Neuroscience

# Fault tolerance in the brain

### Subjects

If stored information is erased from neural circuits in one brain hemisphere in mice, the lost data can be recovered from the other. This finding highlights a safeguarding mechanism at work in the brain. See Article p.459

When we send an e-mail or save a file on our hard drives, information can be lost, owing to dropped data packets or corrupted bits. We typically do not notice such failures because systems are designed with built-in mechanisms to restore the lost data. Dropped packets are retransmitted, and multiple copies of data are saved. The brain also stores and transmits information — is it, too, fault-tolerant? In this issue, Li et al.1 (page 459) report the perturbation of brain activity to erase stored information in mice. They discover that the lost information can be rapidly restored by an unperturbed brain region.

The brain can reorganize itself to restore function after certain types of injury2, but this type of fault tolerance typically takes place over weeks. By contrast, many everyday brain functions, such as putting a name to the face of an acquaintance or hitting a tennis ball, take place on a timescale of seconds or less. Does a fault-tolerance mechanism also operate in neural circuits over these shorter timescales?

Li et al. investigated whether regions present in each of the brain's two hemispheres might act together to produce a rapid back-up system for stored information — a mechanism known as redundancy. Specifically, they tested whether the two premotor cortices of the mouse brain act redundantly to prepare the animal to lick with its tongue in a particular direction, which it has been taught will lead to a reward of water. The authors briefly blocked the activity of premotor neurons in one hemisphere and observed that information about intended licking direction was quickly restored (Fig. 1a). However, if they silenced neurons in both hemispheres in this way, the information was not restored, and so the animal licked left or right at random (Fig. 1b).

These results indicate that, during single-hemisphere silencing, fault tolerance is provided by the unmodified hemisphere. To test this back-up system more directly, the researchers severed the connections between the two hemispheres. When neurons in one hemisphere were silenced in this setting, the information about intended licking direction was not restored.

Next, Li et al. constructed computational network models of neurons in two interacting hemispheres to study how connectivity between the two brain regions enables fault tolerance. In these models, as in the experimental setting, information about movement direction was restored after neuronal silencing in one hemisphere. Together with the experimental evidence, these data suggest that each hemisphere helps the other to restore information about planned movement direction.

Perhaps the most interesting finding in this study is that, after silencing neurons in one hemisphere, not all aspects (called dimensions) of the perturbed neural activity recovered equally. Li and colleagues found that the neural activity that enabled maximal differentiation between left and right licks recovered rapidly. By contrast, other dimensions of neural activity that were not relevant to the task did not always recover. Thus, there was preferential recovery of the dimension that was needed for the animal to succeed at the licking task.

The current study involved both hemispheres controlling a single effector, the tongue. An open question is how these findings apply to brain functions that predominantly involve a single hemisphere, such as control over reaching with one arm. As the authors point out, one possibility is that there are redundant subcircuits within a hemisphere, perhaps spread across multiple brain areas, working together to provide fault tolerance.

Li et al. perturbed neural activity using an optogenetic technique, in which the activity of neurons that harbour light-sensitive ion channels can be modulated using light. This approach allows the silencing or activation of many neurons in unison. To further understand the fault-tolerant properties of neural circuits, more-flexible methods that allow selective activation and silencing of different groups of neurons at different times are needed. Such methods would permit testing of the robustness of a neural circuit to different patterns of perturbation, including those that mimic the random signal disturbances, known as noise, that are a part of normal neuronal signalling3.

The current work demonstrates the power of perturbing neural activity in combination with multidimensional analysis of the activity of a neural population4. By perturbing neural activity in different ways and observing how it recovers, we should be able to gain further insights into fundamental network-level mechanisms that support brain functions5. Advances in methods for perturbing and recording neural activity, for analysing population-wide neural activity and for network modelling are rapidly making such studies possible.Footnote 1

1. 1.

## References

1. 1

Li, N., Daie, K., Svoboda, K. & Druckmann, S. Nature 532, 459–464 (2016).

2. 2

Kolb, B. & Whishaw, I. Q. Prog. Neurobiol. 32, 235–276 (1989).

3. 3

Faisal, A. A., Selen, L. P. J. & Wolpert, D. M. Nature Rev. Neurosci. 9, 292–303 (2008).

4. 4

Cunningham, J. P. & Yu, B. M. Nature Neurosci. 17, 1500–1509 (2014).

5. 5

Brody, C. D., Romo, R. & Kepecs, A. Curr. Opin. Neurobiol. 13, 204–211 (2003).

## Author information

Authors

### Corresponding author

Correspondence to Byron M. Yu.

## Rights and permissions

Reprints and Permissions

Yu, B. Fault tolerance in the brain. Nature 532, 449–450 (2016). https://doi.org/10.1038/nature17886

• Published:

• Issue Date:

• ### An $\ell_0$ -Norm-Based Centers Selection for Failure Tolerant RBF Networks

• Hao Wang
• , Zhanglei Shi
• , Hiu Tung Wong
• , Chi-Sing Leung
• , Hing Cheung So
•  & Ruibin Feng

IEEE Access (2019)

• ### Contributions of Parietal Cortex to the Working Memory of an Obstacle Acquired Visually or Tactilely in the Locomoting Cat

• Carmen Wong
• , Keir G Pearson
•  & Stephen G Lomber

Cerebral Cortex (2018)

• ### Properties and learning algorithms for faulty RBF networks with coexistence of weight and node failures

• Rui-Bin Feng
• , Zi-Fa Han
• , Wai Yan Wan
•  & Chi-Sing Leung

Neurocomputing (2017)