Keeping Track: Time perception and numerosity in brains and machines

Timing and numerosity are fundamental aspects in both biological and synthetic systems, playing critical roles in processes like memory, motion and causal inference. From circadian rhythms to comparing quantities and planning the timing of movement, the nervous system must accurately process and coordinate both temporal and numeric information across different neural areas and time scales. While time perception and numerosity in biological agents are not necessarily straightforward, artificial agents can consistently access objective time and numeric information. However, large language models’ ability to understand time (for example, relative sequences of abstract events) and numerosity (like estimating quantities without counting) seems to be intrinsically poor, unless they are explicitly trained to care about these domains. In general, the ability of agents to acquire a ‘sense’ of time and numbers continues to be challenges for the field. In this symposium, we’ll explore time perception, accuracy, sensibility, and numerosity, from millisecond processing to the lifespan, across genetic, neural, and artificial agents.

Cartoon of a brain in a sand timer with numbers falling through. Text "Keeping track - Time perception & numerosity in brains & machine"

Provisional Schedule* 


Intro by the SWC/Gatsby Student Symposium Team


Speaker 1: Hugo Merchant

Encoding rhythmic time and context in the primate cortico-basal ganglia circuit


Virtual/Interspersion 1: Martin Riemer

Underlying premises for using virtual reality in time perception research


Speaker 2: Jennifer Coull

Dopamine depletion perturbs perception of duration but not temporal order


Lunch break (90 min)


Speaker 3: Andreas Nieder

Number Neurons in the Human Brain


Virtual/Interspersion 2: Karli Nave

A Multi-Lab EEG Replication and Extension of “Tagging the Neural Entrainment to Beat and Meter


Speaker 4: Aida Nematzadeh

Number Sense in Vision-Language Models


Coffee break (10 min)


Student engagement session


Presentations to panel of faculty judges (food and drinks will be provided)

*all timings in BST (UTC+1)



Hugo Merchant
Hugo Merchant is a Professor at the National University of Mexico in the Institute of Neurobiology. His lab investigate the neurophysiological bases of time estimation in cortico-thalamic basal ganglia circuits of primates. To explore temporal behaviour in monkeys, they designed tasks to look at both time perception and the production of time intervals, and compare their abilities with humans carrying out the same task. They characterise behaviour using psychophysical techniques, and measure neural activity through single neuron recordings and neuroimaging methods

Jennifer Coull
Jennifer Coull is a Centre national de la recherche scientifique (CNRS) Senior Research Fellow at Laboratoire des Neurosciences Cognitives (LNC) of Aix-Marseille University. Jennifer's work focuses on the functional and neural substrates of timing and temporal attention, using functional neuroimaging (fMRI) & psychopharmacology to understand the perception of time and the role of dopaminergic (DA) system in it.

Andreas Nieder
Andreas Nieder is a Professor of Animal Physiology and Director of the Institute of Neurobiology at the University of Tübingen. His research focuses around intelligent, goal-directed behaviours and concepts including number representations, consciousness, and vocalisation. To this end he has investigated how these concepts arise in neocortex of highly intelligent species, from primates to songbirds. Read more about Andreas’ journey through the concept of numbers in his book: A Brain for Numbers: The Biology of the Number Instinct

Aida Nematzadeh
Aida Nematzadeh is a research scientist at DeepMind. Before joining DeepMind, she was a postdoctoral researcher at UC Berkely Computational Cognitive Science Lab and Bear, working with Tom Griffiths. She obtained her PhD at University of Toronto, advised by Suzanne Stevenson and Afsaneh Fazly. She has been working on the intersection of computational linguistics, cognitive science and machine learning. Her recent work focuses on understanding and evaluating pre-trained language or vision-language models.


Our virtual speaker sessions highlight early-career researchers.

Martin Riemer
Martin Riemer leads the Time and Space research group at the department of Biological Psychology and Neuroergonomics at Technical University of Berlin. He completed his PhD in Psychology at the University of Mannheim, followed by a post-doctoral training at the German Center for Neurodegenerative Diseases in Magdeburg and at the University of Groningen. His current research aims to understand complex cognitive behavior such as time perception and spatial cognition in both healthy and pathological aging subjects, using a wide range of tools including virtual environments.

Karli Nave
Karli Nave is a postdoctoral associate at Western University in London, Ontario, where she studies neural underpinnings of music and rhythm. She completed her PhD at the University of Nevada Las Vegas in psychological and brain sciences, and has worked on projects in music cognition, auditory rhythm perception, and the neural mechanisms underlying rhythm processing.


Hugo Merchant
Beat induction is the cognitive ability that allows subjects to hear a regular pulse in music and to move in synchrony with it. Although humans are very flexible at perceiving and synchronizing to complex beats, we have shown that macaques are also able to produce predictively and accurately isochronous intervals that are cued by auditory or visual metronomes or when intervals are produced internally without sensory guidance during a synchronization continuation task. In addition, we found that neurons in the medial premotor areas (preSMA and SMA), the putamen and the dorsolateral prefrontal cortex (PF) show mixed selectivity during task performance. Cells of the three areas encode not only elapsed time, but also the tempo of the tapping, the metronome´s modality, and the serial order element in the rhythmic sequence. Comparting the representation of task parameters between areas using decoding methods, we found that MPC acts as a master amodal clock. MPC robustly encoded relative timing and the metronome´s tempo using different neural subpopulations, which also were tuned to the modality of the metronome and the context in which the animals produced the intervals. Hence, MPC is a key player for beat based timing, mapping time with other task parameters by dynamically engaging diverse cell populations.

