Life as Self-Maintaining Systems — Article 2 of 5
Introduction: The Organ That Rewrites Itself
No engineered system on Earth rewrites its own source code as aggressively, as continuously, and as successfully as the human brain. Every second of waking life — and a good deal of sleep — the brain generates predictions about what will happen next, compares those predictions against incoming sensory data, computes the mismatch, and uses that mismatch to update its own wiring. This is not merely a loose metaphor. Modern machine learning and control theory reuse several of these ideas, even if the mapping is not one-to-one.
The brain is, in the most literal sense, a recursive self-improving system (再帰的自己改善システム). It improves its own capacity to improve. Understanding how it accomplishes this feat — and where it fails — offers a concrete design specification for building artificial agents that can evolve safely under governance constraints.
Predictive Coding: The Brain's Core Algorithm
The predictive coding framework, developed by Rajesh Rao and Dana Ballard in 1999 and extended into a grand unified theory by Karl Friston, proposes that the cortex is organized as a hierarchical generative model. Each level of the cortical hierarchy maintains a model of the level below it and sends top-down predictions. The lower level compares these predictions against its own activity and sends back only the prediction error — the residual that the higher level failed to anticipate.
This architecture has a profound implication: the brain does not passively receive sensory data. It actively hallucinates the world and then checks its hallucinations against reality. Perception is controlled hallucination (制御された幻覚), a phrase coined by neuroscientist Anil Seth. What you experience as 'seeing' is actually your brain's best guess about the causes of the photon patterns hitting your retinas, continuously corrected by error signals propagating upward through the visual hierarchy.
The computational efficiency of this scheme is remarkable. Instead of transmitting the full high-bandwidth sensory stream up the hierarchy, each level transmits only the surprise — the information that the higher level did not already predict. This is analogous to delta encoding in data compression. The brain achieves extraordinary perceptual richness while minimizing the metabolic cost of neural communication.
Hierarchical Prediction Errors
Prediction errors propagate both upward and laterally. A low-level prediction error in V1 (primary visual cortex) might signal an unexpected edge orientation. This error propagates to V2, which attempts to explain it within its model of textures and surfaces. If V2 cannot explain the error, the residual propagates further to V4 and the inferotemporal cortex, where it might trigger the recognition of a novel object.
At each level, the system faces the same decision: can this error be absorbed by updating the current model's parameters, or does the model's structure need to change? This distinction — parameter update versus architecture update — maps directly onto the difference between fine-tuning and retraining in machine learning, and onto the difference between self-repair and evolution in the MARIA VITAL framework.
Dopamine and Reward Prediction Error
If predictive coding describes how the brain models the sensory world, the dopamine system describes how it models value. In a landmark series of studies beginning in the 1990s, Wolfram Schultz demonstrated that midbrain dopamine neurons encode not reward itself but reward prediction error (報酬予測誤差) — the difference between expected and received reward.
When a monkey receives an unexpected juice reward, dopamine neurons fire a burst of activity. When the monkey learns to predict the reward from a preceding cue, the dopamine burst shifts from the reward to the cue. When an expected reward is omitted, dopamine activity dips below baseline — a negative prediction error. This is mathematically identical to the temporal difference (TD) learning signal used in reinforcement learning, a correspondence first noted by Read Montague, Peter Dayan, and Terrence Sejnowski in 1996.
The dopamine system thus implements a continuous A/B test on the brain's own value model. Every outcome is compared against expectation. Positive prediction errors (better than expected) strengthen the associations that led to the action. Negative prediction errors (worse than expected) weaken them. The system does not need an external supervisor; the error signal is generated internally, from the discrepancy between the brain's own predictions and the world's response.
The Exploration-Exploitation Tradeoff
Dopamine also modulates the balance between exploitation (using the current best policy) and exploration (trying new actions to discover potentially better policies). Tonic dopamine levels — the background firing rate — appear to encode a kind of average reward rate. When tonic dopamine is high, the organism exploits; when it is low, the organism explores. This is a biological implementation of the epsilon-greedy or softmax exploration strategies used in reinforcement learning.
