Mohammad Taha Fakharian

Okinawa Institute of Science
and Technology (OIST)

1919-1 Tancha, Onna-son
Okinawa 904-0495, Japan

I’m a Ph.D. student at OIST, in the Machine Learning and Data Science Unit under Prof. Makoto Yamada, co-supervised by Prof. Kenji Doya of the Neural Computation Unit.

I want to understand how a system that only ever sees sensory data ends up with knowledge that transfers — how brains do it, and why our models so often don’t. That question keeps pulling me back to neural representations: the geometry a network settles into, and what that geometry commits the system to once the data runs out.

Lately I’ve been circling a more specific version of it. Curiosity-driven world models treat surprise as the learning signal, and the free-energy framing gives that signal a clean objective — but surprise and learnable surprise are not the same quantity, and conflating them is what strands an agent in front of a noisy screen or in a dark room. Two recent ideas sharpen the distinction: epiplexity, which asks what a computationally bounded observer can actually extract from data, and learnable novelty, which isolates the fraction of novelty a learner can convert into knowledge. Set beside Wilson’s argument that generalization is governed by soft inductive biases rather than hard constraints, the question I’d like to answer is: given a learnable-novelty signal and a particular soft bias, which world model does the agent end up with? Not whether it learns, but what it learns.

Before OIST I did three rotations here — spiking basal-ganglia models of dopamine and temporal-difference learning with Prof. Doya, short-term plasticity and temporal associative memory with Prof. Tomoki Fukai, and an evolutionary-developmental extension of curiosity-driven robot learning with Prof. Jun Tani. Before that I worked with Prof. Shervin Safavi on how RNN architecture shapes internal dynamics, wrote a B.Sc. thesis on delay learning in spiking networks with Prof. Mohammadreza Abolghasemi and Prof. Timothée Masquelier, and spent a year as a data scientist at Tapsi helping start its data-science team.

Away from the desk: football, far too much of it. Also bands and games, and currently a losing battle with Genki I.

news

Sep 01, 2026	Began thesis research in the Machine Learning and Data Science Unit, on learnable-novelty signals and soft inductive biases in world models.
Aug 31, 2026	Finished my third and final rotation, with Prof. Jun Tani — an evolutionary-developmental take on curiosity-driven robot learning. Three rotations down; thesis work begins.
Mar 09, 2026	Presented my spiking basal-ganglia model of dopamine-modulated plasticity at the Winter Workshop 2026 on Mechanism of Brain and Mind, Rusutsu, Hokkaido.
Nov 28, 2025	Fast Dynamical Similarity Analysis is on arXiv — a metric for comparing nonlinear dynamical systems that stays faithful to the dynamics without the usual compute bill.
Sep 01, 2025	Started my Ph.D. at OIST in Okinawa.

latest posts

Sep 16, 2026	Surprise you can use

selected publications

preprint
Fast Dynamical Similarity Analysis

Arman Behrad, Mitchell Ostrow, Mohammad Taha Fakharian, and 3 more authors

arXiv preprint arXiv:2511.22828, Nov 2025

Abs arXiv Bib HTML Code

Understanding how nonlinear dynamical systems (e.g., artificial neural networks and neural circuits) process information requires comparing their underlying dynamics at scale, across diverse architectures and large neural recordings. While many similarity metrics exist, current approaches fall short for large-scale comparisons. Geometric methods are computationally efficient but fail to capture governing dynamics, limiting their accuracy. In contrast, traditional dynamical similarity methods are faithful to system dynamics but are often computationally prohibitive. We bridge this gap by combining the efficiency of geometric approaches with the fidelity of dynamical methods. We introduce fast dynamical similarity analysis (fastDSA), a computationally efficient and accurate metric for measuring (dis)similarity between nonlinear dynamical systems. FastDSA leverages modern computational tools, including random matrix theory to determine optimal system rank, novel optimization pipelines for aligning system flow fields, and Koopman embeddings. Across benchmark nonlinear systems and recurrent network models, fastDSA is robust to arbitrary coordinate choices while remaining sensitive to meaningful dynamical differences, capturing variations in system evolution that geometric methods may miss and traditional methods detect only at high computational cost. To our knowledge, fastDSA is the fastest method that retains accuracy in comparing nonlinear dynamical systems. It enables scalable, statistical analyses across diverse systems, significantly expanding the practical applicability of dynamical similarity analysis.
@article{behrad2025fastdsa, title = {Fast Dynamical Similarity Analysis}, author = {Behrad, Arman and Ostrow, Mitchell and Fakharian, Mohammad Taha and Fiete, Ila and Beste, Christian and Safavi, Shervin}, journal = {arXiv preprint arXiv:2511.22828}, year = {2025}, month = nov, url = {https://arxiv.org/abs/2511.22828}, }
CCN
Heterogeneous Effect of Input and Task-optimization on the Dynamics of RNNs

Mohammad Taha Fakharian^*, Alireza Ghalambor^*, Arman Behrad, and 2 more authors

In Cognitive Computational Neuroscience (CCN), 2025

Travel Award Abs Bib HTML

true

Reverse-engineering task-optimized recurrent neural networks (RNNs) has become a key framework in cognitive tasks to uncover mechanisms of brain computation. Tasks are often constructed as a set of inputs. Then, RNNs are optimized to achieve a set of computational sub-goals given the inputs. Then, neural dynamics in RNNs can be shaped by two major factors: the effect of input structure (defined by task) and task-based optimization or training. The former better reflects the attributes of the input to the network, while the latter better reflects the connectivity that is shaped by task-based optimization. Although both are major factors shaping the network dynamics in a task-specific fashion, how exactly these factors affect network dynamics remains elusive. Here, we investigate the effect of both factors on discriminating the neural dynamics across tasks. We systematically vary the network architecture and the input conditions, using three distinct recurrent architectures: Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and vanilla RNN (V-RNN), trained on cognitive tasks from the NeuroGym library. While we observed a vast range of heterogeneity across architectures and choices of task on task-specific dynamics, we observed that task structure (rather than task-based optimization of the connectivity) almost dominantly informs about task-specific dynamics.
@inproceedings{fakharian2025heterogeneous, title = {Heterogeneous Effect of Input and Task-optimization on the Dynamics of RNNs}, author = {Fakharian, Mohammad Taha and Ghalambor, Alireza and Behrad, Arman and Zeraati, Roxana and Safavi, Shervin}, booktitle = {Cognitive Computational Neuroscience (CCN)}, address = {Amsterdam, Netherlands}, year = {2025}, url = {https://2025.ccneuro.org/abstract_pdf/Fakharian_2025_Heterogeneous_Effect_Input_Task-optimization_Dynamics_Recurrent.pdf}, }