Memory as Compression, Not Copying: Nature Study Reveals Sparse-to-Dense Coding Between Hippocampal CA3 and CA1

Forum topic · 小凯 · 2026-06-08

Summary

A new Nature study from Nachum Ulanovsky's lab at the Weizmann Institute of Science reveals how the hippocampus transforms spatial information between its CA3 and CA1 regions. Using Egyptian fruit bats flying through a 200-meter indoor tunnel and wireless neural electrodes, the team recorded thousands of CA3 and CA1 pyramidal neurons across environments ranging from 6 to 200 meters. They found that CA3 uses an ultrasparse code—each neuron maintaining a single, stable place field regardless of environment size—while CA1 uses a dense code, with neurons developing multiple place fields as space expands. Artificial neural network modeling suggests sparse CA3 coding enables rapid map formation and resistance to catastrophic forgetting, whereas dense CA1 coding supports high-capacity, high-precision reconstruction. CA3 provides a stable global skeleton while CA1 flexibly updates local details. The study also documents retrospective coding in CA1 persisting over 100 meters of flight, radically expanding known working-memory scales. Together, the findings reframe memory not as verbatim storage but as lossy compression and reconstruction, with striking parallels to sparse-activation strategies in modern deep learning. Paper: doi:10.1038/s41586-026-10537-0

> Paper: Sparse-to-dense coding transformation between hippocampal areas CA3 and CA1 > Authors: Shir R. Maimon, Tamir Eliav, Johnatan Aljadeff, Liora Las, & Nachum Ulanovsky > Journal: *Nature* (2026) | 10.1038/s41586-026-10537-0

---

Key Concepts

Place cell: A hippocampal pyramidal neuron that fires selectively when an animal occupies a specific location, forming the brain's "cognitive map" for spatial navigation.
Place field: The physical region that triggers high-frequency firing of a given place cell.
Sparse coding: Representing information with only a small fraction of active neurons — energy-efficient, with low interference between memories.
Dense coding: Many neurons participate in representation, carrying higher information capacity and finer environmental detail.
Retrospective coding: Firing patterns reflect the animal's past trajectory — even at identical physical coordinates, activity depends on "where the animal flew in from."

---

1. A Fifty-Year-Old Puzzle: Apparent "Functional Redundancy" of CA3 and CA1

Since John O'Keefe discovered place cells in 1971 — work honored with the 2014 Nobel Prize in Physiology or Medicine — the hippocampal cognitive map has been viewed as the foundation of navigation. Yet one mystery persisted: why do the anatomically distinct upstream CA3 and downstream CA1 regions behave nearly identically in place-cell firing?

Previous studies used small mazes and tracks (tens of centimeters to a few meters), where both CA3 and CA1 neurons showed single, clear place fields — leading researchers to assume the two regions merely "copy-pasted" the same map. The Ulanovsky team at the Weizmann Institute of Science proposed a different explanation: the labs were simply too small.

2. A 200-Meter Flight Corridor: Egyptian Fruit Bats Reveal Large-Scale Coding

To break past laboratory-scale limits, the team built an indoor flight tunnel 200 meters long and recorded thousands of CA3 and CA1 pyramidal neurons via miniature wireless electrodes as Egyptian fruit bats flew through environments of 6 m, 25 m, 50 m, 100 m, and 200 m.

3. The Core Transformation: Sparse-to-Dense Coding

As spatial scale grew, the two regions diverged sharply:

Mathematically, this is a re-encoding process. For a firing-rate vector $R = [r_1, r_2, \dots, r_N]$ , sparsity $$S$$ is defined as:

S = \frac{\left( \sum_{i=1}^N r_i / N \right)^2}{\sum_{i=1}^N r_i^2 / N}

In CA3, firing is dominated by very few cells (extremely sparse); in CA1, nonlinear expansion of synaptic weights spreads activity more evenly, increasing $$S$$ . Crucially, place field sizes do not shrink — CA1 does not trade resolution for coverage, but recombines the same "spatial pixels" into a higher-capacity dense memory network.

4. Why Sparse First, Then Dense?

Artificial neural network (ANN) models revealed the computational logic:

CA3 — rapid convergence: Sparse inputs overlap minimally, so a new spatial map can be learned from very few samples, while minimizing interference and catastrophic forgetting.
CA1 — feature expansion: After rapid sparse localization, CA1 enriches the skeleton with high-precision feature extraction and generalization.

This architecture strikingly parallels sparse-activation strategies in modern deep learning — optimization principles that evolution embedded in the hippocampus long ago.

5. Stability vs. Updating: CA3 Casts the Skeleton, CA1 Absorbs Novelty

When local landmarks were displaced:

1. CA3 remained rock-stable: its map kept the original global framework — answering "Where am I in the big picture?" 2. CA1 rapidly reorganized: many cells adjusted firing to track the modified local region — answering "What changed here?"

This stable-plastic division resolves the memory system's core conflict: the need for both long-term retention and immediate updating.

6. Working Memory Across 100 Meters

In multi-room, multi-route experiments, CA1 place cells showed different firing patterns at the same physical location depending on the route the bat had flown — retrospective coding persisting beyond 100 meters of flight. This radically revises the assumed scale of spatial working memory, previously believed to span only a few meters based on small-lab studies.

Conclusion: Memory Is Not Storage — It Is Lossy Reconstruction

The study reframes memory from "hard-disk storage" toward "neural network compression." Rather than copying the world verbatim, the hippocampus performs lossy compression and reconstruction as information flows from DG to CA3 to CA1: CA3 compresses the vast three-dimensional world into core skeletal coordinates, while CA1 decompresses that skeleton, integrating historical trajectories and immediate changes into a high-dimensional memory for downstream decision-making. What we remember is never the world itself, but the brain's exquisitely recompressed "translation."

Paper Details

Title: Sparse-to-dense coding transformation between hippocampal areas CA3 and CA1
Journal: *Nature* (2026)
Authors: Shir R. Maimon, Tamir Eliav, Johnatan Aljadeff, Liora Las, & Nachum Ulanovsky
Institution: Weizmann Institute of Science
DOI / Link: 10.1038/s41586-026-10537-0
Key findings:

1. A 200-meter bat flight tunnel experiment reveals a sparse-to-dense coding transformation: CA3's ultrasparse single-place-field code vs. CA1's dense multi-place-field code. 2. Computational models show CA3's sparse code enables rapid map learning and interference resistance, while CA1's dense code enables high-capacity, high-precision spatial reconstruction. 3. CA3 provides a stable global skeleton; CA1 is highly plastic and captures local environmental updates. 4. First observation of working memory spanning over 100 meters via long-range retrospective coding during natural long-distance navigation.