NEUROSCIENCE BRAIN SIMULATION PAPER

EvoSpikeNet Whole Brain Simulation: For Neuroscientists

Author: Masahiro Aoki

Copyright: 2026 Moonlight Technologies Inc.

• Document ID; MT2026-AI-01-001
• ORCID ID: 0009-0007-9222-4181
• Affiliation: Moonlight Technologies Co., Ltd.

summary

In this paper, we use a spiking neural network and biological control architecture to A distributed modular simulation platform that approximates the whole human brain We demonstrate the design and implementation of "EvoSpikeNet". What is traditional deep learning and reinforcement learning? Differently, our system uses time-series spike propagation, plasticity-dependent self-regulation, and It is characterized by precise synchronization between geographically dispersed nodes. Existing AI was weak Software architecture that “maintains temporal causality” and “natural hierarchy to higher-order spatial cognition” level, and incorporates the physical latency constraints of neural circuits into the evaluation metrics. The system consists of specialized nodes such as visual, auditory, spatial, linguistic, motor, and memory. Coupled with low-latency middleware (Zenoh) and equipped with quantum-inspired self-modulation Supervised by the prefrontal control module. This paper describes the conceptual framework, functional layers, We discuss mathematical models, data flow, and cortical region correspondence. The remaining functions are listed in the specifications. Biological behavior (hippocampus, amygdala, cerebellum, neuromodulation, etc.) has already been incorporated, It has already been implemented. We present neuron models, plasticity rules, attention mechanisms, and time synchronization equations. Includes detailed block diagrams and brain region maps.

Keywords: Spiking neural network, distributed brain simulation, PFC, ChronoSpikeAttention, STDP, whole brain, cortical mapping

1. Introduction

There are approximately 86 billion neurons and 100 trillion synapses in the human cerebral cortex. It is differentiated into specialized areas such as hearing, language, memory, and motor control. The existing calculation model is Usually focused on a single function, EvoSpikeNet integrates these multiple modules and “Whole-brain simulation” as an extensible distributed software framework This is an attempt to build.

The purpose of this paper is as follows.

1. Maintaining functional decomposition corresponding to cortical areas.
2. Real-time (≤200ms) using a biologically plausible spiking model end-to-end) operation.
3. Continuous learning and self-healing through plasticity (STDP, meta‑STDP) and Supported by the evolutionary genome layer.
4. Expose parameters, timing, and state variables as physiological indicators, To enable neuroscientific verification.
5. Implement biological domains including structures such as the hippocampus, amygdala, and basal ganglia.

The target audience is neuroscientists and others who want to deeply understand this platform. I am a system architect.

Differences with AI

Conventional neural networks are mainly based on dense vector calculations of activation values, Preserving the causal structure of time-series information, dynamic memory management, Not suitable for tasks that require absolute synchronization between geographically distributed processes. EvoSpikeNet treats spike times as physical events on the time axis rather than as binary numbers. By using node synchronization with Zenoh and PTP, Robotics, spatial navigation, real-time EEG closed loop, etc. It has demonstrated performance in "areas where conventional AI was weak."

Specific example: - Preserves temporal causality when tracking and predicting moving objects in 3D space ChronoSpikeAttention provides continuous attention. - For low-latency (<50ms) motion control, PTP synchronization allows each module to Firing timing is aligned to the nanosecond level, which is not possible with traditional batch inference. Achieve reaction speed. - In continuous learning, Meta‑STDP suppresses catastrophic forgetting, Long-term memory becomes possible.

2. System architecture and biological correspondence

Provide a timestamp. Spatial cognition node (Feature 13) is in the cognitive layer It is located in the occipito-parietal cortex.

2.1 Block diagram

flowchart LR
    subgraph "Sensory/Encoding"
        CAM["Camera/Retina (V1‑V5)"]:::implemented
        MIC["Microphone/A1"]:::implemented
        TAS[TAS encoding]:::implemented
    end
    subgraph "Cognitive"
        VIS["Visual module (occipital lobe/IT cortex)"]:::implemented
        AUD["Auditory module<br/>(temporal lobe)"]:::implemented
        SLM["Spiking LM<br/>(Language/Broker – Wernicke)"]:::implemented
        RAG["Hybrid RAG index"]:::implemented
        SPAT["空間層<br/>(parietal lobe)" ]:::implemented
        BM["Biomimetic module group (emotions/rewards/sleep/rhythm etc.)"]:::biomim
    end
    subgraph "Control/PFC"
        PFC["PFC / Q‑PFC<br/>(frontal cortex)"]:::implemented
    end
    subgraph "Memory/State"
        EPI["Episodic memory<br/>(hippocampus)"]:::implemented
        SEM["Semantic memory<br/>(temporal lobe)"]:::implemented
        MINT["メモリインテグレータ<br/>(cingulate gyrus)" ]:::implemented
    end
    subgraph "Motor/Output"
        MTR["Motor planner<br/>(motor cortex)"]:::implemented
    end
    subgraph "Infrastructure"
        ZEN["Zenoh Pub/Sub"]
        PTP["PTP time synchronization"]
    end

