memlnaut-nisps/NISPS_CORE_EXTRACTION_PLAN.md

NISPS Core Extraction Plan

Extract a platform-agnostic C++17 controller library from MEMLNaut-NISPS. This is not a synth or audio engine - it's a parameter mapping engine: control data in → ML → control data out. Use it to drive synths, effects, lights, robots, whatever.

What This Is

NISPS core takes N input parameters (joystick position, sensor data, audio features) and maps them to M output parameters through an interactively-trained neural network. Users teach it by example: "when I'm here in input space, I want these output values."

Dependencies to Remove

Dependency	Replacement
Serial.print*	Optional log callback
queue_t (Pico SDK)	Not needed (single-threaded)
WString.h (Arduino)	std::string
__force_inline, AUDIO_MEM	No-op macros

Structure

nisps/
├── mlp.hpp        # MLP implementation (from memlp, cleaned)
├── dataset.hpp    # Training dataset
├── iml.hpp        # Interactive ML engine (~200 lines)
└── voice_space.hpp # Optional: example parameter mappings

Core API

namespace nisps {

template<typename Float = float>
class IML {
public:
    IML(size_t n_inputs, size_t n_outputs,
        std::vector<size_t> hidden_layers = {10, 10, 14});

    // Input
    void set_input(size_t index, Float value);
    void set_inputs(const Float* values, size_t count);

    // Output (valid after process())
    const Float* get_outputs() const;
    size_t num_outputs() const;

    // Runtime
    void process();  // Run inference, call at control rate

    // Training workflow
    enum class Mode { Inference, Training };
    void set_mode(Mode mode);
    void save_example();      // Store current input→output as training pair
    void clear_dataset();
    void randomise_weights(); // For exploration in training mode
    void train();             // Blocking, runs on current dataset

    // Optional
    using LogFn = void(*)(const char*);
    void set_logger(LogFn fn);
};

} // namespace nisps

Usage

#include "nisps/iml.hpp"

nisps::IML<float> iml(3, 24);  // 3 inputs (x,y,z), 24 outputs

// Control loop
void update(float x, float y, float z) {
    iml.set_input(0, x);
    iml.set_input(1, y);
    iml.set_input(2, z);
    iml.process();

    const float* params = iml.get_outputs();
    my_synth.set_filter_cutoff(params[0] * 10000.f);
    my_synth.set_resonance(params[1]);
    // ... etc
}

// Training (triggered by user interaction)
void on_user_saves_position() {
    iml.save_example();
}

void on_user_exits_training_mode() {
    iml.set_mode(nisps::IML<>::Mode::Inference);
    // This triggers training internally
}

Voice Spaces (Optional)

Voice spaces are just functions that interpret the raw 0-1 output parameters. Not part of core, but useful as examples:

// User-defined mapping
void apply_neve_style(const float* params, MyChannelStrip& strip) {
    strip.pre_gain = 0.5f + params[0] * params[0] * 4.f;
    strip.hp_freq = 30.f + params[8] * params[8] * 270.f;
    strip.comp_threshold = 20.f + params[10] * -40.f;
    // ... etc
}

Phases

Phase 1: Get memlp building standalone (1-2 days)

1. Copy memlp source into nisps/
2. Remove Serial.print calls (or stub them)
3. Remove Arduino String usage
4. Verify it compiles with g++/clang

Phase 2: Wrap in IML interface (2-3 days)

1. Create iml.hpp with the API above
2. Port state machine logic from IMLInterface.hpp
3. Simple test: train on XOR, verify inference works

Phase 3: Example integration (1-2 days)

1. Command-line example that reads CSV input, outputs CSV
2. Or: minimal JUCE/SDL example with mouse input

Total: ~1 week to something usable

Later (only if needed)

• Model serialization (save/load trained weights)
• Thread-safe parameter updates
• Python bindings
• WASM build