SNAILAlgorithm<T, TInput, TOutput>

Models & Types · AiDotNet.MetaLearning.Algorithms

Implementation of SNAIL (Simple Neural Attentive Meta-Learner) for few-shot learning.

Inheritance: ModelBase<T, TInput, TOutput> → MetaLearnerBase<T, TInput, TOutput> → SNAILAlgorithm<T, TInput, TOutput>

For Beginners

SNAIL treats few-shot learning as reading a story:

How it works:

  1. Take all support examples and their labels, arrange them in a sequence
  2. Append query examples (without labels) at the end
  3. Use temporal convolutions to capture local patterns (nearby examples)
  4. Use causal attention to capture global patterns (any previous example)
  5. The model predicts labels for query examples using what it “remembers”

Analogy: Imagine reading a detective novel:

  • Support examples are like clues scattered through the story
  • Temporal convolutions help you remember recent clues (short-term memory)
  • Causal attention lets you recall any clue from the entire story (long-term memory)
  • At the end (query), you must solve the mystery using all clues gathered

Key insight: By treating few-shot learning as sequence modeling, SNAIL can leverage powerful sequence architectures (temporal convolutions + attention) that have been very successful in NLP and speech recognition.

Architecture:

  • Temporal Convolution (TC) blocks: Dilated causal convolutions that aggregate local info
  • Causal Attention blocks: Attend to any previous position in the sequence
  • Stacked TC + Attention blocks form the full SNAIL architecture

How It Works

SNAIL combines temporal convolutions with causal attention to process few-shot learning as a sequence modeling problem. Support examples (with labels) are fed sequentially, then query examples are fed to produce predictions. The model learns which examples to remember and attend to.

Algorithm - SNAIL:

Reference: Mishra, N., Rohaninejad, M., Chen, X., & Abbeel, P. (2018). A Simple Neural Attentive Learner. ICLR 2018.

Constructors

Constructor Summary
SNAILAlgorithm(SNAILOptions<T, TInput, TOutput> options) Initializes a new SNAIL meta-learner.

Properties

Property Summary
MetaLearningAlgorithmType AlgorithmType { get; }

Methods

Method Summary
IModel<TInput, TOutput, ModelMetadata<T>> Adapt(IMetaLearningTask<T, TInput, TOutput> task) Adapts to a new task using SNAIL’s sequence processing.
T MetaTrain(TaskBatch<T, TInput, TOutput> taskBatch) Performs one meta-training step for SNAIL.