top of page

FIRE, FEED, AND FORGE: Insights into the representational drift

Neuroletter, Volume 2 Issue 1

Let’s start with a simple philosophical question - Are you the same you from, say, ten years ago, having undergone significant changes in thoughts, feelings, and physique?

The answer might be a yes or no, depending on how one defines the persistence of identity. What sense does the word ‘representation’ make when it’s used in the context of defining your identity? That is, how the thoughts, feelings, and actions still represent the ‘you’ through a course of change or drift.

If this is too deep into philosophy, let’s go beyond the meta-understanding of representational drift and look into some evidence reported by neuroscientists. Recent studies were conducted at Columbia University on olfactory representation in mice by recording the activity of neurons from the piriform cortex - a brain region involved in identifying smells, while they were made to sniff the same odors for several days and weeks. These trials orchestrated by neuroscientists Schoonover and Fink showed surprising results. The group of neurons firing to encode smelling of the particular odor had a shift in their makeup over time. That is, neurons that represented a specific smell on day 1 of the trial and neurons that represented the same smell on day n of the trial were found to be completely different.

To understand this conclusion in terms of probability, they showed that if a neuron in the piriform cortex fires at a particular smell, the odds that it will do so after 30 days is just one in 15. These ongoing changes in representations of task-related stimuli or correlated neural activity for a particular behavior were termed ‘representational drift’.

This evidence, or more so a timely de-correlation of neural representations and behavior, was also found in other areas of the brain. Experiments by Stanford neuroscientist Driscoll looked at neural activity in rodents while they made a spatial navigation choice in a T-shaped maze. These suggested a representational drift in the posterior parietal cortex (part of the hippocampus) that encodes for spatial reasoning. Particularly, it made sense for representational drift to happen in the hippocampus as it is involved in learning and short-term memory and hence, is expected to overwrite itself.

However, for sensory hubs like the piriform cortex, it is surprising as to make the surrounding stimuli familiar, these representations are expected to be stable in such areas. This suggests that the reorganization of neuronal responses in circuits essential for specific tasks is a fairly common phenomenon in the brain, even after stability in behavior is achieved. That is, drift could be a rule rather than an exception for the brain.

Intuitively, it makes sense for drift to be beneficial in storing information that is continuously useful, which is a manifestation of learning. It is also a biological necessity because brains must be robust to failure in individual neurons and to environmental perturbations.

A possible theoretical understanding of the drift could be a high-dimensional representation of inherently low-dimensional tasks. To simplify, the continuous reorganization of neural representations is encoded in a high dimensional population activity space, which is then projected by the brain in a low dimensional space, thereby encoding the same stable behavior through a drift of representations in time.

For instance, in the Driscoll experiments, the population activity space has many degrees of freedom for the representational structure of spatial activity to move in. Yet, a 2D projection of this population space computationally preserves the topology and local structure of the T shape, thus mapping the navigational trajectories for the rodent. This indicates the importance of understanding drift as a widely-distributed activity examined through the identification of correct coding dimensions relative to a global population of neurons.

These could be the possible theoretical insights into how the brain can encode a stable behavior through representational drift. In simple terms, it corrects for it, as mentioned above, and through a system of errors and predictions. Thereby, drift has the potential to drive the brain towards optimal solutions via sampling through a large pool of possibilities.

However, as rightly pointed out by Krakauer, a neuroscientist at John Hopkins University, it must be noted that stability in behavior does not make it a neural representation. It is a correlation found by observers. So, the premise of understanding drift in the brain itself is prone to drift.

Guess we are back to philosophy, and our brain is engulfed within its own little paradox akin to the Ship of Theseus.


1. ge-this-much/619145/




Author: Shreyas Gadge


bottom of page