Fictional research talk · illustrative data

Stable Adapters for Streaming Forecasts

How small update modules absorb distribution drift without erasing yesterday’s model.

Demo Systems Lab · Methods colloquium · 14-minute talk

One stream
Many regimes
Bounded updates

01 / 08
The operating regime moves faster than the model.
Motivation

A frozen forecast becomes confidently stale.

The illustrative stream shifts every few hours. Full retraining is too slow; unconstrained online updates are too forgetful.

Research question: can a small adapter track change while the base model remains stable?
Regime ARegime BRegime C

Solid: incoming target · Dashed: frozen forecast

Illustrative stream · not an empirical claim
1 · Motivation
02 / 08
Objective: adapt locally, constrain globally.
Formulation

One loss balances fit and memory.

Lt = loss(fbase,At(xt), yt) + lambda ‖At − At−1‖²Illustrative simplified objective; the original expression should remain in speaker notes when adapting a real paper.

Read it aloud

First term: fit the current window.

Second term: pay for abrupt adapter movement.

lambda: decides how much yesterday constrains today.

The base parameters stay fixed.
2 · Formulation
03 / 08
Method: a bounded update loop around a fixed base.
Method

Four operations repeat for every stream window.

01 / OBSERVE

Buffer a window

Collect the newest labelled slice and drift summary.

02 / UPDATE

Fit the adapter

Optimize the small module against the current window.

A_t only
03 / CHECK

Test stability

Reject steps that cross the memory budget.

04 / COMMIT

Advance safely

Promote the adapter and keep the base checkpoint intact.

3 · Method
04 / 08
Why the update stays controlled.
Analysis

The regularizer turns drift into a budget.

Illustrative claim

If the per-window gradient and adapter step are bounded, cumulative change grows with the sum of accepted budgets—not the number of observations alone.

This frame communicates proof intuition, not a formal theorem.

1Bound each local gradient estimate.
2Translate the stability penalty into a step-size ceiling.
3Sum only accepted steps across regime boundaries.
4Keep the full derivation in notes or appendix, not on this frame.
4 · Analysis
05 / 08
Evaluation: stability improves without freezing adaptation.
Results

The adapter tracks three regimes with lower forgetting.

0.80.60.40.2Window 1Window 12Window 24Shift 1Shift 2
Stable adapterUnconstrained online updateIllustrative demo data · 24 windows
−31%forgetting score versus unconstrained update
+4%current-window error versus full retraining

Scope: fictional benchmark values for presentation demonstration only.

5 · Results
06 / 08
The method buys stability; it does not solve every shift.
Discussion

Limits to name before the conclusion.

Current limits

  • Requires delayed labels for each update window.
  • Budget selection is sensitive to shift speed.
  • A fixed base cannot learn entirely new features.
6 · Discussion
07 / 08
Conclusion

Small modules can absorb change without rewriting the whole model.

01Freeze the base; update a compact adapter.
02Price abrupt movement with an explicit stability term.
03Report adaptation and forgetting together.
Next question

When should the base itself be allowed to move?

Open the discussion there.

08 / 08
· ·
Deck studio

Export an editable PPTX

PowerPoint export runs on your computer, next to this file.

  1. If you edited this deck, choose Export → HTML first so the export includes your changes.
  2. Run node scripts/export-pptx.mjs your-deck.html your-deck.pptx from the skill folder, or ask your agent to export this deck to PPTX.
  3. To edit with autosave and export straight from this menu, open the deck with node scripts/serve-deck.mjs your-deck.html.