Biology subjects

Dimitriadis, G.

Publications and source records attributed to Dimitriadis, G..

5 recordsLinked to original sources

bioRxiv2026

From sequences to schemas: low-rank recurrent dynamics underlie abstract relational representations

A hallmark of intelligent behavior is the ability to extract abstract relational structure from temporal sequences -- recognizing, for instance, that aab, ccd, and eef all follow the same underlying pattern, regardless of the specific elements involved. This capacity, observed across species and sensory modalities, is thought to underlie the formation of cognitive schemas: compressed internal models that support rapid generalization to novel experiences. Yet the neural circuit mechanisms by which such abstract, identity-independent representations emerge from sequential experience remain largely unknown. Here, we investigate this question using Recurrent Neural Networks (RNNs) as mechanistic models of neural circuits, trained to classify sequences based on their latent algebraic patterns (e.g., aab, aad [->] AAB; aba, aca [->] ABA) without supervision on intermediate transitions. We demonstrate that RNNs spontaneously learn low-dimensional representations that mirror the hierarchical generative structure of the sequences, and that this abstraction is mechanistically supported by the emergence of low-rank recurrent connectivity. The leading singular component of the recurrent weight matrix integrates relational transition information -- whether consecutive tokens are the same or different -- across time, driving the formation of a structured, tree-like geometry in the population state space. Through singular vector ablation, we establish a causal role for this component: removing it selectively erases memory for earlier transitions while leaving local, single-step sensitivity intact. Finally, while RNNs trained on next-token prediction do not spontaneously acquire these abstract representations, transferring the low-rank scaffold learned from classification significantly accelerates learning and improves generalization -- an effect specific to the abstract structure of the scaffold rather than generic statistical pretraining. These findings offer a computational account of how task demands shape recurrent connectivity to support temporal abstraction, with direct implications for understanding schema formation in biological brains.

neuroscience↗

bioRxiv2023

Heron: A Knowledge Graph editor for intuitive implementation of python based experimental pipelines.

To realise a research project idea, an experimenter faces a series of conflicting design and implementation considerations, regarding both its hardware and software components. For instance, the ease of implementation, in time and expertise, should be balanced against the ease of future reconfigurability and number of black box components. Other, often conflicting, considerations include the level of documentation and ease of reproducibility, resource availability as well as access to online communities. To alleviate this balancing act between opposing requirements we present Heron, a new Python-based platform to construct and run experimental and data analysis pipelines. Herons main principle is to allow researchers to design and implement the experimental flow as close as possible to their mental schemata of the experiment, in the form of a Knowledge Graph. Heron is designed to increase the implementation speed of experiments (and their subsequent updates), while minimising the number of incorporated black box components. It enhances the readability and reproducibility of the final implementation and allows the use of combinations of hardware and software otherwise impossible or too costly to achieve. Given this, Heron offers itself to sciences whose needs involve experiments with a large number of interconnected hardware and software components like robotics, neuroscience, behavioural sciences, physics, chemistry, environmental science, etc.. It is designed with those experimentalists in mind which: i) Demand full control of their setup. ii) Prefer not to have to choose between hardware and software that run only on a specific chip/operating system combination. iii) Appreciate the ease and speed that high-level languages (e.g. Python) and Graphical User Interfaces (GUIs) offer them. It assumes an intermediate knowledge of the Python language and ecosystem, offering a flexible and powerful way to construct experimental setups. It removes any inaccessible corners, yet keeps implementation costs significantly reduced compared to using lower level languages. Finally, its use results in a much cleaner and easier to understand code base, amicable to documentation and reproducibility efforts.

animal behavior and cognition↗

bioRxiv2018

Recording from the same neuron with high-density CMOS probes and patch-clamp: a ground-truth dataset and an experiment in collaboration.