    CAM --> TAS --> VIS --> SLM
    MIC --> TAS --> AUD --> SLM
    SLM --> RAG --> PFC --> MTR
    SPAT --> PFC
    BM --> PFC
    PFC <--> EPI
    PFC <--> SEM
    EPI <--> MINT
    SEM <--> MINT
    ZEN --- PTP
    PFC -.-> ZEN
    VIS -.-> ZEN
    AUD -.-> ZEN
    SPAT -.-> ZEN
    MTR -.-> ZEN
    EPI -.-> ZEN
    SEM -.-> ZEN
    MINT -.-> ZEN
    BM -.-> ZEN

*Figure 1. High-level diagram showing cortical region correspondence. Solid line is data flow, dashed line is Pub/Sub mesh. *

2.2 Node ranks and hub architecture

Nodes are given a rank from 1 to 15 based on functional complexity and connectivity requirements. Ranks 1–5 are sensors such as vision and hearing, and basic cognitive processing, and Ranks 12 and above are advanced Implement spatial awareness and integration. Rank also affects a node's hub role, Rank1 exposes raw spikes as leaf nodes, Rank8–11 exposes raw spikes as leaf nodes, and Rank8–11 exposes raw spikes as leaf nodes and Rank12+, an intermediate hub for LM inference, is responsible for coordination across geographical areas.

A Node Hub is a logical structure within a Zenoh mesh that is responsible for data aggregation and redistribution. When the delay measurement (\(L_{ij}\)) exceeds the threshold, ExecutiveControlEngine will switch to another hub. Elevate and maintain communication below 50ms. All spikes are 64bit PTP timestamps and has a variable length payload in Protobuf format. The control message is A JSON schema called {type,source,target,payload} that can be used between different implementations of languages. But maintain compatibility.

Each hub maintains a list of nodes assigned to its rank area area descriptor have. During operation, this descriptor is broadcast at 1Hz, and other nodes Adjust the routing table. This mechanism eliminates the need for a central directory. Collaboration across clusters becomes possible.

2.3 Explanation of cortical region correspondence

In this version, brain region mapping has been added with the addition of biomimetic functions. Expanded. Emotion/reward circuit (amygdala/nucleus accumbens/VTA), sleep/rhythm synchronization (hippocampal theta waves + ACh), mirror neuron system (premotor cortex vs. observation cortex) , intention/motivation vectors (PFC/ACC/NAcc), and developmental dynamics. (critical period pruning/myelination) etc. have been newly implemented, and the following It constitutes functional cooperation.

• PFC receives emotional signals from the amygdala and VTA dopamine prediction errors. Adjust learning rate and simultaneously perform memory encoding by releasing ACh synchronized with Theta band EEG. Switch.
• The mirror neuron system is implemented as a bidirectional connection between motor cortex and visual cortex, The model also internally generates motor commands and estimates rewards during observation behavior.
• Developmental dynamics are epoch dependent due to DevelopmentalSchedule Controls plasticity/pruning/conduction delay and works with curriculum scheduler The training difficulty level increases gradually.
• Sensory preprocessing is retina DoG → V1 Gabor, cochlear gamma tone, vestibule Provides acceleration/angular velocity normalization and effect copy of motion commands.

These modules are aggregated into the evospikenet/biomimetic package, Shared to all nodes via EEG metadata in DistributedBrainExecutor.

The best EvoGenome generated by the evolution engine Can be deployed directly to each node via DistributedBrainNode.deploy_genome() It is currently implemented. After deployment, within each node's command processing loop: InstantiatedBrain (real network generated by GenomeToBrainConverter) A forward pass is executed and the results of genome evolution are displayed in real time. It is reflected in the inference field. In addition, from the STDP spike government reinforcement history Apply the calculated INT16 delta to the nn.Linear history using apply_weight_delta(). The ability to apply immediately was also implemented, closing the online plasticity loop.

Furthermore, each node mimics the function of the corresponding cortical/subcortical area in software. Parameters (delay, number of neurons, plasticity coefficient, etc.) are based on actual physiological values. I'm adjusting it closer.

Node	Biological Domain	Main Features	Average Latency	Implementation Status
Vision	Occipital Cortex (V1–V5, IT)	Scene Analysis/Object Recognition	45ms	Completed
Hearing	Primary auditory cortex (A1)	Frequency extraction/sound source localization	40ms	Completed
Language/SLM	Broca–Wernicke circuit	Language representation generation in the brain	60ms	Completed
Spatial	Superior parietal lobule, occipito-parietal cortex	Position/object recognition/attention control	50ms	Feature 13 completed
PFC	Dorsolateral/orbitofrontal cortex	Decision making/goal management	30ms	Core module completed
Movement	Primary motor cortex	Motor plan generation/output	25ms	Completed
Episodic	Hippocampus	Temporal Sequence Memory	—	Complete
Semantics	Medial temporal lobe	Conceptual knowledge retention	5ms search	Complete
Integrator	Cingulate/Insular Cortex	Intermodal Integration	10ms	Done

Each explanation includes major papers in the biological field and the corresponding EvoSpikeNet module. Implementation notes are included as internal documentation. as needed Inducible.

