This chapter explains the "simulation-selection model" centered on the hippocampus's memory and future imagination functions, detailing the roles of the CA3 and CA1 regions, how their neural circuits operate, and how this model differs from existing theories. It also draws parallels with the "Dyna" algorithm in artificial intelligence, presenting new perspectives on memory consolidation and reward-based learning. This is essential for understanding how the hippocampus processes information not only about the past but also to prepare for valued futures.

1. The Accumulation of Hippocampal Research and Basic Premises

The hippocampus has been studied across numerous dimensions — memory, imagination, neurological disease — with over 170,000 papers published. Recent research reveals that the hippocampus does more than just remember the past; it also plays a role in imagining and preparing for the future.

"The hippocampus is involved not only in remembering past experiences but also in imagining future events."

Particularly important are "sharp-wave ripples" and neural replay occurring during sleep or rest, as well as the detailed structure and function of the hippocampal CA3 and CA1 subregions.

2. The Core of the Simulation-Selection Model

The simulation-selection model described in this chapter has the following basic structure:

• Simulation: The CA3 region randomly generates sequential series of events — not only past experiences but also novel (never-experienced) scenarios.
• Selection (value-based): The CA1 region preferentially strengthens and favors sequences with particularly high reward (value) from among these various sequences.

"CA3 generates (simulates) various event sequences during rest and sleep based on massive recurrent connections, and CA1 strengthens (selects) neural activity according to value."

Because the two networks are separated with distinct roles, flexible and efficient preparation for various future scenarios becomes possible — this is the key insight.

3. CA3's Role as 'Simulator'

The idea that CA3 serves as a simulator is supported by the following evidence:

• CA3 has strong recurrent projections, allowing the activation of some neurons to trigger chain activation of others.
• In practice, CA3 generates place cell activation patterns for novel paths that have never been experienced, which is well observed when animals rest or sleep.

"The distinguishing feature of CA3 is that it has densely interconnected, individually weak synapses. This means randomness (stochasticity) can play a functionally important role."

In other words, CA3 is not merely a high-fidelity memory storage device but operates as a simulator that "experiments" with diverse future situations using randomness.

4. CA1's Role as 'Value-Based Selector'

How are the sequences generated by CA3 processed? Many researchers expected they would be handled by brain regions associated with "value" (ventral striatum, orbitofrontal cortex, etc.), but the authors emphasize that CA1 itself is the expert in value judgment.

• In CA1, the reactivation of place cells located along high-reward (value) paths is much more prominent
• Both experimental rats and humans show specific hippocampal patterns that are replayed more frequently and more strongly when rewards are present
• CA1 preferentially passes and strengthens high-value replay sequences from among the many generated by CA3

"CA1 preferentially transmits and strengthens high-value replay sequences, and this selection process guides optimal future choices."

In summary, CA3 throws out many possibilities, and CA1 prioritizes learning only the scenarios judged to have the greatest reward.

5. The Role of the Dentate Gyrus

While the simulation-selection model focuses on CA3 and CA1, another hippocampal region — the dentate gyrus — is also important.

• Traditionally regarded as performing pattern separation (making similar input patterns distinct)
• However, the authors and colleagues argue that the dentate gyrus's primary role is "binding" — combining diverse sensory signals into a "spatial context"

"The dentate gyrus combines diverse sensory signals to help us recognize where we are (spatial context), and CA3/CA1 perform simulation-selection within that context."

In other words, the hippocampus prepares high-value choices appropriate to the spatial context through a sequence of information processing: binding → simulation → selection.

6. Implications of the Simulation-Selection Model and Differences from Existing Theories

While this model has not yet been fully experimentally verified, it parsimoniously explains several unresolved phenomena:

• Why the hippocampus handles both memory and imagination simultaneously
• Why memories are easily distorted (constructive nature)
• Why the hippocampus represents value signals
• Why CA1 is needed separately from CA3
• Why place cell properties are both common to and differentiated between the two regions

"The simulation-selection model explains the neural processes of goal-directed behavior and memory consolidation through a single simple mechanism."

The most important point is that this model presents a new perspective in which spatial information and value information are integratively processed within the hippocampus to optimize future choices.

7. Parallel with the Dyna Algorithm: Where AI Meets Biology

Interestingly, this simulation-selection process closely parallels the "Dyna" reinforcement learning algorithm in AI.

• Dyna combines value learning from actual environment interaction (experience) with value learning through internal simulation
• Example: A robot vacuum both cleans in the real world and performs "imagined cleaning," using the absorbed information to find the best cleaning method faster
• If new situations keep arising (e.g., furniture rearrangement) or learning speed is slow, relying solely on real-world experience can never yield the optimal solution → Simulation-selection is the answer!

"Competitors (predators) won't wait for you to finish learning. The environment is always changing. The simulation-selection process can dramatically accelerate learning."

Ultimately, combining actual behavior with internal imagination (simulation) for fast, flexible adaptation is a critical principle in both the brain and artificial intelligence.

Conclusion

This chapter explains the hippocampus's dual mission of "past memory + future preparation" through the unified brain circuit model of "simulation-selection." CA3's random simulation, CA1's value-based selection, and the dentate gyrus's sensory binding function work together to demonstrate that we can systematically prepare for rewards in futures we have never experienced. The fact that this principle deeply connects with the latest learning theories in AI carries significant implications for both neuroscience and machine learning going forward.