Jennifer Coull
Data from animals, healthy humans and patient populations converge to implicate the dopaminergic (DA) system in the perception of time. In a series of within-subject, placebo-controlled studies, we have investigated the effects of DA manipulation on a variety of distinct temporal processes in healthy human participants. We manipulated DA function using acute phenylalanine/tyrosine depletion (APTD), an amino acid drink that temporarily lowers levels of the dopamine precursors phenylalanine and tyrosine.  APTD selectively impaired both perceptual and motor indices of duration estimation, with the direction of effect differing according to individual differences in baseline levels of DA precursor availability. Specifically, APTD caused participants with relatively low baseline DA to overestimate duration but those with high baseline DA to underestimate it. This dissociation highlights the importance of accounting for individual differences in endogenous DA function in psychopharmacological investigations. Our fMRI data further revealed that APTD exerted its effects on perceptual timing by attenuating activity in the basal ganglia and Supplementary Motor Area, key nodes of the timing network. Since these regions are typically associated with motor function, our data indicate a neuroanatomical link between action and time perception. In contrast to its clear effects on duration processing however, APTD had no significant effect on either the ability to resolve the succession of events (simultaneity judgement) or to perceive their relative order (temporal order judgement). This neurochemical dissociation adds to behavioural, developmental and neuroanatomical evidence that duration and order are functionally distinct temporal processes.

Andreas Nieder
Our scientifically and technologically advanced culture would be impossible without our understanding of numbers and mathematics. We addressed the mechanism of how single neurons, the anatomical and functional units of the brain, encode numerical information and recorded from single cells in the medial temporal lobe (MTL) of neurosurgical patients. The recordings revealed neurons tuned to preferred numerical values presented in different formats.  Different neuronal mechanisms  were  seen  for  the  rapid  assessment  of  small numbers as opposed to an approximate estimation of large numbers. While the subjects calculated with numbers, neurons showed an abstract representation of arithmetic (addition and subtraction) rules, but drastically different neuronal codes for arithmetic in different MTL regions. This line of research unveils the workings of the human brain during mental arithmetic, a symbolic competence in humans.

Aida Nematzadeh
The ability to represent and reason about numbers in different contexts is crucial for Artificial Intelligence systems. In this talk, I present our work in evaluating recent vision–language models with respect to their ability to capture number sense–basic numerical competence observed in both humans and animals. First, I will discuss how each modality (vision and language) contributes to the structure of number representations in these models. Next, I will describe the numerical bisection procedure–motivated by studies in animal and infant numerical cognition–to test number discrimination in different families of vision encoders. Our results suggest that vision-specific inductive biases are helpful in numerosity discrimination, as models with such biases have lowest test errors on the task, and often have psychometric curves that qualitatively resemble those of humans and animals performing the task.

Martin Riemer
Technical developments in virtual reality (VR) applications have been immensely fruitful for the investigation of human time perception. The method VR enables the use of realistic stimuli, embedded within an immersive, naturalistic context, while maintaining a high level of experimental control over the presented stimuli. VR allows to manipulate external Zeitgeber (e.g., speed of environmental changes) and to control for environmental factors such as spatial size and the number of surrounding objects, which are known to influence subjective time flow. However, an important premise for the interpretation of the results in VR studies is that the method itself does not affect time perception, a premise that is often neglected or implicitly taken for granted. In my talk I will elaborate on the problems associated with this pre-assumption, present some recent data on the issue, and discuss the implications for the interpretation of VR studies in time perception research.

Karli Nave
Nozaradan et al. (2011) found enhanced frequency-tagged electroencephalographic (EEG) brain activity at beat-related frequencies when listeners imagined a pattern as being in a duple or triple musical meter while presented an ambiguous 
isochronous auditory stimulus. However, it is unclear whether this represents repeatable evidence for musical beat perception reflected in brain activity. This study was replicated in 13 laboratories (N= 154 participants), using a pre-registered
and provisionally-accepted protocol, with an added behavioral task that measured beat perception on each trial. We estimated the meta-analytic effect sizes for differences between imagery conditions (duple vs. passive, triple vs. passive, duple 
vs. triple), as well as moderating effects of music and dance training. Non-registered analyses of variance (ANOVAs) were also performed to detect significant effects of imagery on brain activity in this relatively large sample of participants. Voltage 
differences between different imagery conditions (0.04 uV) were consistently smaller than in the original study (0.16 uV), and all confidence intervals encompassed 0 uV. No moderating effects of musical or dance experience occurred.
Exploratory ANOVAs showed a significant effect of imagery condition for beat-related frequencies, but effect sizes were considerably smaller (partial eta-squared=0.11) than in the original study (partial eta-squared=0.67). There may be a small effect of imagery on beat-related brain activity. Moderating effects of musical or dance training may require much larger samples to detect. Our finding of smaller effect sizes underscores the need to widely embrace practices such as pre-registration, a priori power analysis, and replications.

SWC/GCNU Student Symposium Team

The Student Symposium is organised jointly by PhD students of the Sainsbury Wellcome Centre for Neural Circuits and Behaviour and the Gatsby Computational Neuroscience Unit of University College London. This is the sixth instalment of an annual discussion-based event that aims to bring together neuroscience researchers from the UK and abroad to engage with current and future problems in neuroscience.