We built a rig to perform patch-clamp and extracellular recordings from the same neuron in vivo. In this setup, the axes of two micromanipulators are precisely aligned and their relative position tracked in real-time, allowing us to accurately target patch-clamp recordings to neurons near an extracellular probe. We used this setup to generate a publicly-available dataset where a cortical neurons spiking activity is recorded in patch-clamp next to a dense CMOS Neuropixels probe. "Ground-truth" datasets of this kind are rare but valuable to the neuroscience community, as they power the development and improvement of spike-sorting and analysis algorithms, tethering them to empirical observations. In this article, we describe our approach and report exploratory and descriptive analysis on the resulting dataset. We study the detectability of patch-clamp spikes on the extracellular probe, within-unit reliability of spike features and spatiotemporal dynamics of the action potential waveform. We open discussion and collaboration on this dataset through an online repository, with a view to producing follow-up publications. PrologueOur efforts to record from the same neuron in vivo using patch-clamp and dense extracellular probes have resulted in three outputs: a publicly-available dataset (http://bit.ly/paired_recs), a manuscript, and a code repository (http://bit.ly/paired_git). Together, these three components form the publication arising from the experiments we have performed. The role of the dataset is to be downloaded and re-used. The role of the manuscript is to describe the experimental methods through which we acquired the dataset, explain it and showcase which types of questions it can be used to address. The repository has two roles: first, promoting reproducibility and error correction. By making our analysis and figure-generation code freely-available, we wish to make our analysis procedures clear and enable the reader to reproduce our results from the raw data, alerting us to any potential mistakes. Second, the repository will form a living, dynamic and interactive component of the publication: a forum for open collaboration on this dataset. Any interested scientists can contribute to it, joining us in detailed exploration of these recordings with a view to producing follow-up publications in which they will be credited for their input. Why did we opt to publish this way? The first reason is that the very nature of the project we here describe - recording the same neuron with patch-clamp and extracellular probes - invites an open science and open source approach. This is because the primary use of this type of "ground truth" validation data is to aid the development of new sorting and analysis algorithms, as well as to benchmark and improve existing ones. The second reason is that despite being conceptually very simple, this project generated a large and complex dataset that can be tackled in many ways and used to address different types of question. Some of these questions are beyond the reach of our analytical expertise; others lie even beyond the scope of our scientific imagination. By releasing the dataset and providing a repository for scientific discussion and collaboration, we aim to maximise its scientific return to the community. Instead of having each interested research group work in isolation, we hope that by encouraging collaboration and discussion between peers we can foster synergy between them that will lead to work of greater scientific value. Although datasets like ours are exquisitely suited for such an approach, we believe this publication strategy needs to become more widely adopted in neuroscience. We were pleased to note recent publications spontaneously and independently using similar approaches1-4, in what may well be evidence of convergent thinking. Perhaps the time has come for new publication and collaboration paradigms. We will elaborate on this subject during the Epilogue. For now, let us get back to electrophysiological recordings, before we begin an experiment on scientific collaboration.

neuroscience↗

bioRxiv2018

Why not record from every channel with a CMOS scanning probe?

It is an uninformative truism to state that the brain operates at multiple spatial and temporal scales, each with each own set of emergent phenomena. More worthy of attention is the point that our current understanding of it cannot clearly indicate which of these phenomenological scales are the significant contributors to the brains function and primary output (i.e. behaviour). Apart from the sheer complexity of the problem, a major contributing factor to this state of affairs is the lack of instrumentation that can simultaneously address these multiple scales without causing function altering damages to the underlying tissue. One important facet of this problem is that standard neural recording devices normally require one output connection per electrode. This limits the number of electrodes that can fit along the thin shafts of implantable probes generating a limiting balance between density and spread. Sharing a single output connection between multiple electrodes relaxes this constraint and permits designs of ultra-high density probes. Here we report the design and in-vivo validation of such a device, a complementary metal-oxide-semiconductor (CMOS) scanning probe with 1344 electrodes; the outcome of the European research project NeuroSeeker. We show that this design targets both local and global spatial scales by allowing the simultaneous recording of more than 1000 neurons spanning 7 functional regions with a single shaft. The neurons show similar recording longevity and signal to noise ratio to passive probes of comparable size and no adverse effects in awake or anesthetized animals. Addressing the data management of this device we also present novel visualization and monitoring methods. Using the probe with freely moving animals we show how accessing a number of cortical and subcortical brain regions offers a novel perspective on how the brain operates around salient behavioural events. Finally, we compare this probe with lower density, non CMOS designs (which have to adhere to the one electrode per output line rule). We show that an increase in density results in capturing neural firing patterns, undetectable by lower density devices, which correlate to self-similar structures inherent in complex naturalistic behaviour. To help design electrode configurations for future, even higher density, CMOS probes, recordings from many different brain regions were obtained with an ultra-dense passive probe.

neuroscience↗

bioRxiv2016

T-SNE visualization of large-scale neural recordings.

Electrophysiology is entering the era of Big Data. Multiple probes, each with hundreds to thousands of individual electrodes, are now capable of simultaneously recording from many brain regions. The major challenge confronting these new technologies is transforming the raw data into physiologically meaningful signals, i.e. single unit spikes. Sorting the spike events of individual neurons from a spatiotemporally dense sampling of the extracellular electric field is a problem that has attracted much attention [22, 23], but is still far from solved. Current methods still rely on human input and thus become unfeasible as the size of the data sets grow exponentially.\n\nHere we introduce the t-student stochastic neighbor embedding (t-sne) dimensionality reduction method [27] as a visualization tool in the spike sorting process. T-sne embeds the n-dimensional extracellular spikes (n = number of features by which each spike is decomposed) into a low (usually two) dimensional space. We show that such embeddings, even starting from different feature spaces, form obvious clusters of spikes that can be easily visualized and manually delineated with a high degree of precision. We propose that these clusters represent single units and test this assertion by applying our algorithm on labeled data sets both from hybrid [23] and paired juxtacellular/extracellular recordings [15]. We have released a graphical user interface (gui) written in python as a tool for the manual clustering of the t-sne embedded spikes and as a tool for an informed overview and fast manual curration of results from other clustering algorithms. Furthermore, the generated visualizations offer evidence in favor of the use of probes with higher density and smaller electrodes. They also graphically demonstrate the diverse nature of the sorting problem when spikes are recorded with different methods and arise from regions with different background spiking statistics.

neuroscience↗