Struct StreamingTTT 

pub struct StreamingTTT { /* private fields */ }

Streaming TTT model with RLS readout.

Processes one sample at a time. The TTT layer adapts its internal representation (fast weights) via gradient descent on reconstruction at every step. An RLS readout maps the TTT output to the target.

Example

use irithyll::ttt::{StreamingTTT, TTTConfig};
use irithyll::StreamingLearner;

let config = TTTConfig::builder().d_model(16).eta(0.1).build().unwrap();
let mut model = StreamingTTT::new(config);
model.train(&[1.0, 2.0, 3.0], 4.0);
let pred = model.predict(&[1.0, 2.0, 3.0]);

Implementations

impl StreamingTTT

pub fn new(config: TTTConfig) -> Self

Create a new StreamingTTT from config.

pub fn past_warmup(&self) -> bool

Whether the model has seen enough samples for meaningful predictions.

pub fn config(&self) -> &TTTConfig

Access the config.

pub fn prediction_uncertainty(&self) -> f64

Forward-looking prediction uncertainty from the RLS readout.

Returns the estimated prediction standard deviation, computed as the square root of the RLS noise variance (EWMA of squared residuals). This is a model-level uncertainty signal that does not require transformed features.

Returns 0.0 before any training has occurred.

pub fn output_dim(&self) -> usize

Output dimension of the TTT layer.

pub fn pretrain_projections(&mut self, data: &[(&[f64], f64)], epochs: usize)

Pretrain projection matrices W_K, W_V, W_Q using a warmup data buffer.

This is the “outer loop” learning step that the papers (Titans, TTT) require. Random projections cause the fast weights to learn in an arbitrary subspace. Pretraining aligns the projections with the actual data distribution.

Like MTS (minimum training samples) for trees: collect warmup data, learn the projection basis, then switch to pure streaming.

Arguments

• data — Slice of (features, target) pairs for pretraining.
• epochs — Number of passes over the data (typically 3–10).

How it works

1. Initialize — process data normally to build up W_fast and RLS readout so that readout weights exist for gradient computation.
2. Optimize — for each epoch, accumulate gradients on W_Q (prediction loss) and W_K/W_V (reconstruction loss), clip, then update projections.
3. Clean slate — reset fast weights, RLS, and all tracking state so the model is ready for normal streaming with learned projections.

pub fn set_projections( &mut self, w_k: &[f64], w_v: &[f64], w_q: &[f64], ) -> Result<(), ConfigError>

Inject custom projection matrices for W_K, W_V, W_Q.

Enables researchers to load externally pre-trained projections (e.g. from PyTorch/JAX experiments) before streaming. Each matrix is [d_state x d_input] in row-major order.

Trait Implementations

impl Debug for StreamingTTT

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl DiagnosticSource for StreamingTTT

fn config_diagnostics(&self) -> Option<ConfigDiagnostics>

Return config diagnostics, or None if not supported.

impl StreamingLearner for StreamingTTT

fn train_one(&mut self, features: &[f64], target: f64, weight: f64)

Train on a single observation with explicit sample weight.

fn predict(&self, features: &[f64]) -> f64

Predict the target for the given feature vector.

fn n_samples_seen(&self) -> u64

Total number of observations trained on since creation or last reset.

fn reset(&mut self)

Reset the model to its initial (untrained) state.

fn diagnostics_array(&self) -> [f64; 5]

Raw diagnostic signals for adaptive tuning.

fn adjust_config(&mut self, lr_multiplier: f64, _lambda_delta: f64)

Apply smooth learning rate and regularization adjustments.

fn readout_weights(&self) -> Option<&[f64]>

Return the readout weight vector for supervised projection, if available.

fn train(&mut self, features: &[f64], target: f64)

Train on a single observation with unit weight.

fn predict_batch(&self, feature_matrix: &[&[f64]]) -> Vec<f64>

Predict for each row in a feature matrix.

fn apply_structural_change(&mut self, _depth_delta: i32, _steps_delta: i32)

Apply structural changes at model replacement boundaries.

fn replacement_count(&self) -> u64

Total number of internal model replacements (e.g. tree replacements triggered by drift detection or max-tree-samples).

fn check_proactive_prune(&mut self) -> bool

Manually trigger a proactive prune check.

fn set_prune_half_life(&mut self, _hl: usize)

Dynamically set the contribution accuracy EWMA half-life.

fn tree_structure(&self) -> Vec<(usize, usize, f64, f64, u64)>

Optional tree-level structure diagnostics.