Ben Fulcher
@bendfulcher
I lead the Dynamics and Neural Systems Group at the School of Physics, the University of Sydney. We develop time series tools & physical models to understand the dynamics of complex (usually neural) systems. Also: @bendfulcher@fediscience.org
pyhctsa is out! A native Python port of the majority of the hctsa feature library (highly comparative time-series analysis). Thousands of interpretable time-series features available via a pip install. Built by Joshua Moore. Paper: doi.org/10.21105/jos... Code: github.com/DynamicsAndN...
Kieran explains how all the methods can be understood through these conceptual groupings, derives new relationships between existing methods, and provides some case-study demonstrations/comparisons of how insanely well they can work on data
These powerful methods are underappreciated: A recent review of the field included *0* methods designed for time-series data, instead focusing on generic dimension reduction methods. This paper assembles a diversity of >60 scientific methods for the first time, and unifies them across 7 categories.
The breath of comparison (of both methods and processes) also allowed us to demonstrate that all tested indices of irreversibility had weaknesses: i.e., we could always find an irreversibile process on which any given irreverisbility index will fail to detect irreversibility.
We found key families of algorithmic constructions that were could accurately index irreversibility: (i) generalized autocorrelation functions; (ii) symbolic sequences; and (iii) forecasting-derived metrics. Some recapitulate concepts studied previously but in isolation; others are novel directions
He we compared >6000 time-series metrics to index time reversibility from simulations of 35 different reversible and irreversible processes
New preprint: "Identifying statistical indicators of temporal asymmetry using a data-driven approach" arxiv.org/abs/2511.15991 _Can we statistically distinguish the forward- versus reverse-time dynamics of a system from a finite time series?_
New paper! We introduce an efficient set of statistical features for fMRI time series (calibrated on mouse manipulation experiments and tested on mouse and human data): catchaMouse16. Paper: doi.org/10.52294/001... Code (python/Matlab/C): github.com/DynamicsAndN...
Long-range connections play a large role in shaping dynamics for nearby spatially targeted inputs, but are less important for spatially diffuse inputs like in spontaneous dynamics. This might help explain why we see stronger non-local dynamics in targeted stimulation experiments.
A few key results: The role of long-range connections in shaping dynamics is heavily timescale dependent, concentrated on fast dynamics <~ 30 ms, while slow dynamics resemble the geometric model. This could help explain why they are less important for capturing slower rs-fMRI dynamics.
So Rishi developed a model that allowed us to analyze long-range projections as a perturbation to geometric dynamics and thereby better understand why they might play different roles in different settings. The model dynamics combine wave propagation with 'worm-hole' shortcuts 🕳️🐛
And yet in many experiments, most compellingly in fMRI, 'smearing out' the connectome into a geometric average yields traveling wave dynamics that are a surprisingly strong approximation for data (cf. www.nature.com/articles/s41...)
Long-range connections facilitate rapid communication between networks of distributed cortical populations, are costly, heritable, and we clearly see their functional role in shaping dynamics experimentally, like the rapid non-local 'wormhole' in this whisker-stimulation experiment
New preprint! Why are long-range connectomic interactions in the cortex dominant in shaping dynamics in some experiments but apparently negligible in others? We (w/ R Maran, @elimuller.bsky.social) address this question by studying a new hybrid model of cortical dynamics. arxiv.org/abs/2506.19800
We find that it works: once trained, our MPS approach, "MPSTime", can efficiently learn a time-series model that can be used for time-series classification, imputation of missing data, and synthetic data generation. We demonstrate on synthetic data, medical, industrial, and astronomical data.
We draw an analogy between 1-dimensional (in space) spin chains and 1-dimensional (in time) sequences of measurements: time series. The main idea is to see if a method developed in quantum mechanics—the matrix product state (MPS)—can encode complex correlation structures in time series
New preprint!: "Using matrix-product states for time-series machine learning". arxiv.org/abs/2412.15826 Quick summary below 👇
New preprint by Rishi Maran, Eli Muller: "Analyzing the Brain's Dynamic Response to Targeted Stimulation using Generative Modeling" A review/perspective on why new mechanisms may be found by modeling brain stimulation dynamics 🧠⚡️ arxiv.org/abs/2407.19737
New preprint: "Canonical time-series features for characterizing biologically informative dynamical patterns in fMRI" biorxiv.org/content/10.1... We found a reduced set of time-series features relevant to fMRI (trained in mouse). Code: github.com/DynamicsAndN...
Latest preprint: "Parameter Inference from a Non-stationary Unknown Process" (PINUP) We unify a previously disjoint literature on algorithms for this important problem and introduce new benchmarking results. arxiv.org/abs/2407.089... #timeseriesanalysis #complexsystems
Great work devised and undertaken by Fabiano Baroni comparing >100 statistics of dynamic structure (of synchrony, oscillations, phase relationships, spiking intensity, and variability) of multineuron spike trains, and evaluating them on synthetic and real-world data. www.biorxiv.org/content/10.1...
New preprint: "Tracking the distance to criticality in systems with unknown noise" By Brendan Harris, w/ Leonardo Gollo. We identify new, noise-robust time-series features for tracking the distance to criticality Paper: arxiv.org/abs/2310.14791 Code: github.com/brendanjohnh... #ComplexSystems
New paper: "Neuromodulation of striatal D1 cells shapes BOLD fluctuations in anatomically connected thalamic and cortical regions" in @elife.bsky.social elifesciences.org/articles/78620
catch22 now has a logo! Fast time-series feature-extraction in C, python, Julia, R, and Matlab. github.com/DynamicsAndN...
Some of my time-series feature extraction art now up on the walls of the School of Physics :)