*Table 1. Correspondence between EvoSpikeNet modules and cortical regions. *

2.4 Connectome → Node automatic mapping (Addendum)

In this section, while maintaining neuroscientific validity, we introduce the rank structure of EvoSpikeNet from the public connectome. The principles, verification criteria, and experimental suggestions for automatic mapping are presented. The main purpose is to ensure reliability for research purposes.

• Priority of biological constraints: Keep E/I ratio, layer distribution, and cell type frequency as the highest priority. These are the key metrics that should be preserved during reduction, as they directly affect network dynamics.
• Applying reduction methods: Apply multiple of Policy F (stratified sampling, spectral reduction, cluster representation) and select the method that best matches the source distribution. Selection criteria were E/I difference, KS test of degree distribution, and significance level (p>0.05).
• Verification experiment: Use the generated structural_mask to compare learning curves in short-term learning tasks (visual discrimination, auditory discrimination, etc.). We report the correlation between biological indicators (kurtosis, synchrony, E/I ratio) and task performance.
• Output storage format: Assign version_uuid and etag to NPZ (COO format array: row_indices, col_indices, weights, delays, ei_mask) to ensure reproducibility and version control.
• Ethics/Contract: HCP and other subject data must undergo procedures and approval for research use in accordance with DUC. Regarding the use of public data, please attach a procedural history in the appendix of your paper.

The above is an addendum for the neuroscience community, and it is recommended that it be published as an experiment note along with implementation results (E/I retention rate, KS p-value, task learning curve).

2.4.1 Brain function integration derived from connectome

In this section, we will discuss how to extract and reduce structural information from public connectomes. Specify how to integrate it into EvoSpikeNet's functional module. design guidelines and Clarify verification indicators to ensure reproducibility for research purposes.

• Structural hub (rich‑club): High-degree nodes are high-rank nodes equivalent to PFC (Rank12+) or map to a central hub. Validation indicator: Rich Club Coefficient KS test for conservation rate and degree distribution.
• Modularity/Community Structure: Group nodes based on module detection Register it as an area descriptor and reflect it in the Rank assignment/routing table. Validation metrics: Comparison of modularity \(Q\), cluster consistency score.
• E/I balance and cell type frequency: E/I ratio and cell type frequency at the local circuit level. Retained and stored in structural_mask as ei_mask. Validation metrics: firing rate distribution, Conservation of E/I ratio.
• Layered projection (interlayer connections): Layered connections are retained as meta attributes and are Map to layer parameters (e.g. L2/3→L5 feedforward).
• Delays (long‑range vs local): Keep propagation delays in delays array and time synchronization Used for evaluation and routing optimization.
• Synaptic weight distribution: The weight distribution is saved as an initial condition and is determined by the plasticity rule. Updated while learning. Validation metric: statistical distance of weight distribution (e.g. KL divergence).
• Circuit motif (feedforward/feedback): Frequently occurring motifs are converted into templates. Retained and injected at the node mapping stage. The vibration/synchronization motif is in the hippocampus and cerebellum. Reflect in module design.
• Subgraph injection: Semantic subgraphs such as hippocampal circuits and visual columns are placed in dedicated nodes. Partially inject to ensure functional reproducibility.
• Neuromodulation binding: neuromod metadata to global PFC etc. map to modulator channels and dynamically modulate plasticity gain and learning rate.

Output format and operating rules:

• Save format: Include row_indices, col_indices, weights, delays, ei_mask, layers, region_tags, version_uuid, etag in NPZ (COO).
• Reduction policy: Default to Policy F (stratified sampling, spectral reduction, cluster representation), giving priority to preservation of E/I and degree distribution.
• Validation set: Compare learning curves on visual and auditory short-term learning tasks and report KS test, synchronization level (PLV), and memory usage.
• Reference implementation/configuration: For details, refer to docs-dev/connectome_schema.md, config/connectome_config.yaml, and test group (tests/test_connectome_loader.py, etc.).

The guidelines in this section allow for the analysis of connectome-derived structural information without compromising biological plausibility. Provided for integration into EvoSpikeNet modules. In actual operation, each experiment Explicitly record reduction parameters and verification protocols according to requirements to ensure reproducibility.

This section summarizes the neuron model, learning rules, and attention mechanism.

3. Mathematical models and formulas

3.1 Neuron mechanics

3.1.1 Leakage Integral Firing (LIF)

\[\tau_m \frac{dV(t)}{dt} = -(V(t) - V_{rest}) + R I(t), \\ V(t) \ge V_{th} \Rightarrow \text{Spike}, \quad V \leftarrow V_{reset}. \]

Parameters can be set for each node. See Table 2 for typical values.

