What is the functional role of adult neurogenesis in the hippocampus

Laurenz Wiskott, Malte J. Rasch, Gerd KempermannCogPrints 2004

Abstract

The hippocampus is a critical brain structure for episodic memory, functioning as an intermediate storage system that captures new input patterns for later transfer to the permanent cortical memory. Within the hippocampal formation, the dentate gyrus (DG) is unique for its ability to generate new neurons throughout an adult's life. This paper explores the functional necessity of adult neurogenesis using a computational framework. We hypothesize that neurogenesis allows the dentate gyrus to adapt to novel environmental statistics without suffering from catastrophic interference—the rapid forgetting of previously learned information. By simulating a linear autoencoder model of the hippocampal loop, we demonstrate that adding new, plastic neurons while maintaining fixed older neurons preserves the encoding of past environments while enabling the learning of new ones. This model aligns with several biological observations, including the age-related decline in neurogenesis and its regulation by exploratory behaviors.

Introduction

The hippocampal formation is organized as a loop: sensory information from the entorhinal cortex (EC) passes through the dentate gyrus (DG) to the CA3 and CA1 regions before returning to the EC. In this circuit, the CA3 is widely regarded as the primary memory substrate due to its recurrent connectivity. The DG serves as an encoder, preparing input patterns for storage, while the CA1 and subiculum act as decoders to reconstruct these patterns.

A striking feature of the DG is adult neurogenesis. While most of the brain’s architecture is fixed shortly after development, the DG continues to integrate new neurons. This raises a fundamental question: why is neurogenesis necessary here and not in other hippocampal substructures? We propose that the DG must solve the "stability-plasticity dilemma." It must remain plastic enough to encode new environments but stable enough to prevent the degradation of existing memories.

The Problem: Catastrophic Interference

In artificial neural networks, adapting to a new data distribution typically requires updating synaptic weights. However, this often leads to catastrophic interference, where the weights optimized for an old task are overwritten by the requirements of a new task, leading to the immediate loss of prior knowledge. While machine learning often circumvents this via "interleaved training" (shuffling old and new data), biological organisms encounter environments sequentially and cannot always revisit old contexts to refresh their memories.

We hypothesize that neurogenesis is the biological solution to this problem. By adding new neurons to the DG to handle qualitatively new input distributions while keeping the connections of older neurons fixed, the brain can expand its encoding capacity without overwriting existing representations.

Methods

To test this hypothesis, we utilized a linear autoencoder network representing the hippocampal loop:

  1. Input Layer: Entorhinal Cortex (EC-II/III).
  2. Hidden Layer: Dentate Gyrus (DG) and CA3, acting as an information bottleneck.
  3. Output Layer: CA1 and subiculum, performing the decoding back to the EC (V/VI).

The network was trained to minimize reconstruction error across different environments (A and B), modeled as high-dimensional Gaussian distributions with different statistical orientations. We compared three strategies for DG adaptation when moving from Environment A to Environment B:

  • No DG-adaptation: Weights are frozen after learning Environment A.
  • Full adaptation: All weights are updated to optimize for Environment B.
  • Neurogenesis: Old units are frozen, but a small number of new, plastic units are added to encode the novel features of Environment B.

Key Results

Our simulations revealed that only the neurogenesis strategy successfully managed the transition between environments:

  • Full Adaptation Fails: When the DG fully adapted to Environment B, the retrieval error for Environment A became massive (error > 1.69). The network effectively "forgot" how to encode the original environment.
  • No Adaptation Fails: If the DG remained frozen, the network could not accurately capture the variance of Environment B. Furthermore, when the decoder (CA1) tried to compensate for the new environment using the old encoding, it ruined its ability to decode Environment A (error increased from 0.30 to 0.61).
  • Neurogenesis Succeeds: By adding just 5 new units to a 15-unit hidden layer, the network achieved low error in both Environment B (0.36) and upon returning to Environment A (0.41). The new neurons effectively learned the dimensions of the new environment that were orthogonal to the old ones.

Biological Consistency

This model provides a functional explanation for several observed phenomena in adult neurogenesis:

  1. Age-Related Decline: New neurons are only required for qualitatively new input distributions. As an animal ages and accumulates experience, the likelihood of encountering a truly novel environment decreases, naturally reducing the demand for neurogenesis.
  2. Low Rates of Growth: Only a small percentage of new neurons are needed to capture the missing dimensions of a new environment. This is consistent with the relatively low level of neurogenesis observed in mammals.
  3. Behavioral Regulation: Neurogenesis is upregulated by physical activity and hunger. Since these behaviors are evolutionary precursors to exploration and the discovery of new environments, they serve as logical "unspecific predictors" for the upcoming need for new encoding units.
  4. Cell Survival: The brain generates neurons in "stock." If no novel environment is encountered, these neurons are not integrated and eventually die, which matches experimental findings on the survival rates of newborn cells.

Conclusion

Adult neurogenesis in the dentate gyrus serves as a vital mechanism for avoiding catastrophic interference. By providing a source of fresh plasticity that does not disrupt established circuits, neurogenesis allows the hippocampus to maintain a lifelong library of diverse environmental encodings. Future research should extend this model to non-linear networks to explore how neurogenesis interacts with the DG's known role in sparse coding and pattern separation.

