Aitor Morales-Gregorio
LightNews LightNews
banner
aitormg.bsky.social
Aitor Morales-Gregorio
@aitormg.bsky.social
1.3K followers 1.2K following 160 posts
Computational neuroscientist 🧠🇪🇺 @ Charles Uni (previously @fz-juelich.de) | Catching up on graphic design | Electrical engineer vibes⚡ | he/him 🏳️‍🌈

Personal website: https://aitormg.eu
Posts Replies Media Videos
aitormg.bsky.social
Manifold time 😉
Sara Solla giving the Valentin Braitenberg award lecture at the Bernstein 2025 conference in Frankfurt
September 30, 2025 at 2:15 PM
aitormg.bsky.social
I'll be presenting some of my new work on EEG mental imagery, oscillations and (functional) connectivity! If you are still at #CCN2025 make sure to drop by poster C159 today :)
August 15, 2025 at 10:44 AM
aitormg.bsky.social
Are you at the @cnsorg.bsky.social conference?

Do you like networks? Brain-like networks? Pretty and structure rich networks? 🙇‍♂️🧠🕸

Come to my Poster 204 this afternoon (or anytime today)!!

The poster is right next to the coffee! ☕️ 😉

#CNS2025 #CNS2025Florence 🇮🇹🍝
Brain networks from many animals have nonrandom structure. Including high clustering, short path lengths, heavy tailed weight distributions, and heavy-tailed degree distributions.
July 7, 2025 at 10:05 AM
aitormg.bsky.social
Other methods successfully can tell trials apart, but fail to find the left-right symmetry (cosine sim, frobenius).

And fancier methods (symmetric CKA, angular Procrustes) outright fail to find the differences in some cases.

SAS is quite robust to noise!
Comparing non-square matrices of brain activity with SAS, cosine similarity, and the Frobenius norm Comparing non-square matrices of brain activity with SAS, Symmetric CKA, and Angular Procrustes
May 6, 2025 at 5:49 PM
aitormg.bsky.social
And closest to my heart: SAS can distinguish neural population responses for different stimuli, i.e. different trial types.

Even better, SAS can identify that trials with a bar moving left-right or right-left are more similar than orthogonal bars. Probably, due to orientation tuned neurons! 🧠💫
Comparing nonsquare matrices of brain activity with SAS. (a) Schematic diagram of the neuroscience experiment. (b) Single-trial brain activity at identical recording electrodes (64 electrodes, vertical) for each type as example (trial types distinguished by color). Larger values of matrix entries indicate higher MUAe activity (shading in color bar, arbitrary units). (c) Histograms of SAS between all ( 𝑛=120) individual trials. Filled distributions [color code as in (b)] indicate self-similarities, nonfilled ones indicate cross-similarities. Legends show effect sizes 𝜃 for the comparison between self- and cross-similarities.
May 6, 2025 at 5:49 PM
aitormg.bsky.social
Also works great to distinguish networks (from their adjacency matrix) from each other, when they are created using different generative models.
Self- and cross-similarity of different network models. [(a1)–(f1)] Single instances of the different network models. Colored matrix elements indicate a connection between nodes. For the DCM model, connections with a multiplicity higher than one are shown with the same color intensity as single connections. [(a2)–(f2)] Histograms of SAS between instances of the network models ( 𝑛=100, all pairs compared). Filled distributions indicates self-similarities, nonfilled ones indicate cross-similarities. Legends show effect sizes 𝜃 for the comparison between self- and cross-similarities.
May 6, 2025 at 5:49 PM
aitormg.bsky.social
But is SAS any good? We show that it can tell random matrices apart, even when other metrics (cosine similarity, Frobenius norm) fail to do so!

Why? Because it leverages the 2D structure of the matrix instead of looking at a pixel-by-pixel comparison.

Neat!
Comparison of SAS with standard measures. [(a1)–(e1)] Single instances of the different random matrix classes. For visibility, negative values are shown as white. [(a2)–(e2)] Histograms of SAS, cosine similarity, and the Frobenius norm between instances of the random matrix classes ( 𝑛=100, all pairs compared). Filled distributions indicate self-similarities (self-distances), nonfilled ones indicate cross-similarities (cross-distances). Legends show effect sizes 𝜃 for the comparison between distributions.
May 6, 2025 at 5:49 PM
aitormg.bsky.social
Using singular value decomposition (SVD) allows to generalize to non-square matrices and provides some extra information from the left and right singular vectors.

