MinLSTMLayer<T>

Layers · AiDotNet.NeuralNetworks.Layers.SSM

Implements the minLSTM layer from “Were RNNs All We Needed?” (Feng et al., 2024).

Inheritance: LayerBase<T> → MinLSTMLayer<T>

Implements: IShapeContract

For Beginners

minLSTM is a drastically simplified LSTM that trains much faster.

Traditional LSTMs have a chicken-and-egg problem: to compute the gates at step t, you need the hidden state from step t-1, which means you must process the sequence one step at a time. minLSTM solves this by making gates depend only on the current input:

  • Standard LSTM gates: “What should I remember?” depends on what I currently remember (h_{t-1})
  • minLSTM gates: “What should I remember?” depends only on what I’m currently seeing (x_t)

The normalization trick (f’ + i’ = 1) makes the update a weighted average: new_state = (weight_forget * old_state) + (weight_input * new_candidate) where weight_forget + weight_input = 1

This simple change means the entire sequence can be processed in parallel during training, just like a Transformer, while still maintaining efficient O(1)-per-step inference.

How It Works

minLSTM is a minimal LSTM variant that removes all hidden-state dependencies from the gates, enabling fully parallelizable training via a parallel prefix scan. In a standard LSTM, the forget gate f_t, input gate i_t, and output gate o_t all depend on the previous hidden state h_{t-1}. minLSTM eliminates every one of these dependencies:

Architecture:

Gate Normalization: The normalization f’ = f/(f+i), i’ = i/(f+i) is the key insight that makes minLSTM mathematically equivalent to a linear recurrence. Because f’ + i’ = 1, the cell update is a convex combination of the previous state and the new candidate. This is directly analogous to minGRU’s formulation where z and (1-z) play the same role. This convex structure allows the recurrence to be computed via a parallel scan operation, transforming training complexity from O(T) sequential to O(log T) parallel depth.

Comparison with minGRU:

  • minGRU uses a single gate z: h_t = (1-z)h_{t-1} + ztilde_h
  • minLSTM uses two gates (f, i) that are normalized: c_t = f’*c_{t-1} + i’*c_tilde
  • Both achieve input-only gating for parallel training
  • minLSTM’s two separate gates provide more expressive control over the forget-vs-input

trade-off, whereas minGRU ties them together through a single z

  • Both outperform traditional LSTMs and GRUs on many sequence modeling benchmarks while

being significantly faster to train due to parallel scan

Reference: Feng et al., “Were RNNs All We Needed?”, 2024. https://arxiv.org/abs/2410.01201

Constructors

Constructor Summary
MinLSTMLayer(int sequenceLength, int modelDimension, int expansionFactor, IActivationFunction<T> activationFunction, IInitializationStrategy<T> initializationStrategy) Creates a new minLSTM layer.

Properties

Property Summary
bool SupportsTraining { get; }
int ExpandedDimension { get; } Gets the expanded internal dimension.
int ModelDimension { get; } Gets the model dimension (input/output width).

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].
Tensor<T> GetForgetGateWeights() Gets the forget gate weights for external inspection.
Tensor<T> GetInputGateWeights() Gets the input gate weights for external inspection.
Tensor<T> GetOutputProjectionWeights() Gets the output projection weights for external inspection.
void ClearGradients()
void ResetState()
void SetTrainableParameters(IReadOnlyList<Tensor<T>> parameters) Replaces trainable parameter tensors (e.g., with ParameterBuffer views).
void UpdateParameters(T learningRate)
void ZeroGrad() Clears all gradient fields discovered by convention ({paramName}Gradient).