Summary: To discover shared mind activity patterns, researchers have created a mathematical deep learning technique. The technique, known as MARBLE, learns active motifs from neurological recordings and identifies common tactics employed by various brains to accomplish the same task.
Stone, which was tested on macaques and rats, outperformed different machine learning techniques for decoding neural activity related to movement and routing. The program works by combining neurological information with high-dimensional mathematical spaces, enabling pattern recognition across people and circumstances.
Important Information:
- By encoding neurological signals onto high-dimensional shapes, MARBLE uses geometrical deep learning to identify shared mental activity patterns.
- Cross-Subject Comparisons: The technique was able to identify common neural patterns in various animals that carried out the same job.
- Applications in Brain-Machine Interfaces: By decoding mental action into distinctive patterns, MARBLE could boost supporting robotics and neuroscience research.
Origin: EPFL
A unique aspect of an rhino they are touching, such as a strong horn, a flexible tree, or a large knee, is discussed in the parable of the blind men and the elephant, but the blind men disagree on the animal’s actual nature.
The narrative illustrates the difficulty of understanding an invisible or hidden object based on imperfect individual perceptions.
Similar to how researchers who study head dynamics based on recordings of a small number of neurons may conclude the hidden patterns of mind dynamics that drive these recordings.
” Imagine that you and I both work on a mental process, such as finding a way to work. Can neurons ‘ signs indicate that we employ the same or many different cognitive approaches to accomplish the task? says Pierre Vandergheynst, mind of the Signal Processing Laboratory , LTS2 , in EPFL’s School of Engineering.
Because researchers frequently report data from numerous animals, but we have little information as to whether they represent a given job using the same mind patterns, this is a fundamental problem for science.
A mathematical deep learning strategy, according to Vandergheynst and former doctorate Adam Gosztolai, who is now an associate professor at the AI Institute of the Medical University of Vienna, can be inferred from latent brain activity patterns across experimental subjects in Nature Methods.
This is accomplished by transforming electrical neural activity into dynamic patterns, or motifs, that a geometric neural network can learn.
The researchers used MARBLE to demonstrate that when different animals used the same mental strategy to reach an arm or navigate a maze, their brain dynamics were made up of the same motifs in experiments on macaque and rat brain recordings.
Dynamic data using a geometric neural net
Traditional deep learning, such as firing neurons or flowing fluids, is unsuited to comprehend dynamic systems that frequently change as a function of time.
These activity patterns are so intricate that they best fit the description of them as geometric objects in high-dimensional spaces. A torus, which resembles a donut, is one such object.
MARBLE is unique because, according to Gosztolai, it learns from within curved spaces, which are natural mathematical spaces for intricate patterns of neuronal activity.
The geometric deep learning algorithm is unaware that these spaces are curved inside the curved spaces. Thus, the dynamic motifs it learns are unrelated to the space’s shape, allowing it to pick up the same motifs from various recordings.
The EPFL team recorded recordings of rats ‘ hippocampus and the pre-motor cortex during a reaching task in order to determine whether MARBLE was effective.
They discovered that MARBLE’s representations based on single-neuron population recordings were much more interpretable than those from other machine learning techniques, and that MARBLE could decode brain activity to arm movements with greater precision than other techniques.
Additionally, MARBLE was able to independently combine brain activity recordings from various experimental conditions into a global structure because it is grounded in the mathematic theory of high-dimensional shapes. This gives it an edge over other methods, which must work with a user-defined global shape.
Brain-machine interfaces and beyond
In addition to improving our understanding of how dynamics affect brain computations and behavior, MARBLE could use neural activity data to create decodable representations that could be used to create assistive robotic devices.
The researchers point out, however, that MARBLE is a potent tool for comparing dynamic phenomena across scientific disciplines and datasets.
According to Vandergheynst,” The MARBLE method is primarily aimed at helping neuroscience researchers understand how the brain functions in different individuals or experimental conditions, and to uncover – when they exist – universal patterns.”
Our tool will be useful for researchers in other life and physical sciences who wish to jointly analyze multiple datasets, but its mathematical foundation is by no means limited to brain signals.
About this news about neuroscience and AI research
Author: Celia Luterbacher
Source: EPFL
Contact: Celia Luterbacher – EPFL
Image: The image is credited to Neuroscience News
Original Research: Open access.
” Neural population dynamics and geometry are interpretable statistical representations.” by Pierre Vandergheynst et al. Nature Methods
Abstract
Neural population dynamics and geometry are interpretable statistical representations.
Low-dimensional manifolds are frequently used to control the dynamics of neuron populations. Therefore, we require methods to interpret interpretable and consistent latent representations learned over neural manifolds.
We describe the unsupervised geometric deep learning method MARBLE, which transforms on-manifold dynamics into local flow fields and maps them into a common latent space.
We find emergent low-dimensional latent representations that parametrize high-dimensional neural dynamics during gain modulation, decision-making, and changes in the internal state in simulated nonlinear dynamical systems, recurrent neural networks, and experimental single-neuron recordings from primates and rodents.
These representations are consistent across neural networks and animals, allowing for a thorough analysis of cognitive computations.
Extensive benchmarking demonstrates state-of-the-art within- and across-animal decoding accuracy of MARBLE compared to current representation learning approaches, with minimal user input.
Our findings support the powerful inductive bias used to create decoding algorithms and integrate data across experiments.