3.1.2 Izhikevich model

Using the above formula, more than 20 types of firing patterns can be reproduced.

3.1.3 ChronoSpikeAttention

Attention weight of spike time \(t_i,t_j\):

\[A_{ij} = \frac{\exp\left(\frac{Q_i K_j^T}{\sqrt{d_k}} - \frac{(t_i - t_j)^2}{2\sigma^2}\right)}{\sum_k \exp(\cdot)}\]

Future information is blocked and decays exponentially depending on time distance.

3.2 Plasticity Law

Time lag dependent plasticity (STDP)

\[\Delta w_{ij} = \begin{cases} A_+ e^{-\Delta t/\tau_+} & \Delta t > 0\\ -A_- e^{\Delta t/\tau_-} & \Delta t < 0 \end{cases}\]

4. Connectome integration (CONNECTOME_INTEGRATION_PAPER integration)

This section integrates and augments the contents of docs/CONNECTOME_INTEGRATION_PAPER.md with the EvoSpikeNet system description. The figures use embedded Mermaid, with detailed explanations below each figure.

4.1 Integration overview - Purpose: Initialize the connection topology inside EvoSpikeNet nodes with public connectome data and clearly separate the structural and functional layers to achieve both learning stability and biological validity. - Implementation status (Updated on 2026-03-19 — Phase E-0/E-1/E-2 completed): - ✅ config/connectome_config.yaml: Implemented. - ✅ Document integration (this article): Implemented. - ✅ evospikenet/connectome_loader.py: Phase E-1 completed. load_json · load_npz · save_npz · stratified_sample (F-1 stratified sampling) · spectral_coarsen (F-2 spectral reduction) · load (ETag+TTL cache) has been implemented. - ✅ LIF extension of evospikenet/core.py (ConnectomeLIFLayer): Phase E-1 completed. Implemented structural_mask (bool COO tensor) buffer, connectome_weight parameter, attach_sparse_delay_buffer(), validate_ei_ratio(). Integrate lazy routing (step_int16()) into SNNModel.forward(). - ✅ evospikenet/connectome/node_mapping.py: Phase E-2 completed. get_source_for_node · build_manifest · apply_to_layer implementation (18 tests PASS). - ✅ evospikenet/connectome/delay_buffer.py (SparseDelayBuffer): Phase E-2 completed. COO ring buffer format [max_delay+1, n_neurons], step / step_int16 / from_connectome_data implementation (22 tests PASS). - ✅ evospikenet/zenoh_connectome_publisher.py (ConnectomeMetadataPublisher): Phase E-2 completed. Zenoh topic connectome/metadata/{node_id}, session=None and logo-only mode (30 tests PASS). - ✅ evospikenet/forgetting_controller.py (compute_connectome_density): Phase E-2 completed. - ⬜ scripts/sync_connectome.py: Phase E-3 (not yet started). An automatic differential synchronization pipeline will be implemented in the future.

4.2 Three-layer model (repost and explanation)

graph TB
  subgraph "Layer 1: 構造層 Structural Layer"
    S1["Connectome-derived adjacency matrix A∈{0,1}^{N×N}"]
    S2["E/I neuron type mask"]
    S3["Synaptic delay table delay[i,j]"]
    S1 --- S2 --- S3
  end

  subgraph "Layer 2: 機能層 Functional Layer"
    F1["STDP weight scalar w_scalar"]
    F2["Plasticity update with Meta-STDP"]
    F3["Integration with ChronoSpikeAttention"]
    F1 --- F2 --- F3
  end

  subgraph "Layer 3: 進化層 Evolutionary Layer"
    E1["Structural mask coevolution with EvoGenome"]
    E2["Evolutionary optimization of pruning rate"]
    E3["Cross-scale adaptation"]
    E1 --- E2 --- E3
  end

  S1 -->|"W_ij = A_ij × w_scalar"| F1
  F2 -->|"Update only ΔSTDP × A_ij"| F1
  E1 -.->|"Phase E+ only"| S1

Description: The structure layer maintains a Boolean mask generated from an external connectome (FlyWire/MICrONS/C.elegans/HCP) that limits weight updates during training. The functional layer manages plasticity scalars and attentional weights, and the evolutionary layer handles long-term modification of the mask itself (Phase E+: research use).