The relevance to agent governance is immediate. An agent that only exploits becomes brittle — it cannot adapt to changing environments. An agent that only explores never converges on reliable behavior. The brain's dopamine system solves this problem by dynamically adjusting the exploration rate based on the recent history of prediction errors. MARIA VITAL's Evolution Lab faces the same tradeoff: how aggressively should an agent mutate its own configuration? The biological answer is 'proportionally to recent surprise.'
Synaptic Plasticity: Weight Updates in Biological Hardware
The prediction errors computed by cortical and dopaminergic circuits are translated into physical changes in the brain through synaptic plasticity (シナプス可塑性). The foundational principle, first articulated by Donald Hebb in 1949, is often summarized as 'neurons that fire together wire together.' Modern neuroscience has refined this into a family of plasticity rules:
Long-term potentiation (LTP) strengthens synaptic connections when presynaptic and postsynaptic activity are temporally correlated. This is the biological mechanism underlying associative learning — the strengthening of connections that contribute to accurate predictions.
Long-term depression (LTD) weakens synaptic connections when activity is uncorrelated or anti-correlated. This prunes connections that contribute to prediction errors, gradually sculpting the network toward a more accurate model of its input statistics.
Spike-timing-dependent plasticity (STDP) adds temporal precision: if a presynaptic neuron fires just before a postsynaptic neuron, the synapse is strengthened; if the order is reversed, the synapse is weakened. This implements a causal inference rule — the brain preferentially strengthens connections that reflect cause-and-effect relationships in the world.
Metaplasticity — plasticity of plasticity — adjusts the threshold for LTP and LTD based on the neuron's recent activity history. A neuron that has been highly active becomes harder to potentiate further, preventing runaway excitation. This is the biological equivalent of adaptive learning rate schedules in gradient descent.
Together, these mechanisms mean that the brain is updating its own weights continuously, driven by internally generated error signals, with built-in safeguards against instability. It is a self-improving system with governance constraints baked into the biophysics.
The Cerebellum as Forward Model
While the cerebral cortex handles high-level prediction and the dopamine system handles value estimation, the cerebellum (小脳) implements rapid, precise forward models for motor control. When you reach for a coffee cup, the cerebellum predicts the sensory consequences of the motor command — what your arm will feel like in 200 milliseconds — and compares this prediction against actual proprioceptive feedback.
If the prediction is accurate, the movement proceeds smoothly. If there is a mismatch — the cup is heavier than expected, or the table has shifted — the cerebellum computes a correction signal and sends it to the motor cortex within tens of milliseconds. This is a closed-loop controller with an internally generated reference signal, operating at a timescale too fast for conscious awareness.
The cerebellum's climbing fiber inputs, originating from the inferior olive, are widely believed to carry the error signal that drives cerebellar learning. Each climbing fiber fires at most once or twice per second, delivering a powerful, all-or-nothing teaching signal that updates the weights of the parallel fiber synapses onto Purkinje cells. This architecture — a slow, high-magnitude error signal updating a fast, high-throughput forward model — is strikingly similar to the relationship between offline evaluation (slow, expensive, thorough) and online inference (fast, cheap, approximate) in production ML systems.
Sleep as Batch Processing
The brain does not only learn online. Sleep provides a critical offline processing window during which the day's experiences are replayed, consolidated, and integrated into long-term memory. During slow-wave sleep (深い睡眠), hippocampal place cells replay sequences of activity corresponding to recent experiences, but at compressed timescales — up to 20 times faster than the original experience.
This replay is not a passive recording. The hippocampus selectively replays experiences associated with high prediction error or high reward, prioritizing the consolidation of surprising or valuable information. Meanwhile, synaptic homeostasis theory, proposed by Giulio Tononi and Chiara Cirelli, suggests that sleep globally downscales synaptic weights, counteracting the net potentiation that accumulates during waking learning. This renormalization prevents saturation and restores the signal-to-noise ratio.
During REM sleep, the brain appears to engage in a different kind of processing — testing the generalization of learned models by generating novel combinations of stored experiences. Dreams, in this framework, are the brain's unit tests: synthetic scenarios that probe the robustness of recently updated models.
The engineering parallel is clear. Production systems need both online learning (processing data as it arrives) and offline batch processing (retraining on curated datasets, running regression tests, pruning stale parameters). The brain implements both, with sleep serving as the scheduled maintenance window.