So singular angle similarity (SAS) was born!
Singular value decomposition and singular angles. Schematic representation of the transformations applied on the basis vectors e (yellow) by the components of SVD for two 2×2 matrices Ma (red), and Mb (blue). Small graphs next to the green arrows illustrate the isolated action of the corresponding transformations VT, Σ, and U on e. Below the singular angles of the first [αj,+(V) and αj,+(U)] and second [αj,−(V) and αj,−(U)] kind are shown as the angles between the column vectors of Va and Vb, and Ua and Ub, respectively.
May 6, 2025 at 5:49 PM
aitormg.bsky.social
📣 New & better method to compare matrices just dropped! 📣

Ever wanted to compare things like connectivity matrices or large matrices of neuron responses in trials? There is a new way to do so in town ✨🧪

Singular Angle Similarity (SAS)
journals.aps.org/prxlife/abst...

🧵
Geometric interpretation of Singular Angle Similarity (SAS). Matrices Ma (red) and Mb (blue). Their eigenvectors scaled by the square root of their eigenvalues span the main axes of ellipsoids. These square matrices capture the correlation structure of along the horizontal and vertical axis, respectively (double-headed colored arrows). SAS compares the angles between the corresponding ellipsoids (dashed colored arrows).
May 6, 2025 at 5:49 PM
aitormg.bsky.social
Do you have an external reference? An MRI of the same brain before slicing or at least a brain of the same species (might not work with extremely variable human folding)
Otherwise you risk getting the banana effect:
Aligning slices of a banana (or a brain image) makes the reconstructed 3D volume appear straight instead of curved. A reference volume or some other nonlinear metrics are needed to avoid this. Source: https://doi.org/10.1016/j.media.2017.03.004
March 24, 2025 at 10:24 AM
aitormg.bsky.social
Has someone brought up Georgopoulos 1986 ?

They suggested to first project single neurons to a 3D vector and then add up the vectors for each (directionally tuned) neuron. They did not have multielectrode recordings, but the idea of population coding was there

www.science.org/doi/abs/10.1...
Fig3 from Georgopoulos 1986. Fig5 from Georgopoulos 1986.
February 25, 2025 at 8:03 AM
aitormg.bsky.social
DAY 23: Advent of Comp Neuro 🎄🤖🧠🧪

Connectivity between brain areas depends on distance. Graphs are fascinating! 🕸

“A Predictive Network Model of Cerebral Cortical Connectivity Based on a Distance Rule”
www.sciencedirect.com/science/arti...
Figure 7 from Ercsey-Ravasz et al 2013 Global and Local Communication Efficiency (A and B) Effects of graph density via sequentially deleting weak (blue, green) and strong (black, red) links. Data comparison with (A) EDR (Exponential Distance Rule) model, dashed lines and (B) CDR (Constant Distance Rule) model, dashed lines.
December 23, 2024 at 2:04 PM
aitormg.bsky.social
Wow! Great study and some truly gorgeous figures 🤩
Excerpt from Figure 2 from Michaels et al 2024, showing a very detailed and high quality picture of a macaque using the kinarm setup.
December 23, 2024 at 11:50 AM
aitormg.bsky.social
DAY 22: Advent of Comp Neuro 🎄🤖🧠🧪

Memories in a network can drift at the single neuron level, but persist at the population level! 🍎🍏

