Grants

Reading every brain: applying data-efficient natural image reconstruction approaches to advance cognitive neuroscience

Reconstructing naturalistic visual images from fMRI data presents a challenging task, particularly when dealing with limited data and compute availability. Existing approaches require dozens of hours of 7T fMRI data per participant and extensive compute capability, each of which render these methods inaccessible to traditional cognitive neuroscience labs. This work develops and benchmarks a series of novel approaches to fMRI-based visual image reconstruction using a subject-agnostic common representation space, with the goal of establishing these methods as a tractable tool for 3T cognitive neuroscience studies acquiring data from robust samples of human participants performing cognitive tasks. In preliminary work, we have established that subjects’ brain signals naturally align in this common space during training. In the proposed work, we will leverage this shared representational space to demonstrate that aligning subject-specific adapters to a reference subject is significantly more efficient than traditional end-to-end training methods. We will additionally incorporate aspects of modern cognitive and visual neuroscience (e.g., retinotopic mapping/population receptive field modeling; principled voxel selection) to further optimize the image reconstruction approach, and test these optimizations on new 3T validation data. Finally, we will put this approach to the test by acquiring fMRI data while participants perform cognitive tasks (visual attention and visual working memory) to establish the utility of natural image reconstruction for discriminating between cognitive hypotheses. Upon completion, this work will make fMRI data collection for natural image reconstruction more efficient and practical, reducing the burden on subjects and improving the generalization of fMRI reconstruction models for basic science, diagnostic, and therapeutic endeavors.

AI Institute for Agent-based Cyber Threat Intelligence and Operation

Computer systems are increasingly central to national infrastructure in the financial, medical, manufacturing, defense, and other domains. This infrastructure is at risk from sophisticated cyber-adversaries backed by powerful nation-states, whose capabilities rapidly evolve, demanding equally rapid responses. This calls for advances in artificial intelligence and autonomous reasoning that are tightly integrated with advanced security techniques to identify and correct vulnerabilities, detect threats and attribute them to adversaries, and mitigate and recover from attacks. The ACTION Institute will develop novel approaches that leverage artificial intelligence?informed by and working with experts in security operations?to perform security tasks rapidly and at scale, anticipating the moves of an adversary and taking corrective actions to protect the security of computer networks as well as people?s safety. The Institute will function as a nexus for the AI and cybersecurity communities, and its research efforts will be complemented by innovation in education from K-12 to postdoctoral students, the development of new tools for workforce development, and the creation of new opportunities for collaboration among the Institute?s organizations and with external industry partners.

The AI Institute will initiate a revolutionary approach to cybersecurity, in which AI-enabled intelligent security agents cooperate with humans across the cyber-defense life cycle to jointly improve the security posture of complex computer systems over time. Intelligent security agents will follow a new paradigm of continuous, lifelong learning both autonomously and in collaboration with human experts, supported by a shared knowledge bank and an integrated AI stack that provides novel fundamental primitives for (1) reasoning and learning that incorporates domain knowledge, (2) human-agent interaction, (3) multi-agent collaboration, and (4) strategic gaming and tactical planning. Over time, these intelligent security agents will improve their domain knowledge, becoming increasingly robust and effective in the face of changes in the adversaries? modes of operation, composing defense strategies and tactical plans in the presence of uncertainty, collaborating with each other and with humans for mutually complementary teaming, and adapting to unfamiliar and novel attacks.

More details at https://action.ucsb.edu/.

Cognitive Models and Strategies for High-Performance Human-AI Teams

A key research challenge in human-AI systems is the current limitation in data, models and theories that explain their dynamic behavior, coordination, and performance. We do not yet fully understand the dominant socio-cognitive processes that determine the dynamic, adaptive, and learning behavior of human-AI teams. Of especial interest are decision-making problems in intellective tasks with uncertainty and limited resources: what are the rational, efficient, or irrational strategies and heuristics that humans tend to adopt in such circumstances? Useful socio-cognitive models should inform the design of efficient AI agents that improve the overall human-AI team performance. In other words, empirically-validated models and theories are needed to model and build the human-AI teams of the future and to intervene when their performance deteriorates. The project’s broad objective is the development and experimental validation of a theory of coordination of human-AI teams in complex intellective tasks. We plan to combine fundamental insights and models of team behavior from social sciences with state-of-the-art machine learning and dynamical systems methods. Specifically, our objective include: 1. modeling socio-cognitive structures in human-AI teams, including transactive memory systems, influence systems, and prospect theories; 2. identifying leading cognitive processes, heuristics and biases that underlie the formation of socio-cognitive structures and affect the accuracy of human-AI team decision making; 3. designing supervisory/coordinating AI agents in human-AI teams, based on concepts from applied psychology and machine learning, and testing/validating them in sequential, risky, uncertain decision making tasks; and 4. modeling how human-AI teams cope with limited training data acquired over short sessions, including how they react to various manipulations and intervention schemes.

Funded by: Army Research Office.

Inferring Network Structure and Flow using Partial Observations

Predicting network structure and flows on edges is a precondition to effective planning and disaster response in critical infrastructure networks. Accordingly, we propose to develop algorithms inspired by network science, physical modeling, and machine learning to determine this information from incomplete observations. Using both physical domain-specific and data-driven approaches, these algorithms will address diverse problems including: the reconstruction of network topology and parameters, the estimation of network flows, the optimal selection and scheduling of multiple types of sensors, and the modeling of interdependencies in multilayer networks. Together these methods will develop a rigorous theory for solving the inverse problem of edge and flow determination from observations of critical infrastructure networks.

Funded by: Defense Threat Reduction Agency, Award# HDTRA1-19-1-0017