Retinotopic map, the map between visual inputs on the retina and neuronal activation in brain visual areas, is one of the central topics in visual neuroscience. For human observers, the map is typically obtained by analyzing functional magnetic resonance imaging (fMRI) signals of cortical responses to slowly moving visual stimuli on the retina. Biological evidences show the retinotopic mapping is topology-preserving/topological (i.e. keep the neighboring relationship after human brain process) within each visual region. Unfortunately, due to limited spatial resolution and the signal-noise ratio of fMRI, state of art retinotopic map is not topological. The topic was to model the topology-preserving condition mathematically, fix non-topological retinotopic map with numerical methods, and improve the quality of retinotopic maps. The impose of topological condition, benefits several applications. With the topological retinotopic maps, one may have a better insight on human retinotopic maps, including better cortical magnification factor quantification, more precise description of retinotopic maps, and potentially better exam ways of in Ophthalmology clinic.

Federated Learning (FL) is envisaged to be a promising solution for collaboratively training a machine learning model while keeping the training data decentralized and private. Instead of sharing raw data to the central entity, the participating client devices share focused updates for aggregation to ensure global convergence of the model. Owing to the shortcomings of manually handcrafted neural network architectures, the research community is striving to develop Neural Architecture Search (NAS) approaches to automatically search for optimal networks that fit the clients’ data. Despite the inaccessibility of clients’ data in an FL setting, the federated NAS literature has recently witnessed great progress to apply these NAS techniques to an FL setting. However, one of the key bottlenecks of Federated Learning is the cost of communication between clients and the server, and the state-of-the-art federated NAS techniques search for networks with millions of parameters that require several rounds of communication to find the optimal weight parameters. Also, deploying a network having millions of parameters on edge devices (which are the typical participants in an FL process) is infeasible due to its computational limitations and increased latency. Thus, this work proposes Weight-Agnostic Federated Neural Architecture Search (WFNAS), a novel evolutionary framework to search for well-performing and minimally connected weight-agnostic network architectures in an FL setting. As the connectivity of the networks themselves is the solution, there is no need for weight training and hyperparameter tuning, reducing the communication overhead significantly. The experiments indicate a gain of nearly 40% for orthogonal (vertical FL) data distributions compared to local training. This work is the first federated NAS technique in the literature for vertical FL. Although the experiments are performed in a resource-constrained environment, the aim of this thesis is to show a new direction of research to the FL community.

Natural Language plays a crucial role in human-robot interaction as it is the common ground where human beings and robots can communicate and understand each other. However, most of the work in natural language and robotics is majorly on generating robot actions using a natural language command, which is a unidirectional way of communication. This work focuses on the other direction of communication, where the approach allows a robot to describe its actions from sampled images and joint sequences from the robot task. The importance of this work is that it utilizes multiple modalities, which are the start and end images from the robot task environment and the joint trajectories of the robot arms. The fusion of different modalities is not just about fusing the data but knowing what information to extract from which data sources in such a way that the language description represents the state of the manipulator and the environment that it is performing the task on. From the experimental results of various simulated robot environments, this research demonstrates that utilizing multiple modalities improves the accuracy of the natural language description, and efficiently fusing the modalities is crucial in generating such descriptions by harnessing most of the various data sources.

Traditional Reinforcement Learning (RL) assumes to learn policies with respect to reward available from the environment but sometimes learning in a complex domain requires wisdom which comes from a wide range of experience. In behavior based robotics, it is observed that a complex behavior can be described by a combination of simpler behaviors. It is tempting to apply similar idea such that simpler behaviors can be combined in a meaningful way to tailor the complex combination. Such an approach would enable faster learning and modular design of behaviors. Complex behaviors can be combined with other behaviors to create even more advanced behaviors resulting in a rich set of possibilities. Similar to RL, combined behavior can keep evolving by interacting with the environment. The requirement of this method is to specify a reasonable set of simple behaviors. In this research, I present an algorithm that aims at combining behavior such that the resulting behavior has characteristics of each individual behavior. This approach has been inspired by behavior based robotics, such as the subsumption architecture and motor schema-based design. The combination algorithm outputs n weights to combine behaviors linearly. The weights are state dependent and change dynamically at every step in an episode. This idea is tested on discrete and continuous environments like OpenAI’s “Lunar Lander” and “Biped Walker”. Results are compared with related domains like Multi-objective RL, Hierarchical RL, Transfer learning, and basic RL. It is observed that the combination of behaviors is a novel way of learning which helps the agent achieve required characteristics. A combination is learned for a given state and so the agent is able to learn faster in an efficient manner compared to other similar approaches. Agent beautifully demonstrates characteristics of multiple behaviors which helps the agent to learn and adapt to the environment. Future directions are also suggested as possible extensions to this research.

Understanding the limits and capabilities of an AI system is essential for safe and effective usability of modern AI systems. In the query-based AI assessment paradigm, a personalized assessment module queries a black-box AI system on behalf of a user and returns a user-interpretable model of the AI system’s capabilities. This thesis develops this paradigm to learn interpretable action models of simulator-based agents. Two types of agents are considered: the first uses high-level actions where the user’s vocabulary captures the simulator state perfectly, and the second operates on low-level actions where the user’s vocabulary captures only an abstraction of the simulator state. Methods are developed to interface the assessment module with these agents. Empirical results show that this method is capable of learning interpretable models of agents operating in a range of domains.