“Drifting assemblies for persistent memory: Neuron transitions and unsupervised compensation”
www.pnas.org/doi/full/10....
Figure 2 from Kalle Kossio et al 2021 Drifting assemblies in spiking neural networks. (A) Schematics emphasizing strong synaptic coupling. While an assembly drifts freely (blue-colored assembly neurons) within the interior neurons, its input and output neurons (green and orange) follow it by adapting their synaptic weights. (B) Weights between interior neurons (blue weight matrix), from input and output neurons to the interior neurons (green and orange vertical weight matrices) and from the interior to the input and output neurons (green and orange horizontal weight matrices). Input (output) weights of neuron i are displayed as the ith row (column). Only weights of the four periphery neurons initially (and thus for all times) attached to assembly 1 are shown for clarity. Left column: Network initialization with three assemblies. Center column, after 27 min: Noisy autonomous spiking activity has already driven several interior neurons to attach to a new assembly (blue weight matrix, horizontal and vertical “lines” indicating the changed input and output preference). Right column, after 30 h: The assemblies have drifted away, and the weight matrix is completely remodeled. (C) Like B but with neurons reordered according to assemblies that they belong to, using a clustering algorithm. The assemblies remain intact and the periphery neurons stay strongly coupled to assembly 1. (D, Upper) Spike trains of the input (green) and output (orange) neurons of assembly 1, of 12 neurons from each of the ensembles that initially form assembly 1 (5 to 16), 2 (17 to 28) and 3 (29 to 40) and of 4 inhibitory neurons (black). (D, Lower) Membrane potential of the first interior neuron fluctuates irregularly. Spikes are marked by vertical lines above threshold ; reset is to V0.
December 22, 2024 at 3:07 PM
aitormg.bsky.social
DAY 21: Advent of Comp Neuro 🎄🤖🧠🧪

Aligning stimuli to spontaneous activity creates more robust responses, both in a model and in ferrets!

“The nature-nurture transform underlying the emergence of reliable cortical representations”

By @traegenap.bsky.social et al

www.biorxiv.org/content/10.1...
Figure 4 from Trägenap et al 2023 (A) Schematic illustration of the feedforward-recurrent alignment hypothesis. (B) Top: Trial-to-trial correlation for random input (green) and input maximally aligned to the recurrent network (blue). Bottom: Trial-to-trial correlation as a function of feedforward-recurrent alignment. (C) As (B), but for intra-trial stability. Top shows activity of a network unit. (D) As (B), but for dimensionality. Top shows variance explained by leading PCs of evoked activity (see Fig. 2B,C for comparison to experiment). Input is a 10-dimensional Gaussian distribution. (E) Model of spontaneous activity (schematic, bottom): A broad, high-dimensional distribution of inputs aligned with the recurrent network (see Methods), producing a lower-dimensional output (top). (F) Alignment of evoked to spontaneous activity. (G) Model prediction: trials for which the steady-state evoked response is more aligned to spontaneous activity also show higher intra-trial stability. N=300 samples drawn for the input distribution used in (D) with optimal alignment and temporally varying noise (H) Outline of grating evoked activity 1 s (magenta) and 3 s (cyan) after stimulus onset for a response pattern more weakly aligned (left) and for one more strongly aligned (right) to spontaneous activity 0.5 s after stimulus onset. (I) As (G) but for experimental data. (J) Correlation for the data shown in (I) (large blue circle) and for all other animals in the oldest age group (grey circles; 13 experiments from N=11 animals at EO+4 to +7). Error bars show Mean ± SEM. respectively). (K) Contours show outlines of grating evoked patterns from different trials for a stimulus for which the average pattern is less (left) or more (right) aligned to spontaneous activity. (L) As in (J), but quantifying the relationship illustrated in (K). (M) Evoked dimensionality co-varies with the alignment of the first evoked PC to spontaneous activity. N=11 animals at EO+4 to +7
December 21, 2024 at 12:04 PM
aitormg.bsky.social
DAY 20: Advent of Comp Neuro 🎄🤖🧠🧪

Discovering new plasticity rules using evolutionary algorithms and cartesian genetic programming 🧬

