Total: 17
Chatbots or conversational agents have enjoyed great popularity in recent years. They surprisingly perform sensitive tasks in modern societies. However, despite the fact that they offer help, support, and fellowship, there is a task that is not yet mastered: dealing with complex emotions and simulating human sensations. This research aims to design an architecture for an emotional conversation agent for long-text conversations (multi-turns). This agent is intended to work in areas where the analysis of users feelings plays a leading role. This work refers to natural language understanding and response generation.
Knowledge graph embedding models enable representation learning on multi-relational graphs and are used in security sensitive domains. But, their security analysis has received little attention. I will research security of these models by designing adversarial attacks against them, improving their adversarial robustness and evaluating the effect of proposed improvement on their interpretability.
Neural networks have contributed to tremendous progress in the domains of computer vision, speech processing, and other real-world applications. However, recent studies have shown that these state-of-the-art models can be easily compromised by adding small imperceptible perturbations. My thesis summary frames the problem of adversarial robustness as an equivalent problem of learning suitable features that leads to good generalization in neural networks. This is motivated from learning in humans which is not trivially fooled by such perturbations due to robust feature learning which shows good out-of-sample generalization.
Many mobile applications and virtual conversational agents now aim to recognize and adapt to emotions. These predicted emotions are used in variety of downstream applications: (a) generating more human like dialogues, (b) predicting mental health issues, and (c) hate speech detection and intervention. To enable this, data are transmitted from users' devices and stored on central servers. These data are then processed further, either annotated or used as inputs for training a model for a specific task. Yet, these data contain sensitive information that could be used by mobile applications without user's consent or, maliciously, by an eavesdropping adversary. My work focuses on two major issues that are faced while training emotion recognition algorithms: (a) privacy of the generated representations and, (b) explaining and ensuring that the predictions are robust to various situations. Tackling these issues would lead to emotion based algorithms that are deployable and helpful at a larger scale, thus enabling more human like experience when interacting with AI.
In this research work, the problem of learning abstract rules using neural networks is studied and a solution called ‘Relation Based Patterns’ (RBP) which model abstract relationships based on equality is proposed.
A multivariate time-series forecasting has great potentials in various domains. However, it is challenging to find dependency structure among the time-series variables and appropriate time-lags for each variable, which change dynamically over time. In this study, I suggest partial correlation-based attention mechanism which overcomes the shortcomings of existing pair-wise comparisons-based attention mechanisms. Moreover, I propose data-driven series-wise multi-resolution convolutional layers to represent the input time-series data for domain agnostic learning.
In my PhD project I study the algorithmic aspects of strategic behaviour in collective decision making, with the special focus on voting mechanisms. I investigate two manners of manipulation: (1) strategic selection of candidates from groups of potential representatives and (2) influence on voters located in a social network.
This thesis explores how reinforcement learning (RL) agents can provide explanations for their actions and behaviours. As humans, we build causal models to encode cause-effect relations of events and use these to explain why events happen. Taking inspiration from cognitive psychology and social science literature, I build causal explanation models and explanation dialogue models for RL agents. By mimicking human-like explanation models, these agents can provide explanations that are natural and intuitive to humans.
Automated negotiations based on learning models have been widely applied in different domains of negotiation. Specifically, for resource allocation in decentralised open market environments with multiple vendors and multiple buyers. In such open market environments, there exists dynamically changing supply and demand of resources, with dynamic arrival of buyers in the market. Besides, each buyer has their own set of constraints, such as budget constraints, time constraints, etc. In this context, efficient negotiation policies should be capable of maintaining the equilibrium between the utilities of both the vendors and the buyers. In this research, we aim to design a mechanism for an optimal auction paradigm, considering the existence of interdependent undisclosed preferences of both, buyers and vendors. Therefore, learning-based negotiation models are immensely appropriate for such open market environments; wherein, self-interested autonomous vendors and buyers cooperate/compete to maximize their utilities based on their undisclosed preferences. Toward this end, we present our current proposal, the two-stage learning-based resource allocation mechanism, wherein utilities of vendors and buyers are optimised at each stage. We are aiming to compare our proposed learning-based resource allocation mechanism with two state-of-the-art bidding-based resource allocation mechanism, which are based on, fixed bidding policy (Samimi, Teimouri, and Mukhtar 2016) and demand-based bidding policy (Kong, Zhang, and Ye 2015). The comparison is to be done based on the overall performance of the open market environment and also based on the individual performances of vendors and buyers.
My thesis research incorporates high-level abstract behavioral requirements, called ‘conceptual constraints’, into the modeling processes of robot Learning from Demonstration (LfD) techniques. My most recent work introduces an LfD algorithm called Concept Constrained Learning from Demonstration. This algorithm encodes motion planning constraints as temporal Boolean operators that enforce high-level constraints over portions of the robot's motion plan during learned skill execution. This results in more easily trained, more robust, and safer learned skills. Future work will incorporate conceptual constraints into human-aware motion planning algorithms. Additionally, my research will investigate how these concept constrained algorithms and models are best incorporated into effective interfaces for end-users.
