paper, we introduce a generalized value iteration network (GVIN), which is an end-to-end neural network planning module. GVIN emulates the value iteration algorithm by using a novel graph convolution operator, which enables GVIN to learn and plan on irregular spatial graphs. We propose three novel differentiable kernels as graph convolution operators and show that the embedding-based kernel achieves the best performance. Furthermore, we present episodic Q-learning, an improvement upon traditional n-step Q-learning that stabilizes training for VIN and GVIN. Lastly, we evaluate GVIN on planning problems in 2D mazes, irregular graphs, and real-world street networks, showing that GVIN generalizes well for both arbitrary graphs and unseen graphs of larger scaleand outperforms a naive generalization of VIN (discretizing a spatial graph into a 2D image).