“Evolving interpretable plasticity for spiking networks”
elifesciences.org/articles/66273
Figure 5 from Jordan et al 2021 Cartesian genetic programming evolves diverse correlation-driven learning rules. (A) Network sketch. Multiple inputs project to a single output neuron. The current synaptic weight wj and the eligibility trace Ecj are provided to the plasticity rule that determines the weight update. (B) Membrane potential u of the output neuron over the course of learning using Equation 17. Gray boxes indicate presentation of the frozen-noise pattern. (C) Fitness (Equation 13) of the best individual per generation as a function of the generation index for multiple runs of the evolutionary algorithm with different initial conditions. Blue and red curves correspond to the two representative plasticity rules selected for detailed analysis. Blue and red markers represent fitness of the two representative rules and the orange marker the fitness of the homeostatic STDP rule (Equation 17; Masquelier, 2018), respectively, on 20 validation tasks not used during the evolutionary search. Error bars indicate one standard deviation over tasks. (D, E): Learning rules evolved by two runs of CGP (D: LR1, Equation 19; E: LR2, Equation 20). (F): Homeostatic STDP rule Equation 17 suggested by Masquelier, 2018. Top panels: STDP kernels Δwj as a function of spike timing differences Δtj for three different weights wj. Bottom panels: homeostatic mechanisms for those weights. The colors are specific to the respective learning rules (blue for LR1, red for LR2), with different shades representing the different weights wj. The learning rate is η=0.01.
December 20, 2024 at 3:18 PM
aitormg.bsky.social
DAY 19: Advent of Comp Neuro 🎄🤖🧠🧪

Hyperbolic geometry for the representation of space in rat hippocampus 🐀

“Hippocampal spatial representations exhibit a hyperbolic geometry that expands with experience”
www.nature.com/articles/s41...
Figure 7 from Zhang et al 2023 Schematic illustration of how increased hyperbolic radius affects neural representation of space. As radius grows over time, exponential distribution of place field sizes shifts toward smaller place fields or larger depth in a discrete tree structure, to improve spatial localization.
December 19, 2024 at 6:27 PM
aitormg.bsky.social
DAY 18: Advent of Comp Neuro 🎄🤖🧠🧪

Topology of ripples is one continuous manifold, but different parts for sleep, learning etc. 🤯

“Topological analysis of sharp-wave ripple waveforms reveals input mechanisms behind feature variations”
By @lmprida.bsky.social et al
www.nature.com/articles/s41...
Part of Figure 2 from Sebastian et al 2023 c, Ripples recorded at the SP are represented in a high-dimensional waveform space (127D) determined by the sampling rate (2,500 Hz) and the window size (±25 ms). In this representation, each SWR is a point in the cloud made by all detected events. d, Topology of the SWR cloud was examined with persistent homology, which identifies holes and cavities in the data by using simplicial complexes. These are made of simplices (for example, dots, lines, triangles and tetrahedrons) that connect data points (SWRs in this case) in a given radius. Persistent homology looks for their persistence as the radius around each point varies in the high-dimensional space. The different homology groups (H) are defined from the number of cuts that separate simplices in pieces. e, Barcodes for the three homology groups (Ho, H1 and H2) show persistence of only one continuous component in the experimental SWR data cloud (data from a random subset of n = 3,500 events).
December 18, 2024 at 3:27 PM
aitormg.bsky.social
DAY 17: Advent of Comp Neuro 🎄🤖🧠🧪

Clusters of neurons shape the attractor dynamics of neural networks.

Plus I am a simple person, I see eigenvalue spectra and I instantly like the paper 10 times more.

“Emergence of Slow-Switching Assemblies in Structured Neuronal Networks”
tinyurl.com/4zazbxhj
Figure 1 from Schaub et al Network dynamics and eigenvalue spectra of two LIF networks. One network with uniform synaptic connections (left), and one with 20 groups of clustered excitatory connections (right). A Visualization of the network topologies (top) and exemplars of raster plots (bottom). The dynamics of the clustered network exhibit the banded structure associated with slow-switching group activity. The magnitude of this switching can be characterized statistically a posteriori from the data through the spike-rate variability metric Ŝ, defined in Materials and Methods Eq (18), as discussed in the text. In this case, the unclustered network has Ŝ=0.035 while the clustered network has a much larger value Ŝ=8.23. B Eigenvalue spectra of the network weight matrices W. The weighted connectivity matrix of the clustered network exhibits a clear eigengap Δλ separating the 19 eigenvalues with largest real parts from the cloud of eigenvalues in the bulk. There is no such eigengap for the unclustered network. As indicated by the two arrows, both matrices have a pair of complex conjugate eigenvalues associated with the (damped) global activation modes of the networks characteristic of balanced networks.
December 17, 2024 at 2:09 PM
aitormg.bsky.social
DAY 16: Advent of Comp Neuro 🎄🤖🧠🧪

There is a network of oscillating neurons in the frontoparietal areas during movement!

