Motor Control: Sensory feedback can give rise to neural rotations
Investigating how an artificial network of neurons controls a simulated arm suggests that rotational patterns of activity in the motor cortex may rely on sensory feedback from the moving limb.
- Ming Hsieh Department of Electrical and Computer Engineering, Viterbi School of Engineering, University of Southern California, United States
- Department of Biomedical Engineering, Viterbi School of Engineering, University of Southern California, United States
- Neuroscience Graduate Program, University of Southern California, United States
Each time you move your arm, populations of neurons in the motor cortex perform an intricate and coordinated dance that leads to the generation of movement. A large portion of this coordinated activity can be described as having a ‘rotational pattern’ over time, which is often not directly visible in the neural activity and is uncovered using dimensionality reduction methods such as principal component analysis (Figure 1A; Churchland et al., 2012). Such rotational dynamics have been observed in the motor cortex in many studies involving arm reaching or reach and grasp movements (see, for example Kao et al., 2015; Pandarinath et al., 2015; Suresh et al., 2020; Susilaradeya et al., 2019; Abbaspourazad et al., 2021; Sani et al., 2021), but were absent in the supplementary motor area (Lara et al., 2018).
Figure 1
Using an artificial network to investigate how rotational patterns are generated in the motor cortex.
(A) The brain and the arm together can be viewed as a closed-loop feedback control system. When the brain receives instructions for a task, neurons in the motor cortex (red inset) send a command to the arm, which moves and returns sensory information back to the cortex. During arm movements, the activity of neurons in the motor cortex exhibits rotational patterns, which may not be visible directly, but usually emerge after neural activity (red graph) has been subjected to dimensionality reduction methods and averaged across several repetitions of the same movement (different movements are shown with different colors). (B) A similar closed-loop system can be constructed in simulations with an artificial neural network (magenta, left) replacing the brain and a musculoskeletal model (right) replacing the arm. Kalidindi et al. show that such a system generates rotational patterns in the artificial neural network that resemble those observed in the motor cortex, regardless of the presence or absence of recurrent connections (purple).
Several studies have investigated the mathematics behind how these rotational patterns arise from the activity of individual neurons (e.g., Michaels et al., 2016; Elsayed and Cunningham, 2017). Some of these reports (Sussillo et al., 2015; Michaels et al., 2016) rely on artificial neural networks: simplified computational representations of interconnected neurons, or groups of neurons, that allow researchers to study how patterns of activity in the brain may emerge. These networks make it possible to explore how the nervous system might perform certain tasks without using real brains, which are harder to observe and difficult or impractical to manipulate.
One question that remains a topic of lively investigation is whether rotational patterns of activity are generated autonomously within the motor cortex itself or whether they reflect ongoing inputs from other regions of the brain (Vyas et al., 2020). It is known that networks with recurrent connections – this is, networks with ‘memory’, in which the output can be affected by previous inputs – can generate patterns autonomously. Indeed, a group of researchers discovered that when they trained a recurrent artificial network to generate the motor activities needed for movement, the patterns resembled the rotational activity seen in the motor cortex (Sussillo et al., 2015). Now, in eLife, Hari Teja Kalidindi (Scuola Superiore Sant'Anna), Kevin P Cross (Queen's University) and colleagues report that it is also possible to train a neural network to control an artificial arm without using any recurrent connections inside the network (Kalidindi et al., 2021).
The team (who are based in Italy, Canada, the United Kingdom and the United States) constructed an artificial neural network that can activate muscles on a simulated arm which then sends sensory information, such as its position and muscle activations, back into the network (Figure 1B). The artificial neural network was trained to perform reaching movements or to counter disturbances from the environment, such as forces that suddenly pushed the arm aside. This approach replicated key findings from non-human primate experiments in that the activity of neurons in the brain showed rotational patterns, whereas muscle activations in the arm did not. Intriguingly, Kalidindi et al. also found that when they stripped recurrent connections from the neural network, it could still learn to move the artificial arm, and still generated rotational patterns in its activity.
So how can a network with absolutely no recurrent connections produce rotational patterns similar to those observed in the motor cortex? Since non-recurrent networks always return the same output when they receive a specific input, the only way they can produce patterns that vary over time is if the input to the network also changes over time. Indeed, Kalidindi et al. found that the sensory feedback signals from the arm, which act as the input to the neural network, also show rotational patterns. Thus, feedback from the arm’s position and from muscle activations is sufficient to generate rotational patterns of activity in the brain. When Kalidindi et al. repeated the experiment with monkeys performing the same tasks, they observed rotational patterns not only in the motor cortex, but also in the somatosensory cortex, the region of the brain that receives and processes sensory information from the environment. This suggests that sensory feedback to the real brain may also contain rotational dynamics, as was the case in the artificial network simulations.
