TNeuralNetworkHandler — Neural Networks in Ascoos OS

Introduction

TNeuralNetworkHandler is the native class in Ascoos OS for building, training, and evaluating neural networks. It fully supports deep learning architectures and operates without external libraries or cloud dependencies.

Integrated into the Web5 Kernel, it works seamlessly with macro scripts, semantic profilers, and DSL commands. This documentation outlines its capabilities, supported architectures, and practical examples.

1️⃣ Network Capabilities

• Create and train neural networks via macro or DSL scripts
• Support for forward and backward propagation
• Automatic optimization using gradient descent and Adam
• Dynamic layer and activation function configuration
• Performance evaluation with metrics and confusion matrix

---

2️⃣ Supported Network Types

TNeuralNetworkHandler supports multiple neural network architectures, offering flexibility and adaptability across tasks:

• Dense (Fully Connected): Ideal for classification and regression tasks
• Convolutional Neural Networks (CNN): Suitable for image, signal, and spatial data processing
• Recurrent Neural Networks (RNN): Useful for sequences, language, and time-series analysis
• Long Short-Term Memory (LSTM): RNN extensions for long-term memory retention
• Autoencoders: For compression, reconstruction, and anomaly detection
• Residual Networks (ResNet): For deep networks with skip connections

All networks can be combined, trained, and evaluated natively without external libraries. The class also supports custom layers and activation functions.

---

3️⃣ Usage Examples

Below are practical examples of how to use TNeuralNetworkHandler in real-world scenarios:

3.1 Creating and Training a Network

$nn = new TNeuralNetworkHandler();
$nn->addLayer('dense', ['units' => 128, 'activation' => 'relu']);
$nn->addLayer('dense', ['units' => 64, 'activation' => 'relu']);
$nn->addLayer('dense', ['units' => 10, 'activation' => 'softmax']);
$nn->compile(['optimizer' => 'adam', 'loss' => 'categoricalCrossEntropy']);
$nn->fit($data, $labels, ['epochs' => 20, 'batchSize' => 32]);
        

3.2 Evaluating Performance

$metrics = $nn->evaluate($testData, $testLabels);
echo $metrics['accuracy']; // e.g., 0.87
        

3.3 Predicting with a Trained Network

$prediction = $nn->predict($newInput);
print_r($prediction); // Array of probabilities or labels
        

---

4️⃣ Integration with Web5

TNeuralNetworkHandler is deeply integrated into the Web5 Kernel of Ascoos OS, enabling neural networks to participate in semantic and autonomous workflows.

• Macro Engine: Networks can be created and trained via macro scripts for dynamic decision-making and adaptive logic.
• Semantic Profiler: Neural outputs can be semantically scored and influence execution paths and system behavior.
• DSL Scripting: Developers can embed neural logic into DSL scripts for intelligent branching and contextual adaptation.
• Data Modules: Networks can be trained on internal datasets such as logs, user inputs, macro history, and structured data.

This integration transforms Ascoos OS into a native deep learning platform, where logic and learning are core system capabilities.

---

📚 References

• 1 : Native AI in Ascoos OS — The Foundation of Web5
• 2 : Ascoos OS Case Studies
• 3 : Ascoos OS Meets Web5

✍️ Written by Christos Drogkidis — Developer, thinker, and advocate of logic-driven software design.