“Uniting functional network topology and oscillations in the fronto-parietal single unit network of behaving primates”
@benjamindann.bsky.social et al
tinyurl.com/53ydvdxh
The fronto-parietal network of neurons exhibits strong beta and low frequency synchronisation during planned and executed grasping movements. Summary figure kindly provided by Benjamin Dann :)
December 17, 2024 at 1:59 PM
aitormg.bsky.social
DAY 15: Advent of Comp Neuro 🎄🤖🧠🧪

Heterogeneity of neural properties can improve some computational functions, but homogeneity also has its advantages! It's complicated 😅

“Neural heterogeneity controls computations in spiking neural networks”
@rgast.bsky.social et al
www.pnas.org/doi/epub/10....
Figure 1 from Gast et al Heterogeneity linearizes neural population dynamics. (A–D) Two-dimensional (2D) bifurcation diagrams are depicted for different cell types. Regions colored in gray and green depict bistable and oscillatory regimes, respectively. The black and orange crosses depict approximate bifurcation points, estimated from the dynamics of simulated SNNs with a Lorentzian and a Gaussian distribution of the spiking thresholds, respectively. Spiking neural network dynamics were obtained from simulating networks of = 1,000 neurons connected by sparse, random couplings (coupling probability of 20%) The y-axis on the Left (Right) depicts the width of the Lorentzian (Gaussian) distribution used to generate the SNNs that result on the bifurcation points shown by the black (orange) crosses. (A) Excitatory regular-spiking neurons with low spike-frequency adaptation ( pA). (B) Excitatory regular-spiking neurons with high spike-frequency adaptation ( pA). (C) Inhibitory fast-spiking neurons. (D) Inhibitory low-threshold-spiking neurons.
December 17, 2024 at 1:06 PM
aitormg.bsky.social
DAY 14: Advent of Comp Neuro 🎄🤖🧠🧪

Neurotransmitter expression is very different throughout the cortex, and those differences are aligned with the hierarchy!

“Gradients of neurotransmitter receptor expression in the macaque cortex”
By @seanfw.bsky.social et al

www.nature.com/articles/s41...
Figure 4 from Froudist-Walsh et al An inverse relationship between cortical myelin and receptor density. a, Cortical T1w/T2w ratio, a proposed marker for myelin content, is strongly negatively correlated with the principal receptor gradient across 109 cortical areas (r(107) = −0.72 (range: −0.80 to −0.62), P = 0.02, corrected for spatial autocorrelation and multiple comparisons (Bonferroni)). b, Receptor-, cyto-architecture and myelo-architecture of the macaque primary visual cortex (V1). Changes in Gray Level Index, which represents a measure of the volume fraction of cell bodies, in myelin density and in receptor concentration (in fmol mg−1 of protein) throughout the cortical depth are provided by the profile curve overlaid onto each section. Note that the scale has been optimized for each profile to provide the best visualization of changes in receptor densities throughout the cortical ribbon. Roman and Arabic numerals indicate cyto-architectonic and myelo-architectonic layers, respectively. Positions of cyto-architectonic layers were transferred to the neighboring receptor images.
December 16, 2024 at 5:18 PM
aitormg.bsky.social
DAY 13: Advent of Comp Neuro 🎄🤖🧠🧪

It is surprisingly hard to achieve signal propagation in large recurrent networks!