In classification applications, such as medical disease diagnosis, the cost of one type of error (false negative) could greatly outweigh the other (false positive) enabling the need of asymmetric error control. Due to this unique nature of the problem, traditional machine learning techniques, even with much improved accuracy, may not be ideal as they do not provide a way to control the false negatives below a certain threshold. To address this need, a classification algorithm that can provide asymmetric error control is proposed. The theoretical foundation for this algorithm is based on Neyman-Pearson (NP) Lemma and it is complemented with sample splitting and order statistics to pick a threshold that enables an upper bound on the number of false negatives. Additionally, this classifier addresses the imbalance of the data, which is common in medical datasets, by using Hellinger distance as the splitting criterion. This eliminates the need of sampling methods, which add complexity and the need for parameter selection. This approach is used to create a novel tree-based classifier that enables asymmetric error control. Applications, such as prediction of the severity of cardiac arrhythmia, require classification over multiple classes. The NP oracle inequalities for binary classes are not immediately applicable for the multiclass NP classification, leading to a multi-step procedure proposed in this dissertation to extend the algorithm in the context of multiple classes. This classifier is used in predicting various forms of cardiac disease for both binary and multi-class classification problems with not only comparable accuracy metrics but also with full control over the number of false negatives. Moreover, this research allows us to pick the threshold for the classifier in a data adaptive way. This dissertation also shows that this methodology can be extended to non medical applications that require classification with asymmetric error control.

In settings where a human and an embodied AI (artificially intelligent) agent coexist, the AI agent has to be capable of reasoning with the human's preconceived notions about the environment as well as with the human's perception limitations. In addition, it should be capable of communicating intentions and objectives effectively to the human-in-the-loop. While acting in the presence of human observers, the AI agent can synthesize interpretable behaviors like explicable, legible, and assistive behaviors by accounting for the human's mental model (inclusive of her sensor model) in its reasoning process. This thesis will study different behavior synthesis algorithms which focus on improving the interpretability of the agent's behavior in the presence of a human observer. Further, this thesis will study how environment redesign strategies can be leveraged to improve the overall interpretability of the agent's behavior. At times, the agent's environment may also consist of purely adversarial entities or mixed entities (i.e. adversarial as well as cooperative entities), that are trying to infer information from the AI agent's behavior. In such settings, it is crucial for the agent to exhibit obfuscatory behavior that prevents sensitive information from falling into the hands of the adversarial entities. This thesis will show that it is possible to synthesize interpretable as well as obfuscatory behaviors using a single underlying algorithmic framework.

Most planning agents assume complete knowledge of the domain, which may not be the case in scenarios where certain domain knowledge is missing. This problem could be due to design flaws or arise from domain ramifications or qualifications. In such cases, planning algorithms could produce highly undesirable behaviors. Planning with incomplete domain knowledge is more challenging than partial observability in the sense that the planning agent is unaware of the existence of such knowledge, in contrast to it being just unobservable or partially observable. That is the difference between known unknowns and unknown unknowns.

In this thesis, I introduce and formulate this as the problem of Domain Concretization, which is inverse to domain abstraction studied extensively before. Furthermore, I present a solution that starts from the incomplete domain model provided to the agent by the designer and uses teacher traces from human users to determine the candidate model set under a minimalistic model assumption. A robust plan is then generated for the maximum probability of success under the set of candidate models. In addition to a standard search formulation in the model-space, I propose a sample-based search method and also an online version of it to improve search time. The solution presented has been evaluated on various International Planning Competition domains where incompleteness was introduced by deleting certain predicates from the complete domain model. The solution is also tested in a robot simulation domain to illustrate its effectiveness in handling incomplete domain knowledge. The results show that the plan generated by the algorithm increases the plan success rate without impacting action cost too much.

Many real-world planning problems can be modeled as Markov Decision Processes (MDPs) which provide a framework for handling uncertainty in outcomes of action executions. A solution to such a planning problem is a policy that handles possible contingencies that could arise during execution. MDP solvers typically construct policies for a problem instance without re-using information from previously solved instances. Research in generalized planning has demonstrated the utility of constructing algorithm-like plans that reuse such information. However, using such techniques in an MDP setting has not been adequately explored.

This thesis presents a novel approach for learning generalized partial policies that can be used to solve problems with different object names and/or object quantities using very few example policies for learning. This approach uses abstraction for state representation, which allows the identification of patterns in solutions such as loops that are agnostic to problem-specific properties. This thesis also presents some theoretical results related to the uniqueness and succinctness of the policies computed using such a representation. The presented algorithm can be used as fast, yet greedy and incomplete method for policy computation while falling back to a complete policy search algorithm when needed. Extensive empirical evaluation on discrete MDP benchmarks shows that this approach generalizes effectively and is often able to solve problems much faster than existing state-of-art discrete MDP solvers. Finally, the practical applicability of this approach is demonstrated by incorporating it in an anytime stochastic task and motion planning framework to successfully construct free-standing tower structures using Keva planks.

Robot motion planning requires computing a sequence of waypoints from an initial configuration of the robot to the goal configuration. Solving a motion planning problem optimally is proven to be NP-Complete. Sampling-based motion planners efficiently compute an approximation of the optimal solution. They sample the configuration space uniformly and hence fail to sample regions of the environment that have narrow passages or pinch points. These critical regions are analogous to landmarks from planning literature as the robot is required to pass through them to reach the goal.

This work proposes a deep learning approach that identifies critical regions in the environment and learns a sampling distribution to effectively sample them in high dimensional configuration spaces.

A classification-based approach is used to learn the distributions. The robot degrees of freedom (DOF) limits are binned and a distribution is generated from sampling motion plan solutions. Conditional information like goal configuration and robot location encoded in the network inputs showcase the network learning to bias the identified critical regions towards the goal configuration. Empirical evaluations are performed against the state of the art sampling-based motion planners on a variety of tasks requiring the robot to pass through critical regions. An empirical analysis of robotic systems with three to eight degrees of freedom indicates that this approach effectively improves planning performance.