architecture/network/runtime

Runtime control utilities for advanced network inference features.

Provides:

• Weight noise injection (global and per-layer).
• DropConnect regularization during inference.
• Stochastic depth (layer skipping) for layered networks.
• Iterative magnitude-based weight pruning with configurable schedules.

architecture/network/runtime/network.runtime.controls.utils.ts

clearStochasticDepthSchedule

clearStochasticDepthSchedule(): void

Clear the stochastic-depth schedule function so runtime behavior reverts to the currently stored static survival probabilities without additional per-step schedule adjustments.

Parameters:

• this - Target network instance.

Returns: Nothing.

clearWeightNoiseSchedule

clearWeightNoiseSchedule(): void

Clear the dynamic global weight-noise schedule so future steps stop applying schedule-driven standard-deviation updates and keep only explicit static configuration.

Parameters:

• this - Target network instance.

Returns: Nothing.

configurePruning

configurePruning(
  configuration: PruningConfiguration,
): void

Configure scheduled pruning during training.

This stores the pruning window and target policy on the network so the training loop can opportunistically apply structured sparsification later.

Parameters:

• this - Target network instance.
• configuration - Pruning schedule and ranking configuration.

Returns: Nothing.

disableStochasticDepth

disableStochasticDepth(): void

Disable stochastic depth entirely so all hidden layers participate in every pass and no layer-skipping regularization is applied at runtime.

Parameters:

• this - Target network instance.

Returns: Nothing.

disableWeightNoise

disableWeightNoise(): void

Disable all configured weight-noise mechanisms so subsequent training and inference passes execute without global or per-hidden-layer perturbation state, schedule updates, or hidden-layer noise carryover.

Parameters:

• this - Target network instance.

Returns: Nothing.

enableWeightNoise

enableWeightNoise(
  configuration: WeightNoiseConfiguration,
): void

Enable weight noise using either one global standard deviation or per-hidden-layer values.

A single global value is useful for quick experiments, while the per-hidden schedule keeps layered models explicit about which hidden stage receives how much perturbation.

Parameters:

• this - Target network instance.
• configuration - Global standard deviation or one value per hidden layer.

Returns: Nothing.

getLastSkippedLayers

getLastSkippedLayers(): number[]

Read the last hidden-layer indices skipped by stochastic depth so diagnostics can inspect which layers were bypassed in the most recent forward pass.

Parameters:

• this - Target network instance.

Returns: Snapshot of the last skipped hidden-layer indices.

getRuntimeRegularizationStats

getRuntimeRegularizationStats(): Record<string, unknown> | null

Read regularization statistics collected during training so callers can inspect dropout, noise, and penalty telemetry without direct access to internal runtime fields.

Parameters:

• this - Target network instance.

Returns: Last regularization stats payload or null when none exists yet.

getTrainingStep

getTrainingStep(): number

Read the current training-step counter so external schedulers, dashboards, and callback logic can align runtime control decisions with iteration progress.

Parameters:

• this - Target network instance.

Returns: Current training step.

setRandom

setRandom(
  randomFunction: () => number,
): void

Replace the network random number generator.

This lets advanced callers share one deterministic source across mutation, stochastic depth, DropConnect, and other runtime randomness.

Parameters:

• this - Target network instance.
• randomFunction - RNG function returning values in $[0,1)$.

Returns: Nothing.

setStochasticDepth

setStochasticDepth(
  survivalProbabilities: number[],
): void

Configure stochastic depth with one survival probability per hidden layer.

Matching survival values to hidden layers keeps the runtime contract explicit and avoids silently applying one layer's policy to another.

Parameters:

• this - Target network instance.
• survivalProbabilities - Survival probabilities for each hidden layer.

Returns: Nothing.

setStochasticDepthSchedule

setStochasticDepthSchedule(
  schedule: StochasticDepthSchedule,
): void

Set the stochastic-depth schedule function that updates survival probabilities over time using the current training step and previous schedule state.

Parameters:

• this - Target network instance.
• schedule - Function mapping the current step and schedule to a new schedule.

Returns: Nothing.

setWeightNoiseSchedule

setWeightNoiseSchedule(
  schedule: (step: number) => number,
): void

Set a dynamic scheduler for global weight noise so each training step can derive a new standard deviation from one explicit and testable policy function.

Parameters:

• this - Target network instance.
• schedule - Function mapping the current training step to a standard deviation.

Returns: Nothing.

testForceOverflow

testForceOverflow(): void

Force the next mixed-precision overflow path.

This is a test-oriented hook used to exercise loss-scale recovery logic without waiting for a real floating-point overflow.

Parameters:

• this - Target network instance.

Returns: Nothing.

architecture/network/runtime/network.runtime.diagnostics.utils.ts

Runtime diagnostics and safety helpers for the public Network class.

This chapter owns the public readers and small runtime controls that expose training-health state, activation-ordering diagnostics, DropConnect policy, and dropout-mask cleanup without changing the network topology itself.

appendSuggestion

appendSuggestion(
  suggestions: string[],
  suggestion: string,
): string[]

Append one suggestion string only when it is not already present.

Parameters:

• suggestions - Existing suggestions.
• suggestion - Suggested next action.

Returns: Updated suggestions list.

cloneSchedulingDiagnostics

cloneSchedulingDiagnostics(
  diagnostics: ActivationSchedulingDiagnostics,
): ActivationSchedulingDiagnostics

Clone the diagnostics snapshot so callers cannot mutate runtime state.

