Table of Contents

Interface INeuralNetworkModel<T>

Namespace
AiDotNet.Interfaces
Assembly
AiDotNet.dll

Defines the contract for neural network models with advanced architectural introspection capabilities.

public interface INeuralNetworkModel<T> : INeuralNetwork<T>, IFullModel<T, Tensor<T>, Tensor<T>>, IModel<Tensor<T>, Tensor<T>, ModelMetadata<T>>, IModelSerializer, ICheckpointableModel, IParameterizable<T, Tensor<T>, Tensor<T>>, IFeatureAware, IFeatureImportance<T>, ICloneable<IFullModel<T, Tensor<T>, Tensor<T>>>, IGradientComputable<T, Tensor<T>, Tensor<T>>, IJitCompilable<T>

Type Parameters

T

The numeric data type used for calculations (e.g., float, double).

Inherited Members
Extension Methods

Remarks

This interface extends the basic neural network functionality with methods for accessing the internal architecture and layer-wise activations of neural networks.

For Beginners: This interface represents a neural network that can tell you about its structure.

Think of INeuralNetworkModel as a neural network with "x-ray vision" into its own structure:

  • It can show you what each layer produces (activations)
  • It can describe its own architecture (how it's built)
  • It combines all the basic neural network abilities with introspection capabilities

For example, if you're debugging or analyzing a neural network:

  • You can see what each layer outputs for a given input
  • You can examine the network's structure (number of layers, layer types, connections)
  • You can understand how information flows through the network

This is particularly useful for:

  • Debugging neural networks (seeing where things go wrong)
  • Understanding what the network has learned
  • Visualizing how the network processes information
  • Implementing advanced techniques like transfer learning or feature extraction

This interface is typically implemented by neural network base classes that provide comprehensive access to their internal structure and computations.

Methods

GetArchitecture()

Gets the architectural structure of the neural network.

NeuralNetworkArchitecture<T> GetArchitecture()

Returns

NeuralNetworkArchitecture<T>

The architecture object describing the network's structure.

Remarks

This method returns an object that describes the complete structure of the neural network, including all layers, their configurations, and how they connect to each other.

For Beginners: This gives you the "blueprint" of how the neural network is built.

Just like a building has an architectural blueprint showing:

  • How many floors it has
  • The layout of each floor
  • How rooms connect to each other

A neural network architecture describes:

  • How many layers the network has
  • What type each layer is (dense, convolutional, etc.)
  • How layers are connected
  • The size and shape of each layer

This information is useful for:

  • Understanding how the network is structured
  • Documenting your model
  • Recreating the same architecture with different parameters
  • Comparing different network designs
  • Implementing model serialization and deserialization

The architecture is typically set when the network is created and remains constant throughout the network's lifetime (though the parameters/weights change during training).

GetNamedLayerActivations(Tensor<T>)

Gets the intermediate activations from each layer when processing the given input with named keys.

Dictionary<string, Tensor<T>> GetNamedLayerActivations(Tensor<T> input)

Parameters

input Tensor<T>

The input tensor to process through the network.

Returns

Dictionary<string, Tensor<T>>

A dictionary mapping layer names to their activation tensors.

Remarks

This method processes the input through all layers of the network and returns a dictionary where each key is a descriptive layer name and each value is the output (activation) of that layer.

For Beginners: This shows you what each layer "sees" or produces when given an input.

Think of a neural network as a series of transformations:

  • The input enters the first layer
  • Each layer transforms the data and passes it to the next layer
  • The final layer produces the output

This method lets you see the intermediate results at each step. For example, in an image recognition network:

  • Layer 1 might detect edges (its activation shows edge patterns)
  • Layer 2 might detect simple shapes (its activation shows shape patterns)
  • Layer 3 might detect object parts (its activation shows parts like eyes or wheels)
  • Final layer produces the classification result

Each layer's name includes its position and type (e.g., "Layer_0_DenseLayer", "Layer_1_ConvolutionalLayer"). This is useful for:

  • Visualizing what the network learned
  • Debugging why the network makes certain predictions
  • Extracting features from intermediate layers
  • Understanding the network's decision-making process
Edit this page