Hello, I am Aziz.

Psychology graduate studying cognition, behavior, and mental health.

I am a McGill Psychology graduate interested in how cognitive processes, brain activity, and individual differences relate to mental health. My work uses computational modeling, behavioral data, fMRI, and neural-network methods as tools for asking psychological and clinical questions. I am especially interested in research that connects careful measurement with clinically meaningful variation in cognition, physiology, and pathology.

Get in touch View research

Research Interests

My research interests grew out of philosophy, especially questions about how mind, body, brain, and behavior relate to one another.

My interest in psychology began with philosophy. I was especially drawn to the mind-body problem, which I explored in a CEGEP philosophy mémoire on the soul-body problem. In that work, I contrasted two major positions: the dualist view, where body and soul are distinct, and the materialist view, where all is grounded in matter. I then developed my own reasoning on the question and ended by opening toward a possible theory of where the soul could be situated.

At university, those questions opened into new forms of research. I became interested in psychology, cognition, and mental health, and in empirical ways of studying how subjective experience, thought, emotion, and behavior relate to the brain and to the biological systems that support them.

In research, I am excited by methods that let us connect behavior, brain activity, and computational models. NeuroAI made this especially concrete for me: training neural networks on cognitive tasks, then asking whether the way those models solve problems resembles the way humans solve them. Comparing model representations with patterns of brain activity feels like a powerful way to study cognition, because it gives us another object to inspect, test, and question.

I am also interested in the broader possibilities that AI and computational modeling are opening in health and clinical psychology. These tools can help researchers ask questions about individual differences, treatment response, brain structure, physiological markers, and patient-specific risk. I like that this space is not limited to one method or one domain; it sits between psychology, neuroscience, medicine, statistics, and computer science.

At the core, I enjoy discovering things and building things. I like coding, making models, looking at brain plots, and working through the practical hurdles that turn an idea into a real research project. The part that keeps pulling me in is the combination of curiosity and construction: using computational tools not just because they are powerful, but because they can help us ask better questions about the mind, the brain, and mental health.

Clinical psychology Philosophy of mind Cognitive neuroscience NeuroAI Computational modeling

Publications

Published and accepted work.

Zero-shot neural predictivity in human prefrontal cortex with a massively multi-task multimodal transformer

Published

Lucas Gomez, Aziz Ktari, Hao Yuan, Bai Pouya Bashivan. Data on the Brain and Mind, NeurIPS Workshop, 2025. Link

Abstract

Working memory supports a broad spectrum of behaviors and higher cognitive abilities, with the prefrontal cortex playing a central role in this capacity. Although prior work has identified which brain regions are engaged in specific working memory tasks, and in some cases how they contribute, we still lack a general framework that can predict which regions will be recruited in novel tasks, what information they represent, and the computations they perform. To address this gap, we trained a single neural network on millions of visual decision-making tasks with sensory-realistic inputs, aiming to build a generalized model of working memory. We evaluated the model against an fMRI dataset spanning 12 tasks and hundreds of distinct conditions, testing its ability to capture neural activity across the brain, with a focus on the prefrontal cortex. Our results show that large models trained on a broad distribution of tasks can predict brain activity zero-shot, outperforming even models trained directly on the target tasks. This ability improves further with model size, which consistently enhances prediction accuracy. Furthermore, analyses of layer-to-region correspondences largely conformed with the theories of hierarchical organization along the rostro-caudal axis of the prefrontal cortex. These findings suggest that neural network models hold significant potential not only for simulating neural activity in regions previously difficult to model, but also for revealing how the brain encodes, organizes, and manipulates task information during working memory.

Conference Presentations

Poster presentations and accepted abstracts.

A Computational Pipeline for Identifiability and Parameter Recoverability in Cognitive Models

Presented

Aziz Ktari, Yanan Liu, Emir Sahin, Paul Masset. Canadian Association for Neuroscience, 2026.

Abstract

Computational models are widely used to explain behavior and link latent processes to neural measures. Two standard safeguards are simulation and parameter recovery, which extend classic power analysis to complex models that lack analytical solutions, revealing misspecified models, coding errors, and non-identifiable parameters before real data are collected. In practice, however, these steps are often implemented with ad hoc scripts, making results inconsistent. Thus, we take a bootstrap approach and present a modular diagnostic pipeline, with a Python package coming soon, that standardizes simulation and recovery analyses and provides actionable checks for model validity and recoverability. The pipeline helps researchers evaluate whether a model can be reliably fit under realistic noise and data constraints, and how experimental design choices affect identifiability. Specifically, it characterizes error as a function of trials and fixed budget constraints with varying participant counts, and summarizes parameter interactions through correlation matrices and derived summaries. To make non-identifiability interpretable, it includes richer pairwise diagnostics such as density maps and bias vector fields in two-dimensional parameter subspaces. In a Kalman-bandit case study, the diagnostics reveal parameter non-identifiability and budget-dependent recoverability. Together, these outputs help researchers pinpoint why recovery fails and choose designs that make targeted mechanisms reliably identifiable before collecting data.

Zero-shot neural predictivity in human prefrontal cortex with a massively multi-task multimodal transformer

Presented

Lucas Gomez, Aziz Ktari, Hao Yuan, Bai Pouya Bashivan. Presented at Data on the Brain and Mind, NeurIPS Workshop, 2025, and COSYNE, 2026.

Awards

Recognition and funded research experience.

Mackey & Glass Summer Research Award, Department of Physiology

2025

Research on ways to improve multiple sequence alignment tools using neural-network methods.

Work in Progress

Active research and software directions across cognition, behavior, and computational biology.

Working-Memory and fMRI

Studying how the prefrontal cortex supports working-memory and decision-making tasks. The project trains neural networks on cognitive tasks, then compares their activations with human brain activity. When model layers resemble activity in specific PFC regions, the model can become a tool for testing theories of hierarchical organization and for asking what computations different parts of the PFC may be performing.

Working memory fMRI PFC organization

Behavioral Modeling Toolkit

A computational pipeline for diagnosing behavioral models before collecting new data, inspired by Wilson and Collins' "Ten simple rules for the computational modeling of behavioral data" (2019). The goal is to help researchers test whether model parameters are identifiable, recoverable from simulated data, and meaningful enough to support interpretation in real experiments.

Parameter recovery Model diagnosis

Neural MSA

Improving multiple sequence alignment methods with neural networks. Classical MSA tools often rely on heuristics and iterative procedures; this project explores whether neural networks can learn broader structural patterns across DNA sequences and approach the alignment problem more globally.

DNA alignment Neural networks

Contact

Open to collaborations, mentorship discussions, and research conversations in clinical psychology, cognition, and computational modeling.

Email

aziz.ktari@mail.mcgill.ca

Montreal, Canada

Collaboration

I am open to collaborations on clinical and cognitive research that uses computational methods to study behavior, physiology, brain activity, and patient-specific differences.