“Inter-areal Balanced Amplification Enhances Signal Propagation in a Large-Scale Circuit Model of the Primate Cortex”
www.cell.com/neuron/fullt...
Figure 2 from Joglekar et al Signal Propagation in a Large-Scale Network with Recurrent Excitatory Loops Leads to Either Strong Attenuation or Instability (A) Network model of the 29 cortical areas with the strongest inter-areal projection strengths (FLN values >0.005). Input is injected to the excitatory population in area V1 at the lowest position in the cortical hierarchy, and the peak firing rate in area 24c at the top of the hierarchy is recorded. (B) The response of the excitatory populations in areas V1 and 24c to a current pulse of 250 ms to V1, using two close values of the excitatory global coupling parameter 𝜇𝐸⁢𝐸. Response (left) in 24c shows an attenuation of four orders of magnitude when 𝜇𝐸⁢𝐸=34. With 𝜇𝐸⁢𝐸=36 (right), excitatory firing rates in V1 and 24c exponentially grow, and the system becomes unstable.
December 16, 2024 at 5:04 PM
aitormg.bsky.social
DAY 12: Advent of Comp Neuro 🎄🤖🧠🧪

Simulating spiking neuron networks is quite challenging, now the simulations can be done faster than biologic time! ⚡

"Sub-realtime simulation of a neuronal network of natural density"
By @ackurth.bsky.social et al

iopscience.iop.org/article/10.1...
Figure 1 from Kurth et al Strong scaling of a cortical microcircuit model on a conventional compute node. (a) Sketch of the microcircuit model with about 80 000 neurons and 300 million synapses organized into four layers of excitatory (blue) and inhibitory (red) populations of neurons. (b) Strong scaling for two placing schemes. Top graph shows realtime factor (RTF) over total number of threads; dashed horizontal line indicates realtime; black solid line indicates linear scaling. The sequential scheme (blue) minimizes distance of threads on hardware, the distant scheme (brown) maximizes it both using 1 MPI process per node; pale coloured line segments connect data points, dashed vertical lines indicate number of cores per processor (64) and node (128). On node hybrid parallelization by 2 MPI processes (1 per socket) for full nodes for comparison (green). Bottom graphs show fractions of wall-clock time consumed by different stages of the simulation cycle; update: integrates state of neurons, deliver: distributes spike events to target neurons, communicate: transfers spikes between MPI processes (for shared and distributed memory setups), other: not accounted for by timers. (c) Top three graphs: power measurements of a compute node during 100 s of model time in three configurations. The measurements are aligned to the start of the simulation phase starting at t = 0. Bottom graph: cumulative energy consumption of the simulations.
December 16, 2024 at 4:32 PM
aitormg.bsky.social
DAY 11: Advent of Comp Neuro 🎄🤖🧠🧪

A revisited classic! Perfectly functional networks, with disparate parameters! 🦞

“Energy-efficient network activity from disparate circuit parameters”
By @deismic.bsky.social , @jakhmack.bsky.social and @ppjgoncalves.bsky.social
www.pnas.org/doi/abs/10.1...
Figure 2 from Deistler et al Bayesian inference reveals wide range of energy consumption. (A) Inferring the posterior distribution by combining a rejection classifier and a deep neural density estimator. First, a classifier (trained on an initial set of simulations) predicts which circuit parameters sampled from the prior produce valid simulation outputs. We then proceed by sampling from the part of the parameter space that is accepted by the classifier, i.e., the restricted prior. All valid data (green) are used to train a deep neural density estimator, and all invalid data are discarded (red). Once this estimator is trained, it can be evaluated on experimental data to return the posterior distribution over model parameters. (B) Experimental data recorded from the pyloric network. Arrows indicate a subset of the physiologically relevant features, namely, the cycle period (arrow 1), phase delays (arrow 2), phase gaps (arrow 3), and burst durations (arrow 4). (C) Simulation output from a parameter set sampled from the prior distribution. The traces are AB/PD (Top), LP (Middle), and PY (Bottom). (D) Subset of the marginals and pairwise marginals of the 31-dimensional restricted prior, i.e., the subspace of parameters for which the model produces bursting activity. All maximal conductances are given in mS/. (E) Subset of the marginals and pairwise marginals of the posterior distribution, i.e., the subspace of parameters for which the model matches experimental data shown in B. (F) Sample from the restricted prior producing bursting activity but not matching experimental data. (G) Sample from the posterior distribution closely matching features of the experimental data. (H) Histograms over energy consumed by each neuron (blue, orange, and green) as well as by the entire circuit (black), divided by the duration of the simulation. Trace with lowest energy consumes nine times less energy than trace with highest energy. (I) Same as in H but for energy per spike.
December 11, 2024 at 5:35 PM