Shape Network (Python Edition)

An experimental neural network architecture ported from Java that utilizes sophisticated feature extraction and "shaping" functions. This network is designed for reliable classification tasks, such as facial recognition, by leveraging cosine similarity measures.

Unlike traditional backpropagation-based networks, this architecture performs well with single-class/label training, eliminating the need for complex weight balancing to achieve accurate predictions.

Features

• Shaping Functions: Uses mathematical "shapes" (Mean, Log-Mean, Min, Max) to adapt weights.
• Cosine Similarity: Measures the orientation of vectors rather than just magnitude, providing robustness against scaling.
• Partition Layers: Automatically selects and optimizes the best-fitting neurons for specific features.
• NumPy Backend: Optimized for numerical vector operations in Python.

Installation

This project requires NumPy.

pip install numpy

Usage

1. Build the Network

Use the ShapingNetworkBuilder to define the architecture. You can connect layers to specific target ranges.

from shape_network import ShapingNetworkBuilder, LayerBuilder, Shapes

model = ShapingNetworkBuilder() \
    .connect_in_range(0, 1, # Adds layers for targets 0 through 1
        LayerBuilder()
        .add_layer(Shapes.MEAN, 2) # Shaping function and number of nodes
    ).build()

2. Initialize and Randomize

Set the input dimension (number of features) and randomize the initial sum vectors.

# Initialize with 3 input nodes
model.initialize(3)
model.randomize()

3. Prepare Training Data

Data is fed into the model using ContextVector objects.

from shape_network import ContextVector

training_data = {
    0: [ # Class 0
        ContextVector([0.001, 2.531, 1.523]),
        ContextVector([0.009, 2.231, 1.241])
    ],
    1: [ # Class 1
        ContextVector([1.8, 0.001, 4.9]),
        ContextVector([2.5, 0.002, 4.5])
    ]
}

4. Training

Training is performed by "feeding" vectors to the model along with their target labels.

epochs = 100
for i in range(epochs):
    for target, vectors in training_data.items():
        for vec in vectors:
            model.feed(vec, target) # Adjusts the weights/shapes for that target

5. Predict

The predict method returns a list of probabilities (similarity scores) for each layer/class.

prediction = model.predict(test_vector)
# Get the index of the highest probability
max_index = prediction.index(max(prediction))
print(f"Predicted Class: {max_index}")

Example Output

The network converges quickly to high accuracy:

Accuracy: 0.998521
Class 0: 
Index=0, Probability=0.999873

Class 1: 
Index=1, Probability=0.997032

Mathematical Overview

The core of the network is the Shaping Function. Instead of traditional gradient descent, neurons maintain a sum_vector and a shape_vector. When feed() is called:

1. The sum_vector accumulates the input data.
2. The shaping_function (e.g., LOG_MEAN) is applied to derive a new shape_vector.
3. Prediction is calculated as the Cosine Similarity between the input vector and the neuron's shape_vector.

License

This project is licensed under the MIT License - see the LICENSE file for details.