Alex Petrov's Research Projects
My research combines behavioral experimentation with computational and mathematical modeling of cognition, paying close attention to neurological constraints. I am particularly interested in learning, adaptation, and the dynamic aspects of perception, memory, categorization, reasoning, and executive control, as well as their neural implementation. I often use state-of-the-art psychophysical methods to track the dynamics of cognition at time scales from tens of milliseconds to several days. I favor decentralized and interactive models -- connectionist, symbolic, or mathematical -- that incorporate biologically plausible learning mechanisms and are informed by an integrated cognitive architecture. Coherent global behavior emerges from local interactions in a steady state that changes adaptively whenever the environment changes.
We live in very exciting times -- for the first time in human history a consistent and empirically grounded picture of how the brain creates the mind begins to emerge. Cognitive science today seems at the threshold of a breakthrough comparable to that of genetics in the 1950s. This is an adventure of a lifetime and I am very enthusiastic to participate.
Perceptual Learning via Hebbian Reweighting
This interdiciplinary project combines experimentation and modeling to study the neural mechanisms of perceptual learning. A series of psychophysical studies probe the dynamics of perceptual learning in non-stationary environments, both with and without feedback. Practicing an orientation-discrimination task gradually improves performance but there are significant switch costs (interference) whenever the context surrounding the target stimuli changes. In a recent Psychological Review article (Petrov, Dosher, & Lu, 2005) we provide an existence proof that incremental Hebbian reweighting can account quantitatively for the complex pattern of learning and switch-costs in our non-stationary training protocol. The model takes grayscale images as inputs, produces binary responses as outputs, and improves its discrimination accuracy incrementally with practice with no need for external feedback. The model performance is thus directly comparable to the human data. Learning occurs only in the read-out connections to a decision unit in a neural network; the stimulus representations never change.
Project page Data sets Software
ANCHOR: A Memory-Based Scaling Model
This interdiciplinary project combines experimentation and modeling at the intersection of psychophysics and memory. A series of studies revealed that human response distributions are markedly non-stationary and non-uniform even when the stimulus distributions are stationary and uniform. Moreover, skewed stimulus distributions induce context effects in opposite directions depending on the presence or absence of feedback. In a recent Psychological Review article (Petrov & Anderson, 2005) we demonstrated that a memory-based model accounts naturally and quantitatively for these and many other dynamic effects in category rating and absolute identification. The ANCHOR model maintains a set of adjustable anchors that compete to match the perceived magnitude of the target stimulus. An explicit correction strategy determines the final response. Two incremental learning mechanisms track the statistics of the stimulus distribution and make the system very adaptive but also prone to sequential, context, transfer, and priming effects. The rating scale unfolds as an adaptive map from a single arbitrarily placed anchor with no need for external feedback.
Project page Data sets Software
AMBR: Analogy Making with Decentralized Representations
My Ph.D. dissertation involved a model of analogy making called AMBR. It explores the fundamental cognitive ability of interpreting a novel situation in terms of a similar situation that is already familiar. The model is based on the hybrid symbolic/connectionist architecture DUAL and employs its dynamic emergent style of computation to account for the flexible and context-sensitive nature of human analogy making. The decentralized representations of episodes support gradual and reconstructive memory retrieval that is inextricably intertwined with the process of analogical mapping (Kokinov & Petrov, 2001).
