580 California St., Suite 400
San Francisco, CA, 94104
Key research themes
1. How can credit assignment be optimized in multi-agent systems and neural networks for effective resource allocation and learning?
This research theme explores algorithmic and computational frameworks for credit assignment tailored to multi-agent settings and neural networks, focusing on how credit (or reward) attribution to individual agents or neurons can be optimized to enhance system efficiency and learning performance. It addresses challenges such as distributing global rewards among multiple agents, biologically plausible credit assignment in deep networks, and credit assignment under constraints like task start thresholds and multi-score rewards, relevant to complex environments such as smart cities and artificial intelligence systems.
Key finding: This paper proposes a novel solution to resource allocation in smart cities by formulating the problem as a multi-agent credit assignment (MCA) issue mapped to a bankruptcy game, introducing a task start threshold (TST)... Read more
Key finding: The authors introduce a self-adaptive mechanism leveraging a dynamic extension of the multi-armed bandit framework to guide the allocation of computational resources among different species in a cooperative coevolutionary... Read more
Key finding: The study decomposes traditional Backpropagation into interacting local learning agents and identifies that error signals in deep networks factorize into a global scalar error and a complex, biologically implausible... Read more
2. What mathematical and computational frameworks can improve credit scoring and optimal credit allocation in financial systems?
This theme investigates advanced modeling approaches and mathematical frameworks to enhance credit risk assessment and credit allocation decisions in financial contexts. It examines how to move beyond traditional dichotomous credit scoring classification into optimal credit allocation methodologies maximizing financial returns, integrating machine learning techniques for credit scoring with explainability, and minimizing total costs in credit decisions through staged evaluation models. The insights contribute to bridging the gap between predictive accuracy and practical credit allocation, regulatory requirements, and explainability in risk management.
Key finding: This paper introduces a novel conceptual framework connecting probability of default (PD) to optimal credit allocation through the Kelly criterion, shifting focus from binary default classification to maximizing risk-adjusted... Read more
Key finding: This work presents a unified mathematical framework combining various machine learning algorithms (logistic regression, decision trees, SVM, neural networks) with advanced optimization techniques (Particle Swarm Optimization,... Read more
Key finding: The paper develops a two-stage credit scoring model minimizing the total cost associated with granting credit, including costs of defaults, denials of good applicants, and information acquisition. The first stage classifies... Read more
3. How do neural mechanisms and learning models address the credit assignment problem in biological and artificial learning systems?
This theme focuses on the neurobiological and computational foundations of credit assignment in the brain and artificial networks, probing how neurons encode and link outcomes to preceding causal factors in learning. It includes studies of prefrontal cortex representations that support credit assignment over time, the modulation of synaptic plasticity by neuromodulators and glial cells in spike-timing dependent plasticity (STDP), and cerebellar learning models that implement stochastic gradient descent via complex spikes as perturbations. These insights elucidate mechanisms underlying associative learning, reinforcement, and maladaptive behaviors, and inform biologically inspired algorithms for credit assignment.
Key finding: Using electrophysiological recordings from the dorsolateral prefrontal cortex (dlPFC) of rhesus macaques engaged in a task requiring credit assignment, this study found that dlPFC neurons maintain stable representations of... Read more
Key finding: This review consolidates evidence that spike-timing dependent plasticity (STDP), traditionally viewed as a two-factor Hebbian mechanism dependent on pre- and postsynaptic activity timing, is modulated by a third factor... Read more
Key finding: The paper proposes a novel cerebellar learning algorithm termed stochastic gradient descent with estimated global errors (SGDEGE), wherein spontaneous complex spikes perturb ongoing movements, create eligibility traces, and... Read more
Key finding: This study models obsessive-compulsive disorder (OCD) symptoms as maladaptive implicit behaviors arising from an imbalance in memory trace decay timescales for positive and negative prediction errors within reinforcement... Read more