4.3 Modules and data flows (integration)

The diagram below shows the end-to-end flow from connectome data acquisition to node initialization and distributed distribution. Describe the specific behavior after each block.

flowchart TB
  subgraph "コネクトームデータ層"
    CE_DB["C. elegans DB<br/>WormAtlas JSON"]
    FLY_DB["FlyWire DB<br/>CAVE API"]
    MIC_DB["MICrONS DB<br/>CAVEclient"]
    HCP_DB["HCP S1200<br/>FSL/MRtrix3"]
  end

  subgraph "コネクトームローダー層"
    LOADER["connectome_loader.py<br/>Sparse COO transformation"]
    SYNC["sync_connectome.py<br/>Auto-update pipeline"]
    CONFIG["connectome_config.yaml<br/>Source control"]
  end

  subgraph "構造層（変更不可）"
    MASK_V["Visual structural mask<br/>structural_mask[V1]"]
    MASK_A["Auditory structural mask<br/>structural_mask[A1]"]
    MASK_M["Motor structural mask<br/>structural_mask[M1]"]
    MASK_E["Episodic structural mask<br/>structural_mask[HPC]"]
    ZENOH_TOPO["Zenoh connection priority<br/>HCP derived weight"]
  end

  subgraph "機能層（STDP可変）"
    LIF_V["LIFNeuronLayer<br/>Visual Node<br/>W = mask × scalar"]
    LIF_A["LIFNeuronLayer<br/>Auditory Node"]
    LIF_M["LIFNeuronLayer<br/>Motor Node"]
    LIF_E["LIFNeuronLayer<br/>Episodic Node"]
  end

  CE_DB --> LOADER
  FLY_DB --> LOADER
  MIC_DB --> LOADER
  HCP_DB --> LOADER
  LOADER --> MASK_V & MASK_A & MASK_M & MASK_E & ZENOH_TOPO
  SYNC -->|"Weekly Difference"| LOADER

  MASK_V --> LIF_V
  MASK_A --> LIF_A
  MASK_M --> LIF_M
  MASK_E --> LIF_E

  LIF_V & LIF_A & LIF_M & LIF_E --> ZENOH_M["Zenoh Mesh"]
  ZENOH_TOPO --> ZENOH_M
  ZENOH_M --> PFC_C["PFC Control"]
  PFC_C --> STDP_C["Meta-STDP"] --> LIF_V & LIF_A & LIF_M & LIF_E
  EVO_C["EvoGenome"] -.->|"Structural mask update (Phase E+)"| MASK_V & MASK_A & MASK_M & MASK_E

Details: connectome_loader.py is responsible for authentication, acquisition, normalization, and COO conversion for each source. sync_connectome.py detects version differences and generates a difference COO for apply_weight_delta(). structural_mask represents the effective synapse set within each node, and LIFNeuronLayer uses that mask to generate initial weights as W = A × w_scalar.

4.4 Automatic updates and validation

An auto-update pipeline satisfies the following: - Difference detection: Compare the version number of FlyWire etc. and ETag. - Validation: Verify that the new ei_mask and E/I ratio are within the specified range (e.g. within ±0.5). - Rollback: Return to previous version when verification fails. - Distribution: Distribute weight_delta via Zenoh, and each node applies it after validation locally.

4.5 Initialization/PFC control sequence

In the initialization sequence, main.py reads connectome_config.yaml, initializes connectome_loader and retrieves micro/macro data in parallel. After obtaining, generate ConnectomeBundle{mask, weight, delay, ei_mask, hcp} and instantiate LIFNeuronLayer(...bundle[node_type]) for each node type.

PFC control operates in a 50ms loop, aggregates spike groups subscribed by Zenoh, and calculates route_probs and cognitive entropy H_t. Adjust the plasticity gain based on H_t, and in the learning state, use ForgettingController to prune connections with low contribution using a threshold.

4.6 Mathematical notes regarding learning - Definition of initial weights: - $\(W_{ij}^{(0)} = A_{ij}\cdot (s_{ij}\cdot \alpha\cdot e_{ij})\)$ - STDP with structural constraints: - $\(\Delta W_{ij} = A_{ij} \cdot \Delta_{STDP}(t_i,t_j)\)$ - E/I ratio constraint: Keep the local E/I ratio \(R = N_E/N_I\) from the source.

Meta‑STDP objective function

\[\mathcal{L}_{meta} = \mathcal{L}_{task} + \lambda_E E + \lambda_S \mathrm{Var}(r).\]

3.3 Control equation

Cognitive entropy: $\(H_t = -\sum_i p_i(t) \log p_i(t).\)$

Quantum modulation coefficient is obtained from a simulated quantum circuit with a Hamiltonian that depends on \(H_t\). Details in patent MT25‑EV003.

3.4 Communication and time synchronization

All spike events are timestamped with the PTP synchronization clock. Measure the inter-node delay \(\Delta_{ij}\) and calculate \(\sum_{i,j} \Delta_{ij} w_{ij}\) Perform routing to minimize.

