ssc_architecture.mmd

SSC architecture diagram - 43 lines. View on GitHub (speech-neuro/ssc_architecture.mmd).

Rendered diagram

Rendering diagram…

Source

graph TD
    INPUT["SSC Input<br/><i>Cochlear spike events (t, channel)</i>"]

    subgraph FE["Histogram Binning + Encoding"]
        direction LR
        BIN["Dense Histogram<br/><i>700 channels × 100 bins</i>"]
        DY["DyNED<br/>32 levels"]
        BIN --> DY
    end

    subgraph SL["cAdLIF Spiking Layers"]
        direction LR
        S1["Linear + BN + Delay + cAdLIF<br/>Drop 0.23<br/><i>700 → 512</i><br/>Surrogate: ArcTan"]
        S2["Linear + BN + Delay + cAdLIF<br/>Drop 0.23<br/><i>512 → 512</i><br/>Surrogate: ArcTan"]
        S3["Linear + BN + Delay + cAdLIF<br/>Drop 0.23<br/><i>512 → 512</i><br/>Surrogate: ArcTan"]
        S4["Readout: Linear + Leaky Integrator<br/><i>512 → 35</i>"]
        S1 --> S2 --> S3 --> S4
    end

    TS["100 Time Steps per Sample<br/><i>One histogram bin = one SNN timestep</i>"]
    SC["Softmax Vote<br/><i>Accumulated over 100 steps</i>"]
    CLASS["35 SSC Commands"]

    INPUT --> FE
    FE --> SL
    TS -.- SL
    SL --> SC
    SC --> CLASS

    classDef inputStyle fill:#27AE60,stroke:#1E8449,color:white,font-weight:bold,font-size:20px
    classDef freqStyle fill:#3498DB,stroke:#2471A3,color:white,font-weight:bold,font-size:20px
    classDef dynedStyle fill:#E67E22,stroke:#CA6F1E,color:white,font-weight:bold,font-size:20px
    classDef snnStyle fill:#C0392B,stroke:#922B21,color:white,font-weight:bold,font-size:20px
    classDef classStyle fill:#2C3E50,stroke:#1B2631,color:white,font-weight:bold,font-size:20px
    classDef noteStyle fill:#F5F5F5,stroke:#999,color:#333,font-style:italic,font-size:20px

    class INPUT inputStyle
    class BIN freqStyle
    class DY dynedStyle
    class S1,S2,S3,S4 snnStyle
    class SC,CLASS classStyle
    class TS noteStyle