Another important consideration is the structure of the behavioral task. Usually, the tasks used to study the activity of the motor cortex involve the arm being stationary at an initial position and ending up stationary in another. So, could the rotational patterns observed in these tasks be due to the movement starting and ending at the same zero-velocity state? To investigate, Kalidindi et al. trained neural networks to perform a new task in which the arm continuously tracks a target moving at a constant velocity. Both recurrent and non-recurrent networks performed well, but this experiment led to substantially less rotational dynamics than the previous tasks, suggesting that the design of the behavioral task can play a critical role in the prominence of rotational patterns in neural activity.
The work of Kalidindi et al. cleverly uses artificial neural networks and real-world data to highlight the importance of studying the motor cortex in the context of the entire closed feedback loop between the brain and the body, and the need to study this system using different types of tasks. Critically, the experiments suggest that rotational patterns observed in the motor cortex can be influenced not only by internal autonomous dynamics, but also by external inputs such as sensory feedback. An important future research direction is to tease apart the extent to which these two contributing factors influence neural activity in the motor cortex.
References
-
Structure in neural population recordings: an expected byproduct of simpler phenomena?Nature Neuroscience 20:1310–1318.https://doi.org/10.1038/nn.4617
-
A neural network that finds a naturalistic solution for the production of muscle activityNature Neuroscience 18:1025–1033.https://doi.org/10.1038/nn.4042
-
Computation through neural population dynamicsAnnual Review of Neuroscience 43:249–275.https://doi.org/10.1146/annurev-neuro-092619-094115
Article and author information
Author details
Publication history
- Version of Record published: December 21, 2021 (version 1)
Copyright
© 2021, Sani and Shanechi
This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.
Metrics
-
- 2,205
- views
-
- 215
- downloads
-
- 3
- citations
Views, downloads and citations are aggregated across all versions of this paper published by eLife.
Download links
A two-part list of links to download the article, or parts of the article, in various formats.
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Motor Control: Sensory feedback can give rise to neural rotations
eLife 10:e75469.
https://doi.org/10.7554/eLife.75469
Further reading
-
- Computational and Systems Biology
- Developmental Biology
Organisms utilize gene regulatory networks (GRN) to make fate decisions, but the regulatory mechanisms of transcription factors (TF) in GRNs are exceedingly intricate. A longstanding question in this field is how these tangled interactions synergistically contribute to decision-making procedures. To comprehensively understand the role of regulatory logic in cell fate decisions, we constructed a logic-incorporated GRN model and examined its behavior under two distinct driving forces (noise-driven and signal-driven). Under the noise-driven mode, we distilled the relationship among fate bias, regulatory logic, and noise profile. Under the signal-driven mode, we bridged regulatory logic and progression-accuracy trade-off, and uncovered distinctive trajectories of reprogramming influenced by logic motifs. In differentiation, we characterized a special logic-dependent priming stage by the solution landscape. Finally, we applied our findings to decipher three biological instances: hematopoiesis, embryogenesis, and trans-differentiation. Orthogonal to the classical analysis of expression profile, we harnessed noise patterns to construct the GRN corresponding to fate transition. Our work presents a generalizable framework for top-down fate-decision studies and a practical approach to the taxonomy of cell fate decisions.
-
- Computational and Systems Biology
- Genetics and Genomics
We propose a new framework for human genetic association studies: at each locus, a deep learning model (in this study, Sei) is used to calculate the functional genomic activity score for two haplotypes per individual. This score, defined as the Haplotype Function Score (HFS), replaces the original genotype in association studies. Applying the HFS framework to 14 complex traits in the UK Biobank, we identified 3619 independent HFS–trait associations with a significance of p < 5 × 10−8. Fine-mapping revealed 2699 causal associations, corresponding to a median increase of 63 causal findings per trait compared with single-nucleotide polymorphism (SNP)-based analysis. HFS-based enrichment analysis uncovered 727 pathway–trait associations and 153 tissue–trait associations with strong biological interpretability, including ‘circadian pathway-chronotype’ and ‘arachidonic acid-intelligence’. Lastly, we applied least absolute shrinkage and selection operator (LASSO) regression to integrate HFS prediction score with SNP-based polygenic risk scores, which showed an improvement of 16.1–39.8% in cross-ancestry polygenic prediction. We concluded that HFS is a promising strategy for understanding the genetic basis of human complex traits.
