Publications

The remapping of spatial firing fields of cells in the hippocampal formation between rooms is a well-known and extensively studied phenomenon; however, the organizing principle for this remapping is unknown. While CA1 shows random remapping, regions such as the subiculum and medial entorhinal cortex (MEC) display a more organized transformation between rooms. Still, the structure of this transformation is unknown. To study this, we used high-density chronic Neuropixels recordings of the subiculum and MEC in freely moving mice and investigated the structure of the transformation at the population level. We implemented an advanced decoder to estimate the existence and characteristics of the remapping transformation. We deliberately designed the decoder as non-linear and time-dependent to capture rich and complex transformations. To our surprise, we discovered that the ensemble activity often underwent a simple, smooth, low-dimensional transformation captured by an affine transformation (i.e., rotation, scaling, and shear). The population code is thus flexibly adapted to a new context while retaining a stable spatial representation at the network level. This principle was reproduced across the subiculum and the MEC, as well as across spatial cell types such as border, head direction, and other spatially modulated cells. These results provide new insights into the computational principles of hippocampal formation remapping, suggesting that spatial cognition is subserved by adaptive ensemble codes governed by a simple affine coordinate transformation. Our findings establish population structure as a critical organizing principle for spatial memory, suggesting avenues for decoding spatial information in unvisited environments by embedding transformation rules into population-level decoders. We thus demonstrate for the first time a simple organizing principle for the representation of the transformation between diverse environments in the hippocampal formation.

Shai Abramson, Daniel Zur, Gaya Tzadok, Shachaf Kolan, Shiladitya Laskar, Ohad Rechnitz, Vijay Balasubramanian, Genela Morris, Hadas Benisty, Dori Derdikman (Preprint)

Unsupervised feature selection (FS) is essential for high-dimensional learning tasks where labels are not available. It helps reduce noise, improve generalization, and enhance interpretability. However, most existing unsupervised FS methods evaluate features in isolation, even though informative signals often emerge from groups of related features. For example, adjacent pixels, functionally connected brain regions, or correlated financial indicators tend to act together, making independent evaluation suboptimal. Although some methods attempt to capture group structure, they typically rely on predefined partitions or label supervision, limiting their applicability. We propose GroupFS, an end-to-end, fully differentiable framework that jointly discovers latent feature groups and selects the most informative groups among them, without relying on fixed a priori groups or label supervision. GroupFS enforces Laplacian smoothness on both feature and sample graphs and applies a group sparsity regularizer to learn a compact, structured representation. Across nine benchmarks spanning images, tabular data, and biological datasets, GroupFS consistently outperforms state-of-the-art unsupervised FS in clustering and selects groups of features that align with meaningful patterns.

Shira Lifshitz, Ofir Lindenbaum, Gal Mishne, Ron Meir and Hadas Benisty

(Accepted to AAAI, preprint)

Understanding how patterns of neural interactions evolve over time is a central challenge in neuroscience. Traditional analyses often focus on how neural activity changes, while assuming that the correlations between neurons or brain regions remain static. Yet, in reality, these interactions are highly dynamic, reflecting ongoing processes such as learning, attention, or adaptation.

Here, we introduce a new computational framework for examining the temporal dynamics of functional connectivity. Our approach represents each time point as a correlation matrix, capturing how neural elements co-fluctuate, and then analyzes how these matrices change over time using a geometry-aware, multi-resolution method. This framework allows us to separate slow and fast components of network reorganization and to identify specific sub-networks that drive these changes.

We demonstrate the utility of our approach on large-scale population recordings collected via various modalities such as Calcium imaging, electrophysiology and EEG, spanning from local dendritic signals to large-scale cortical networks. Across these diverse systems, our method uncovers meaningful patterns of connectivity dynamics that correspond to behavioral changes and learning-driven reorganization. These results demonstrate our method as a principled and interpretable framework for studying the internal mechanisms of temporal evolution of neural networks.

Yonatan Kleerekoper, Mohammad Kurtam, Yonatan Keselman, Shai Abramson, Dori Derdikman, Yitzhak Schiller, Simon Musall, and Hadas Benisty (Preprint)

The primary motor cortex (M1) is crucial for motor skill learning. Previous studies demonstrated that skill acquisition requires dopaminergic VTA (ventral-tegmental area) signaling in M1, however little is known regarding the effect of these inputs at the neuronal and network levels. Using dexterity task, calcium imaging, chemogenetic inhibiting, and geometric data analysis, we demonstrate VTA-dependent reorganization of M1 layer 2-3 during motor learning. While average activity and average functional connectivity of layer 2-3 network remain stable during learning, activity kinetics, correlational configuration of functional connectivity, and average connectivity strength of layer 2-3 neurons gradually transform towards an expert configuration. Additionally, sensory tone representation gradually shifts to success-failure outcome signaling. Inhibiting VTA dopaminergic inputs to M1 during learning, prevents all these changes. Our findings demonstrate dopaminergic VTA-dependent formation of outcome signaling and new connectivity configuration of the layer 2-3 network, supporting reorganization of the M1 network for storing new motor skills.

Amir Ghanayim*, Hadas Benisty,*, Avigail Cohen-Rimon, Sivan Schwartz, Sally Dabdoob, Shira Lifshitz, Ronen Talmon and Jackie Schiller

Nature Communications, 2025

We propose a novel architecture for learning the importance of each input variable for prediction of the target variable(s) as a function of a given context. Our extensive benchmark shows that c-STG improves feature selection, enhances prediction accuracy as well as model interpretability across multiple real-world domains.

ICML 2024