4. Full brain operation scenario

In distributed brain simulation, multiple nodes operate in parallel, Enable event-driven data flow through Zenoh meshes. Below is It shows a typical sequence of operations and the average latency required for each step.

sequenceDiagram
    participant SENS as Sensory Nodes
    participant ENC as TAS Encoder
    participant COG as Cognitive Layer
    participant PFC as PFC/Q-PFC
    participant MEM as Memory Layer
    participant MOT as Motor Node
    participant ZEN as Zenoh Mesh

    SENS->>ENC: 生データ取得 (camera/mic) [5ms]
    ENC->>COG: スパイク列送信 [10ms]
    COG->>ZEN: 各モジュール間スパイク交換 [2-8ms]
    ZEN->>PFC: 統合入力到着
    PFC->>PFC: route_probs 計算、認知エントロピー測定 [30ms]
    PFC->>PFC: Q-PFC 量子変調実行
    PFC->>MOT: 命令送信 [25ms]
    MOT->>ZEN: 結果公開
    PFC->>MEM: 経験保存要求 [5ms]
    MEM->>COG: 過去情報検索要求 [5-10ms]
    COG->>PFC: 補完データフィードバック
    PFC-->>ZEN: 学習信号 (STDP/Meta-STDP)
    ZEN->>COG: 重み更新通知

• Initialization The Zenoh network is formed by an auto-discovery protocol, and each node Advertise your capabilities.
• Input The raw signal from the camera/microphone is received in real time.
• Encoding The TAS encoder generates a spike train with 1ms resolution and multiple Distribute to the recognition module.
• Processing Cognitive layer nodes fire independently and communicate with other nodes via Zenoh as necessary. Replace spikes. ChronoSpikeAttention and spatial nodes come into play here.
• Decision Making PFC integrates the received information and calculates route_probs and cognitive entropy. Q-PFC generates gating parameters according to the confidence level and determines the final Determine the output command.
• Output Motor nodes receive PFC commands and execute or Returns a vector.
• Memory update Experiences are recorded in an episodic/semantic store, When a search request is made, relevant data will be provided through the RAG.
• Learn Synapse weights are updated online based on STDP/Meta-STDP, EvoGenome will attempt to modify the architecture offline if necessary.

Memory retrieval and spatial attention are sandwiched between processing stages depending on task demands. There are cases. The current deployment supports a full brain configuration of 21 nodes, Nodes such as the hippocampus, amygdala, and cerebellum have also been integrated.

Memory retrieval and spatial attention are sandwiched between steps 4–6 depending on task demands. There are cases. The current deployment supports full brain simulation with 21 nodes, It is composed of nodes such as the hippocampus, amygdala, and cerebellum.

5. Residual biological behavior and planned expansion

5.1 EvoGenome configuration and collaboration

EvoGenome is a compact network architecture and plasticity parameter It is an expression and consists of the following three clauses.

1. Topology – Neighbor list including node ID, rank, edge weight \((n_i, r_i, w_{ij})\).
2. Parameters – Node-specific dictionary such as time constants, thresholds, STDP coefficients, etc.
3. Meta – Evolutionary metadata including fitness scores and mutation history.

During distributed operation, two-phase commit is performed on Zenoh to synchronize the genome. negotiator (usually the highest ranked PFC) proposes the update, and the follower node Perform checks (e.g. rank does not exceed current maximum). After the commit, all nodes are The new genome is read atomically, and nodes that are behind the old version can read the difference from any peer. You can request it.

The significance of EvoGenome lies in its ability to perform structural adaptation without centralized control. The node is When persistent performance degradation is detected, the genome is mutated locally and the changes are gossiped about. Spread it with. Other nodes accept or reject proposals based on their own metrics. With this federated evolution, even on geographically and environmentally disparate hardware, A distributed brain can adapt.

EvoGenome Deployment Pipeline to Distributed Nodes has been implemented. After DistributedEvolutionEngine.run_evolution() finishes, deploy_to_nodes([node1, node2, ...]) to all nodes. best_genome can be expanded all at once. The deploy_genome() method of each node is Use GenomeToBrainConverter and as an InstantiatedBrain instance. hold. From then on, in _process_brain_command() of all nodes Genome-driven forward pass is executed.

Remaining_Functionality.md has the following biologically motivated extensions: It is listed.

• Hippocampus: Sequence encoding, pattern separation/completion, theta-gamma combination. Oscillations implemented as episodic memory nodes to model coupling mechanics $\(V_{hp}(t)=A\sin(2\pi \theta t)+B\sin(2\pi \gamma t)\)$ Introduced (\(\theta\approx8\,\mathrm{Hz}\), \(\gamma\approx40\,\mathrm{Hz}\)). Cell assembly encodes time order through LTP/LTD, Search for similar episodes using cosine similarity.
• Amygdala: Emotional valence tagging and fear conditioning. Spiking to sensory input Encode the positive and negative values in the rate and set the PFC gating coefficient as $\(g_{PFC}=1+\alpha \cdot \mathrm{valence}\)$ By modulating like Controlling memory salience.
• Cerebellum: Main function is rapid sensorimotor error correction. Between PFC and Motor node seat, delta law $\(\Delta w = \eta (r - \hat r) x\)$ Works as a supervised learning element with Calculate the prediction error \(r-\hat r\) in real time. Generate and update parameters iteratively.
• basal ganglia: Behavioral selection in competitive spiking populations with Go/No‑Go pathways. Each route has a weight update rule according to the reward signal \(R(t)\) $\(\Delta w = \alpha R(t) - \beta\)$ , and controls behavioral firing when the threshold is exceeded.
• Neuromodulators: Global time series of dopamine/serotonin levels \(D(t),S(t)\), and the STDP coefficients \(A_+,A_-\) are Modulate. $\(A_+(t)=A_{+,0}(1+\gamma_D D(t)-\gamma_S S(t)).\)$
• Space/Time Hierarchy: Additional rank nodes for planning, language, and abstract reasoning. By introducing Higher‑rank nodes, we can solve long-term dependencies. Meta-learning becomes possible.