This doctoral consortium presents an overview of my anticipated PhD dissertation which focuses on employing quantum Bayesian networks for social learning. The project, mainly, aims to expand the use of current quantum probabilistic models in human decision-making from two agents to multi-agent systems. First, I cultivate the classical Bayesian networks which are used to understand information diffusion through human interaction on online social networks (OSNs) by taking into account the relevance of multitude of social, psychological, behavioral and cognitive factors influencing the process of information transmission. Since quantum like models require quantum probability amplitudes, the complexity will be exponentially increased with increasing uncertainty in the complex system. Therefore, the research will be followed by a study on optimization of heuristics. Here, I suggest to use an belief entropy based heuristic approach. This research is an interdisciplinary research which is related with the branches of complex systems, quantum physics, network science, information theory, cognitive science and mathematics. Therefore, findings can contribute significantly to the areas related mainly with social learning behavior of people, and also to the aforementioned branches of complex systems. In addition, understanding the interactions in complex systems might be more viable via the findings of this research since probabilistic approaches are not only used for predictive purposes but also for explanatory aims.
This paper states the challenges in fine-grained target-dependent Sentiment Analysis for social media data using recurrent neural networks. First, the problem statement is outlined and an overview of related work in the area is given. Then a summary of progress and results achieved to date and a research plan and future directions of this work are given.
The study of social networks has increased rapidly in the past few decades. Of recent interest are the dynamics of changing opinions over a network. Some research has investigated how interpersonal influence can affect opinion change, how to maximize/minimize the spread of opinion change over a network, and recently, if/how agents can act strategically to effect some outcome in the network's opinion distribution. This latter problem can be modeled and addressed as a reinforcement learning problem; we introduce an approach to help network agents find strategies that outperform hand-crafted policies. Our preliminary results show that our approach is promising in networks with dynamic topologies.
The abundance of temporal data generated by mankind in recent years gives us the opportunity to better understand human behaviors along with the similarities and differences in groups of people. Better understanding of human behaviors could be very beneficial in choosing strategies, from group-level to society-level depending on the domain. This type of data could range from physiological data collected from sensors to activity patterns in social media. Identifying frequent behavioral patterns in sensor data could give more insight into the health of a community and provoke strategies towards improving it; By analyzing patterns of behaviors in social media, platform's attributes could be adjusted to the user's needs. This type of modeling introduces numerous challenges that varies depending on the data. The goal of my doctoral research is to introduce ways to better understand and capture human behavior by modeling individual's behaviors as time series and extracting interesting patterns within them.
The research presented herein addresses the topic of explainability in autonomous pedagogical agents. We will be investigating possible ways to explain the decision-making process of such pedagogical agents (which can be embodied as robots) with a focus on the effect of these explanations in concrete learning scenarios for children. The hypothesis is that the agents' explanations about their decision making will support mutual modeling and a better understanding of the learning tasks and how learners perceive them. The objective is to develop a computational model that will allow agents to express internal states and actions and adapt to the human expectations of cooperative behavior accordingly. In addition, we would like to provide a comprehensive taxonomy of both the desiderata and methods in the explainable AI research applied to children's learning scenarios.
Deep Probabilistic Models (DPM) based on arithmetic circuits representation, such as Sum-Product Networks (SPN) and Probabilistic Sentential Decision Diagrams (PSDD), have shown competitive performance in several machine learning tasks with interesting properties (Poon and Domingos 2011; Kisa et al. 2014). Due to the high number of parameters and scarce data, DPMs can produce unreliable and overconfident inference. This research aims at increasing the robustness of predictive inference with DPMs by obtaining new estimators of the predictive uncertainty. This problem is not new and the literature on deep models contains many solutions. However the probabilistic nature of DPMs offer new possibilities to achieve accurate estimates at low computational costs, but also new challenges, as the range of different types of predictions is much larger than with traditional deep models. To cope with such issues, we plan on investigating two different approaches. The first approach is to perform a global sensitivity analysis on the parameters, measuring the variability of the output to perturbations of the model weights. The second approach is to capture the variability of the prediction with respect to changes in the model architecture. Our approaches shall be evaluated on challenging tasks such as image completion, multilabel classification.
With the rising number and complexity of cancer therapies, it is increasingly difficult for clinicians to identity an optimal combination of treatments for a patient. Our research aims to provide a decision support tool to optimize and supplant cancer treatment decisions. Leveraging machine learning, causal inference, and decision analysis, we will utilize electronic medical records to develop dynamic cancer treatment strategies that advice clinicians and patients based on patient characteristics, medical history, and etc. The research hopes to bridge the understanding between causal inference and decision analysis and ultimately develops an artificial intelligence tool that improves clinical outcomes over current practices.