PrototypeAlignmentLayer<T>

Layers · AiDotNet.NeuralNetworks.Layers

A learned prototype-alignment layer per Sun et al. 2024 “TEST: Text Prototype Aligned Embedding to Activate LLM’s Ability for Time Series”. Maintains a bank of numPrototypes learnable embeddings of dimension embedDim. For each input token, computes cosine similarity to every prototype, softmax-normalizes, then aggregates prototypes by weight — producing an aligned representation that lives in the same prototype subspace as a frozen LLM’s text-token embeddings.

Inheritance: LayerBase<T> → PrototypeAlignmentLayer<T>

Implements: IShapeContract

For Beginners

Think of this as a “translator” that maps each input patch to the closest matches in a small library of learned reference embeddings (prototypes). The output is a weighted blend of those reference embeddings, so downstream layers (often a frozen language model) see inputs in a vocabulary they already understand.

How It Works

Forward: input [B, N, D] → cosine similarity with prototypes [K, D] → softmax over K → weights [B, N, K] → aggregate prototypes → output [B, N, D].

The prototype bank is trainable and initialized with small random values. During training, the prototypes learn to represent the “grammar” of time-series patches in a way that is compatible with a downstream frozen LLM.

Reference: Sun, C. et al., “TEST: Text Prototype Aligned Embedding to Activate LLM’s Ability for Time Series”, ICLR 2024. https://openreview.net/forum?id=Tuh4nZVb0g.

Constructors

Constructor Summary
PrototypeAlignmentLayer(int embedDim, int numPrototypes) Initializes a new PrototypeAlignmentLayer.

Properties

Property Summary
bool SupportsTraining { get; }

Methods

Method Summary
Dictionary<string, string> GetParameterRoles() Returns parameter roles for per-role optimizer configuration (e.g., weight decay exemption for biases).
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank) Derived from this type’s [TensorLayout] declarations.
IReadOnlyList<Tensor<T>> GetTrainableParameters() Returns all trainable parameter tensors marked with [TrainableParameter].
void ResetState()
void SetTrainableParameters(IReadOnlyList<Tensor<T>> parameters) Replaces trainable parameter tensors (e.g., with ParameterBuffer views).
void ZeroGrad() Clears all gradient fields discovered by convention ({paramName}Gradient).