5.1 Cortical/subcortical area composition diagram

flowchart TB
    %% cortical modules already implemented
    Occipital["Vision<br/>(V1‑V5, IT)"]
    Temporal["Hearing<br/>(A1)"]
    Broca["Language<br/>(Broker area)"]
    Wernike["Language<br/>(Wernicke area)"]
    Parietal["Space<br/>(parietal lobe)"]
    Frontal["PFC<br/>(frontal cortex)"]
    Motor["motor cortex"]

    %% subcortical and neuromodulatory regions added
    Hippocampus["海馬<br/>(Sequence memory/
Pattern separation/completion)" ]
    Amygdala["扁桃体<br/>(emotional tagging,
fear conditioning)"]
    Cerebellum["Cerebellum (rapid sensorimotor error correction)"]
    BasalGanglia["Basal ganglia<br/>(Go/No‑Go behavior selection)"]
    Neuromod["神経調節<br/>(DA/5‑HT/ACh)"
]

    %% information flow
    Occipital --> Parietal
    Temporal --> Broca
    Broca --> Wernike
    Parietal --> Frontal
    Frontal --> Motor

    Hippocampus --> Frontal
    Amygdala --> Frontal
    Cerebellum --> Motor
    BasalGanglia --> Frontal
    BasalGanglia --> Motor

    %% neuromodulators broadcast globally
    Neuromod -.-> Occipital
    Neuromod -.-> Temporal
    Neuromod -.-> Parietal
    Neuromod -.-> Frontal
    Neuromod -.-> Motor
    Neuromod -.-> Hippocampus
    Neuromod -.-> Amygdala
    Neuromod -.-> Cerebellum
    Neuromod -.-> BasalGanglia

*Figure 2. Conceptual diagram of EvoSpikeNet modules and corresponding brain regions. Solid line is information Transmission path, dashed line indicates planned expansion. Emotion tag, error correction in the cerebellum, action selection in the basal ganglia, and dopamine/serotonin It reflects how neuromodulators such as nin/acetylcholine act on the whole brain. *

6.7 Individual subsystem diagram and detailed explanation

Below is a block diagram (mermaid), data flow, cortical correspondence, and implementation file references for the main biomimetic subsystems.

6.7.1 Hippocampus — Sequence encoding and replay

flowchart LR
  Input["Context / Sequence Input"] --> CA3["CA3 / Pattern Separation"]
  CA3 --> CA1["CA1 / Sequence Output"]
  CA1 --> Buffer["HippocampalBuffer\n(evospikenet/biomimetic/hippocampal_memory.py)"]
  Buffer -->|prioritized_replay| Sleep["SleepConsolidation\n(evospikenet/biomimetic/sleep_consolidation.py)"]
  Sleep --> Cortex["Cortical Targets (PFC/Temporal)"]

• Role: Time-series episode encoding, pattern separation/completion, prioritized replay.
• Implementation: evospikenet/biomimetic/hippocampal_memory.py, evospikenet/biomimetic/sleep_consolidation.py.
• Important parameters: replay batch size, priority criteria, SWR period (100–200 Hz imitation).

6.7.2 Prefrontal Cortex (PFC/Q‑PFC) — Intention/Decision Making and Gating

flowchart LR
  SensoryFeat["Features (Visual/Auditory/Spatial)"] --> PFCcore["PFC Core\n(evospikenet/biomimetic/intention_module.py)"]
  PFCcore --> Policy["Route_probs / Policy\n(Q‑PFC gating)"]
  Policy --> Motor["Motor Planner & Efference"]
  Reward["VTA TD Error\n(evospikenet/biomimetic/reward_circuit.py)"] --> PFCcore
  Neuromod["DA / ACh / Oxytocin\n(evospikenet/biomimetic/neuromodulators.py)"] --- PFCcore

• Role: Goal management (IntentionModule), routing, plasticity gating (via PlasticityGate).
• Implementation: evospikenet/biomimetic/intention_module.py, evospikenet/biomimetic/modulatory.py.
• Important parameters: Intention priority decay half-life, gating threshold, PFC inner loop delay.