Parameters:

• diagnostics - Diagnostics snapshot to clone.

Returns: Detached diagnostics snapshot.

createCompiledSchedulingDiagnostics

createCompiledSchedulingDiagnostics(
  network: default,
  activationSchedule: ActivationSchedule,
): ActivationSchedulingDiagnostics

Build the standard diagnostics snapshot for a compiled activation schedule.

Parameters:

• network - Target network instance.
• activationSchedule - Cached compiled schedule.

Returns: Scheduling diagnostics snapshot.

createDefaultSchedulingDiagnostics

createDefaultSchedulingDiagnostics(
  network: default,
  runtimeNetwork: NetworkRuntimeDiagnosticsInternals,
): ActivationSchedulingDiagnostics

Build a safe default diagnostics snapshot when no explicit scheduling record exists yet.

Parameters:

• network - Target network instance.
• runtimeNetwork - Runtime diagnostics internals.

Returns: Default scheduling diagnostics snapshot.

disableDropConnect

disableDropConnect(): void

Disable DropConnect by resetting the drop probability to zero on the target network instance.

Parameters:

• this - Target network instance.

Returns: Nothing.

enableDropConnect

enableDropConnect(
  probability: number,
): void

Enable DropConnect with a probability in $[0,1)$.

Parameters:

• this - Target network instance.
• probability - DropConnect probability.

Returns: Nothing.

getActivationSchedulingDiagnostics

getActivationSchedulingDiagnostics(): ActivationSchedulingDiagnostics

Read a human-friendly snapshot of the current activation-ordering contract.

The snapshot explains whether activation is using a compiled schedule or a fallback path, whether topology is currently dirty, and what callers should do next when cycles or stale caches prevent the preferred schedule.

Parameters:

• this - Target network instance.

Returns: Activation scheduling diagnostics snapshot.

getLastGradClipGroupCount

getLastGradClipGroupCount(): number

Read the last recorded gradient-clipping group count from the most recent optimizer step.

Parameters:

• this - Target network instance.

Returns: Last gradient-clipping group count.

getLossScale

getLossScale(): number

Read the currently active mixed-precision dynamic loss scale from the network training state.

Parameters:

• this - Target network instance.

Returns: Current loss scale.

getRawGradientNorm

getRawGradientNorm(): number

Read the raw (pre-clip) gradient norm recorded during the most recent backward pass.

Parameters:

• this - Target network instance.

Returns: Last raw gradient norm.

getTrainingStats

getTrainingStats(): TrainingStatsSnapshot

Read a consolidated training-health snapshot including gradient norms, loss scale, and mixed-precision event counters.

Parameters:

• this - Target network instance.

Returns: Training statistics snapshot.

resetDropoutMasks

resetDropoutMasks(): void

Reset every dropout mask to 1.

This is useful after training so later inference does not inherit transient node-level dropout state from a previous activation pass.

Parameters:

• this - Target network instance.

Returns: Nothing.

TrainingStatsSnapshot

Snapshot of key training-health metrics collected from the most recent backward pass.

Aggregates gradient norm readings, dynamic loss scale state, optimizer step index, and mixed-precision event counters so the public diagnostic reader can return a coherent bundle instead of requiring separate calls for each individual metric.

architecture/network/runtime/network.runtime.errors.ts

Raised when a pruning schedule window size is zero, negative, or otherwise falls outside the valid range accepted by the activation-ordering runtime.

NetworkRuntimeDropConnectProbabilityRangeError

Raised when the DropConnect drop probability falls outside the required half-open interval [0, 1) for valid stochastic connection masking.

NetworkRuntimeLayeredWeightNoiseRequiredError

Raised when per-hidden-layer weight noise is requested but the target network was not constructed with an explicit layered topology.

NetworkRuntimePruningScheduleWindowError

Raised when a pruning schedule window size is zero, negative, or otherwise falls outside the valid range accepted by the activation-ordering runtime.

NetworkRuntimeStochasticDepthEntryCountError

Raised when the count of stochastic-depth survival entries does not match the number of hidden layers in the network topology.

NetworkRuntimeStochasticDepthLayeredNetworkRequiredError

Raised when stochastic depth is requested but the target network was not constructed with an explicit layered topology as required.

NetworkRuntimeStochasticDepthSurvivalArrayError

Raised when the stochastic-depth survival probability input is not an array of per-layer probability values as required by the runtime.

NetworkRuntimeStochasticDepthSurvivalRangeError

Raised when a stochastic-depth survival probability value falls outside the open-closed interval (0, 1] required for valid layer retention.

NetworkRuntimeTargetSparsityRangeError

Raised when pruning target sparsity is outside the open interval (0, 1).

NetworkRuntimeWeightNoiseConfigurationError

Raised when the weight-noise configuration contains an unexpected shape, is missing required fields, or carries incompatible type combinations.

NetworkRuntimeWeightNoiseEntryCountError

Raised when the number of hidden-layer weight-noise entries does not match the hidden-layer count in the network topology.

NetworkRuntimeWeightNoisePerLayerRangeError

Raised when a per-hidden-layer weight-noise standard deviation is negative; each layer entry must be zero or a positive value.

NetworkRuntimeWeightNoiseStdDevRangeError

Raised when the weight-noise standard deviation is negative; only non-negative values produce a well-defined Gaussian noise distribution.