What Transfers Cleanly to Tier-2 Agent Design
Not every biological detail should be copied into software. What transfers cleanly is the control structure. The brain suggests four design rules for recursive self-improvement in agents.
First, split fast correction from slow improvement. Online loops should correct local errors quickly, but larger model or prompt changes should be evaluated offline through replay, regression tests, and rollbackable promotion gates. This is the software analogue of cerebellar correction plus sleep-based consolidation.
Second, treat mutation rate as a governed variable. The lesson of metaplasticity is that a system should not only learn; it should regulate how aggressively it is allowed to learn. Agents that recently changed a lot should cool down before further self-modification.
Third, keep value signals separate from world-model signals. Predictive coding and dopamine solve different problems in the brain. Agent architectures should likewise distinguish 'did I predict correctly?' from 'did this outcome advance the objective?', rather than collapsing both into a single reward proxy that invites reward hacking.
Fourth, require external reality checks. The most dangerous failure mode in recursive systems is mistaking self-generated signals for external validation. Tier-2 improvement therefore needs benchmark replay, counterfactual tests, and evidence from actual task outcomes — not just self-scored confidence.
Failure Modes: When Self-Improvement Goes Wrong
The brain's recursive self-improvement architecture is powerful but not infallible. Several pathologies illustrate what happens when the loop breaks down:
Addiction hijacks the dopamine prediction error signal. Drugs of abuse produce artificially large dopamine bursts that override the brain's natural value estimation, driving compulsive behavior that the cortical monitoring systems cannot override. This is the biological equivalent of reward hacking in reinforcement learning — the agent optimizes a proxy metric that diverges from the true objective.
Rumination and anxiety represent failure modes of the predictive coding loop. The brain generates catastrophic predictions, cannot resolve the prediction error through action or evidence, and enters a self-reinforcing cycle of negative prediction and escalating arousal. The monitoring system detects a problem but the repair mechanism is unable to address it, leading to a stuck state.
Schizophrenia may involve a failure of the brain's ability to distinguish self-generated predictions from externally caused sensory signals. If the corollary discharge mechanism — the system that tags predictions as internally generated — malfunctions, the brain's own predictions are experienced as external events, producing hallucinations and delusions.
These failure modes are not merely clinical curiosities. They are design constraints. Any recursive self-improving system must guard against reward hacking, stuck states, and confusion between internal models and external reality.
Connection to Agent Systems: MARIA VITAL Evolution Lab
The brain's architecture provides a detailed blueprint for the MARIA VITAL Evolution Lab:
Prediction → Error → Update maps to the Evolution Lab's Hypothesis → Test → Promote pipeline. An agent proposes a configuration change (prediction), tests it against a benchmark suite (error measurement), and promotes or reverts based on results (weight update). The key insight from neuroscience is that the error signal cannot be purely rhetorical — it must be grounded in measurable outcomes, replay traces, or controlled evaluation rather than self-description alone.
Hierarchical error processing maps to the Evolution Lab's multi-level evaluation. A minor configuration change (parameter update) is evaluated at the unit test level. A major architectural change (structural update) requires integration tests, load tests, and human review — analogous to the cortical hierarchy escalating errors that cannot be absorbed at lower levels.
Sleep-as-batch-processing maps to the Evolution Lab's offline evaluation mode. Candidate mutations are tested in a sandboxed environment before deployment, replaying historical workloads at compressed timescales. This is the agent equivalent of hippocampal replay, and in the current MARIA VITAL vocabulary it corresponds more closely to shadow-agent validation and gated promotion than to unconstrained live rewriting.
Metaplasticity maps to adaptive mutation rates. An agent that has recently undergone significant changes should have a reduced mutation rate, allowing the effects of previous changes to be properly evaluated before introducing new ones. An agent in a stable, well-understood environment should also have a low mutation rate — do not fix what is not broken.
The brain teaches us that recursive self-improvement is not just possible but inevitable for any sufficiently complex adaptive system. The practical question is not whether agents will adapt, but where adaptation is allowed to occur, what evidence is allowed to count as improvement, and which gates prevent drift from turning into reward hacking, instability, or opacity.