6.7.3 Sleep Consolidation — Offline playback and memory transfer

sequenceDiagram
  participant H as HippocampalBuffer
  participant S as SleepConsolidation
  participant C as Cortex
  H->>S: prioritized episodes
  S->>C: replayed spike sequences (SWR bursts)
  Note right of C: PlasticityGate opens during replay (ACh modulation)

• Role: Long-term memory transfer from hippocampus to cortex by prioritized replay, interaction between SWR and δ/θ.
• Implementation: evospikenet/biomimetic/sleep_consolidation.py.

6.7.4 Reward/Emotion Circuit (VTA/NAcc/Amygdala)

flowchart LR
  Stimulus --> Amy["Amygdala\n(evospikenet/biomimetic/emotion_system.py)"]
  Amy --> Valence["valence/arousal"]
  RewardPredict["Value Estimator\n(VTADopamineModel)"] --> DA["Dopamine Signal"]
  DA --> Plasticity["PlasticityGate\n(evospikenet/biomimetic/modulatory.py)"]
  Plasticity --> Synapses

• Role: Dopamine release based on emotional tagging (valence/arousal) of stimuli and TD error, dynamic modulation of plasticity.
• Implementation: evospikenet/biomimetic/emotion_system.py, evospikenet/biomimetic/reward_circuit.py, evospikenet/biomimetic/modulatory.py.

6.7.5 Rhythm synchronization (θ/γ/δ) and EEG bands

flowchart TB
  HippocampusTheta["Hippocampal θ (4–8 Hz)"] --> AChTrigger["ACh Release\n(evospikenet/biomimetic/neuromodulators.py)"]
  CortexGamma["Cortical γ (30–80 Hz)"] --> Coupling["θ–γ Coupling\n(evospikenet/biomimetic/rhythm_sync.py)"]
  Coupling --> Memory

• Role: Sequence segmentation by θ–γ coupling, offline integration trigger by δ wave.
• Implementation: evospikenet/biomimetic/rhythm_sync.py, trigger for ACh module.

6.7.6 Mirror neurons and motor output (imitation learning)

flowchart LR
  ObservedAction["Observed Action Embedding"] --> Classifier["Action Classifier"]
  Classifier --> Mirror["MirrorNeuronSystem\n(evospikenet/biomimetic/mirror_neurons.py)"]
  Mirror --> MotorPrimitives["Motor Primitives (M1)"]
  MotorPrimitives --> ImitationReward

• Role: Activate motor primitives from observation and generate imitative rewards to promote learning.
• Implementation: evospikenet/biomimetic/mirror_neurons.py.

---

Each subsystem diagram can be disassembled and converted to high-resolution diagrams (SVG/PNG) for publication in papers. Suggested work next: - Add unit tests and expected input/output examples for each diagram to tests/unit/ (recommended) - Generate SVG of diagram and save to docs/assets/

6. Data flow and communication

All inter-node traffic uses Zenoh Pub/Sub, topic naming is It is in the <node_type>/<region>/<signal> format. Example: Visual spikes are Published to vision/v1/spikes. Subscribers can filter by ID or neighborhood, The SpatialWhere node subscribes to vision/*/spikes and performs local processing.

Clock synchronization is done through PTP grandmaster election and jitter is typically <1µs. Spike events contain 64bit timestamps and Zenoh is within the zone The order is guaranteed, but the entire mesh is asynchronous. master director node (monitored by ExecutiveControlEngine) logs \(L_{ij}\) delays every second, Used for adaptive routing.

7. Discussion and future prospects

EvoSpikeNet is a distributed spiking architecture that runs on commodity hardware. We demonstrated that a comprehensive set of brain functions can be simulated in real time. Hierarchical design and biologically plausible models ensure neuroscientific testability Facilitate: Parameters such as membrane time constants, plasticity time windows, and attentional decay contribute to experimental observations. Respond directly.

• EvoGenome → Distributed Node Deployment Bridge: DistributedEvolutionEngine.deploy_to_nodes() allows evolution results to be displayed immediately. This is reflected in the real-time reasoning of each node.
• STDP delta real weight application: InstantiatedBrain.apply_weight_delta() gives INT16 delta Applied in-place to the nn.Linear layer, closing the online plasticity loop.
• BrainSimulation import error fixed: The coupling of DistributedBrainNode and BrainSimulationFramework is now complete. Now it works.

A remaining limitation is that some subsystems require additional expansion and stabilization. Please refer to Remaining_Functionality.md for detailed implementation status and phase information. (The main functions of the hippocampus, amygdala, cerebellum, and neuromodulatory system have already been implemented). Large-scale scaling of PTP synchronization and fine-grained control of neuromodulators remain to be explored. This is an ongoing challenge.

Future research will include close examination of the parameters used in the connectome, closed-loop experiments integrating the EEG/BMI interface, Deployment to geographically distributed clusters (federated simulation), and fine-grained control of neuromodulators and adaptive learning evaluation in complex environments. Can be mentioned.