Between-subject prediction reveals a shared representational geometry in the rodent hippocampus

Abstract

How a memory system encodes related experiences has consequences for what operations the system supports. For instance, independent coding enables retention of potentially important idiosyncratic details by reducing interference, but makes it difficult to generalize across experiences. Strikingly, the rodent hippocampus constructs statistically independent representations across environments ("global remapping") and assigns individual neuron firing fields to locations within an environment in an apparently random fashion, processes thought to contribute to the role of the hippocampus in episodic memory. This random mapping implies that it should be challenging to predict hippocampal encoding of a given experience in a one subject based on the encoding of that same experience in another subject. Contrary to this prediction, we find that by constructing a common representational space across rats ("hyperalignment"), we can consistently predict data of "right" trials (R) on a T-maze in a target rat based on 1) the "left" trials (L) of the target rat, and 2) the relationship between L and R trials from a different source rat. These cross-subject predictions outperformed a number of control mappings, such as those based on permuted data that broke the relationship between L and R activity for individual neurons, and those based solely on within-subject prediction. This work constitutes proof-of-principle for successful cross-subject prediction of ensemble activity patterns in the hippocampus. This novel approach provides new insights in understanding how different experiences are structured, and suggests further work identifying what aspects of experience encoding are shared vs. unique to an individual.

Across-animal odor decoding by probabilistic manifold alignment

“…We compared the decoding performance of our method to two state-of-the-art methods: 1) across condition alignment by canonical correlation analysis (CCA) [21,22] and 2) Procrustes alignment with FA-based dimensionality reduction (FA+Procrustes) [8,6]. CCA takes pairs of measures from two animals and projects them into a shared low dimensional manifold so as to maximize the correlation between them.…”

“…4 Instead, they rely on a sequence of independent steps: first a (usually off-the-shelf) dimensionality reduction procedure. Dimensionality reduction is followed by across-condition data alignment within the latent manifold, using approaches such as CCA (and its multivariate generalizations [37,38]), or Procrustes alignment [6,39,8]. Lastly, a common decoder is trained on the aligned across-condition data [21,9,8].…”

“…Lastly, a common decoder is trained on the aligned across-condition data [21,9,8]. These procedures are popular as a replacement for costly and inconvenient BCI re-calibration in prosthetics [40], but their application to across-animal alignment remains rare [7,6]. The multi-step procedure, although formally suboptimal, allows for some degree of flexibility in the individual component steps.…”

“…Collective network dynamics are the foundation of neural computation, from early sensory encoding [1], to working memory [2], decision making [3], or motor control [4]. Neural population activity often has low-dimensional structure that is qualitatively preserved across sessions [3], behavioral states [5], and even animals [6,7]. Aligning neural datasets from multiple recordings and animal subjects into a common latent space could provide a powerful tool for extracting unifying principles of brain computations.…”

Conserved structures of neural activity in sensorimotor cortex of freely moving rats allow cross-subject decoding

“…For the decoding analysis, we divided the six behavioral classes in two disjoint sets: one “align set” that was used to align the neural structures, and one “decode set” that was used for training and testing a classifier. The mean neural vectors (four dimensions) corresponding to the behavioral classes in the “align set” were used to compute a Procrustes transformation between two sessions to align the population activity structures ( 27, 28 ). Procrustes transformations involve translation, scaling, reflection and rotation and thus preserve the shape of a set of points.…”