How Memories are ‘Rotated’ to Avoid Sensory Interference

Post by Lincoln Tracy

What's the science?

Being able to maintain short-term memory of recent stimuli like sights, sounds, and smells is vital to cognition. These memories provide the context for making important decisions and are especially critical in developing predictions about future events. Predictions are based on expectations, which one learns by associating the current stimulus with the memory of what happened previously. Although sensory and memory information both play a key role in cognition, scientists are unsure how the brain incorporates new sensory information and memory representations without interference. This week in Nature Neuroscience, Libby and Buschman used an auditory-based implicit sequence-learning paradigm to explore the brain mechanisms responsible for avoiding interference between sensory and memory representations.

How did they do it?

The authors began by inserting silicon recording arrays into the right auditory cortex of adult male mice. After surgery, mice were exposed to an implicit sequence-learning paradigm involving sequences of four auditory chords over four consecutive days (1,500 sequences per day). Each sequence began with a pair of contextual chords, either A and B (AB context) or X and Y (XY context). The AB and XY context pairs predicted what chord would follow: in two-thirds of cases the AB context was followed by a C chord and the XY context was followed by a C* chord. However, in 20% of cases, the context pair was followed by the other C/C* stimulus (i.e., ABC* and XYC). All sequences ended with the same D chord. Auditory cortex neuronal activity was recorded during the sequence-learning paradigm. The authors then trained linear support vector machine classifiers (a type of machine learning model) to discriminate between neuronal firing rate responses to each pair of stimuli. Each classifier defined an encoding axis; stimulus information in different neuronal populations could be estimated at each moment by projecting the firing rates onto the encoding axes.

What did they find?

First, the authors found that experience over time led to the mice learning the auditory sequences. When the A or X context chords were presented on day 4, there was predictive neural encoding of the expected C or C* stimulus. In addition, the presentation of the C/C* stimulus evoked a response along the A/X sensory axis, a phenomenon known as postdiction. Postdiction is when new information updates the perception of previously experienced events. Second, they found that while the alignment of A/X and C/C* sensory representations allowed for prediction and postdiction, it also led to interference between current and previous sensory inputs and representation. The authors determined that an orthogonal (i.e. rotated to be perpendicular to sensory inputs) memory representation was being created to avoid interference. Third, they found that the orthogonalization of memory representation occurred due to two clusters of conjunctive neurons: stable and switching. Stable neurons maintained contextual preference across the chord sequence while switching neurons switched their A/X contextual preference during the sequence. Finally, using a computational model, the authors found the combination of stable and switching neurons was the most efficient way to rotate sensory representations into orthogonal memory representations, avoiding interference. In other words, memories were rotated to avoid new incoming sensory information that could interfere with memory formation.

lincoln (4).jpg

What's the impact?

This study revealed that the brain avoids interference between sensory and memory representations by rotating the memory representations to become orthogonal to incoming sensory inputs. As mice became familiar with the sequence of sounds, the neural representations of associated stimuli became aligned in the auditory cortex. This alignment explains postdiction, where new information is used to update the perception of previously experienced events. Further work is required to better understand the mechanisms underlying stable and switching neuronal populations. 

Timothy_quote_April20.jpg

Libby and Buschman. Rotational dynamics reduce interference between sensory and memory representations. Nature Neuroscience (2021). Access the original scientific publication here.