Summary updated

Abstract

The hippocampus is a brain structure that is instrumental for episodic memory, i.e. for memorizing facts and events. It is often thought to be an intermediate memory that can store new input patterns quickly, which subsequently get transferred into more permanent cortical memory. The entorhinal cortex serves as an interface between hippocampus and other cortical areas. Within the hippocampal formation the different substructures form a loop: entorhinal cortex (layers II/III) - dentate gyrus - CA3 - CA1 - subiculum - entorhinal cortex (layers V/VI). Because of its recurrent connectivity, CA3 is thought to be the actual memory. Dentate gyrus would then be an encoding network, preparing the input patterns for storage in CA3; CA1 and subiculum would perform the decoding to reconstruct the stored patterns in entorhinal cortex. The dentate gyrus is special in that it generates new neurons throughout life, a phenomenon referred to as adult neurogenesis. Why does adult neurogenesis occur in the dentate gyrus and not in any of the other structures? Assume the dentate gyrus adapts to the environment the animal lives in in order to optimize its encoding for the input-pattern distribution encountered in this environment. If the animal moves to another environment, new adaptation takes place and the dentate gyrus is faces with the problem of catastrophic interference. As a new encoding is learned the old encoding degrades quickly and as a consequence old patterns could not be addressed and retrieved from the CA3-memory. In artificial neural networks catastrophic interference is usually avoided by interleaved training, i.e. the training patterns are presented repeatedly in an alternating fashion. However, this is not possible in real life, because many patterns occur only once. How can the dentate gyrus solve this problem? We hypothesize that new neurons are the solution to this problem. If the dentate gyrus keeps old neurons and their connections fixed but adds new neurons that are plastic, it can adapt to qualitatively new input patterns but at the same time maintain the encoding capabilities for old patterns. Note that new neurons are required only for qualitatively new patterns and not for new patterns that belong to the old input distribution, because we assume the encoding to be characteristic for a distribution and not for individual patterns. As a proof of principle we have simulated a linear auto-encoder network modelling the loop within the hippocampal formation. We assume that the animal first lives in environment A with a certain input-pattern statistics, then moves to a new environment B with a different input-pattern statistics, and finally returns to environment A. We assume that the animal has time to adapt to environment A and then B, but when it returns to A we only test the performance without giving the time for new adaptation. We also assume that the decoding (CA1/subiculum) stays plastic and can adapt to environment A and B in any case (but not when the animal returns to A). We have considered three different scenarios: (a) No DG-adaptation: The dentate gyrus adapts to environment A and keeps the synaptic weights fixed after that. No adaptation to environment B occurs. (b) Neurogenesis: The dentate gyrus starts with fewer units and first adapts to environment A. In environment B the old units and connections are fixed but a few new units are added and used to adapt to the new input-pattern distribution. (c) Full adaptation: The dentate gyrus always fully adapts to the current environment, first A then B. In the simulations we find that the networks always perform reasonably well on the pattern distributions they are adapted to. However, in scenario (a) the performance is poor in environment B, because although the decoding can adapt to environment B the encoding is still optimal for A and misses important dimensions of B. Performance is also poor when the animal returns to A, because the decoding has adapted to B. In scenario (c) performance is particularly poor when the animal returns to environment A, because the network is then fully adapted to B. Only in scenario (b) is the effect of catastrophic interference largely avoided and the performance good in environments A and B and also as the animal returns to A. Our model is consistent with a number of anatomical and physiological facts: New neurons are found to be more plastic than old ones as required by our model. Since new units are only required for qualitatively new input patterns, there is decreasing need for neurogenesis with age, because the animal has more and more experience and encounters fewer and fewer qualitatively new stimuli. This is consistent with the decrease in neurogenesis observed experimentally. A relatively small number of newly added units can have a large effect, since only missing dimensions have to be newly encoded. This is consistent with the relatively low level of neurogenesis of 30 new neurons in mice over the whole lifetime. Since the generation of new neurons takes weeks but the demand for new neurons can be on a much shorter time scale when the animal changes its environment, it is reasonable that new neurons are generated all the time to have some in stock when needed. New neurons not needed die after some time. This is what is found experimentally. The level of neurogenesis is regulated by rather unspecific factors such as physical activity or hunger. It is clear that if new neurons are only needed if qualitatively new input-pattern distributions are encountered, there are no specific factors that would be available earlyenough. Thus the unspecific factors might actually be fairly descent predictors for the need of new neurons, since hunger andrunning fosters exploration of new environments. In summary we hypothesize that adult neurogenesis in the dentate gyrus helps to solve the problem of catastrophic interference when an animal adapts to new environments. Our network simulations confirm that adding new neurons can reduce the effect of catastrophic interference significantly. The model is also qualitatively consistent with a number of anatomical and physiological facts about adult neurogenesis. See http://cogprints.org/4012/ for more information.
BibTeX
@inproceedings{wiskott2004what,
  author = {Laurenz Wiskott and Malte J. Rasch and Gerd Kempermann},
  title = {{What is the functional role of adult neurogenesis in the hippocampus}},
  booktitle = {CogPrints},
  year = {2004},
}
Figure