adative learning and early stoping along with refactor and small touches

[tndmadman]

1. Refactored Neuron and Compartment Logic

   Preserved original neuron and dendritic compartment structure.

   Cleaned up variable naming for clarity and consistency.

   Added optional sigmoid and tanh explanations as inline comments.

2. Enhanced Training Routine

   Added adaptive learning rate decay.

   Introduced early stopping using patience and minDelta.

   Error is now tracked and logged every 1000 epochs for long runs.

3. Extended Task Support

   Expanded from just XOR and circle classification to:

    Logic gates (AND, OR, NOT, NAND, NOR, XOR, XNOR, BUFFER)
   
    Half Adder (SUM, CARRY)
   
    Full Adder (3-input SUM and CARRY)
   

   Implemented label-based multi-output support (universal neuron model).

4. Multi-Label Training Architecture

   Introduced []float64 expected outputs per training sample.

   Built a universal label-to-index map (labelIndexMap) and corresponding output vector for each logic label.

   All logic gates and combinational circuits are trained simultaneously.

5. Multithreaded Training

   Training now runs across multiple cores using Go’s sync.WaitGroup and runtime.GOMAXPROCS().

   Training batches are divided and updated in parallel using per-thread gradient accumulation and a mutex-protected global weight update step.

6. Improved Accuracy Evaluation

   Added clean, structured printout showing input, expected, predicted, rounded output, and correctness.

   Final output includes overall accuracy across all tasks.

7. New Data Generation Functions

   makeMultiLabelData() produces samples for logic gates and adders.

   generateCircleData() enables spatial classification for non-linear tests.

8. Other Enhancements

   Added reusable functions like normalizeInput() for data prep.

   Replaced hard-coded gates with scalable, indexed label definitions.

   Added detailed inline comments and structure for clarity and learning.

📁 File Structure & Usage

All logic is contained in a single file (main.go) for now.
Future refactor might include:

neuron.go

training.go

data.go

main.go as entry point

💡 Future Work

Optional: Convert to support GPU via CUDA with cgo/C bindings (complex).

Visualization of decision boundaries (circle classification) or logic outputs.

Export/import weights for persistence.

Build unit tests for each logic block.

[silver0a1]

Check out the latest implementations. The most recent successes of tying the neurons together are dann_tumor_bio.go, dann_tumor.go and the sine approximation test with a standard neural network as a control case.

Unaddressed-
Multi-threading(not there yet sorry)
Early-stop
Adaptive learning