← All AI/ML open problems

Learning dynamics & scaling · Data-efficient learning / Open question

Learning Without Losing the Ability to Learn

Saturation and redundant neurons can trap training dynamics, so a compact network may cease adapting even when new data are informative. Efficiency payoff: Preserve useful features while rejuvenating only the capacity needed for change, reducing full retraining and replay requirements.

OpenEditorially scopedFramework needs definitionCompiled 2026-11-20

Seed catalog · Showing the bundled record.

Problem statement

Observe xt∼N(0,Id)x_t\sim N(0,I_d) and yt=ft(xt)+ξty_t=f_t(x_t)+\xi_t, where ft(x)=∑j=1kaj,ttanh⁡(uj,t⊤x)f_t(x)=\sum_{j=1}^k a_{j,t}\tanh(u_{j,t}^\top x), ∣aj,t∣,∥uj,t∥≤1|a_{j,t}|,\|u_{j,t}\|\leq1, and independent noise has variance at most σ2\sigma^2. The teacher path is fixed in advance with variation VT=∑t<T(∥at+1−at∥2+∥Ut+1−Ut∥F)V_T=\sum_{t<T}(\|a_{t+1}-a_t\|_2+\|U_{t+1}-U_t\|_F). For a width-mm student using one stochastic-gradient update per example and at most RR neuron resets, determine achievable tracking regret ∑t=1TE[(f^t(xt)−ft(xt))2]\sum_{t=1}^T\mathbb E[(\hat f_t(x_t)-f_t(x_t))^2] as a function of T,VT,m,RT,V_T,m,R.

Assumptions and scope

  • Predict before seeing y_t; independent inputs/noise and an oblivious teacher path prevent anticipatory leakage. Teacher unit labels are fixed when defining variation. A useful subfamily may impose a quantified feature-separation condition, but must report its necessity and constants.
  • Fix initialization, learning-rate/projection rules, which units may be reset and their reset distribution before a positive theorem. Count all model/optimizer state in O(md) storage and prohibit replay in the base comparison; an added replay buffer must be charged.
  • Require a vanishing average tracking error in a specified nontrivial regime such as V_T=o(T), with static-teacher estimation cost and reset cost explicit. Compare to retraining and to the same-width update without resets; do not infer learning from a covariance-rank statistic alone.

What would settle it?

A finite-width update/reset scheme with a quantitative tracking bound and a failure or lower-bound regime for too little plasticity. Connect the guarantee to escaping or avoiding the identified invariant/trapping sets.

Selected for a theoretical or computational target. A fitted scaling curve or improved benchmark score alone does not establish the stated learning-theoretic guarantee.

Status and evidence

Stable-manifold explanations of plasticity loss and utility-weighted perturbation methods already exist. The proposed teacher-path and compute constraints sharpen the empirical agenda into a research benchmark; the optimizer/reset design remains to be fixed and the exact frontier is not certified novel. This is an editorial, source-motivated research target. The preliminary review through 2026-11-20 does not certify novelty or exclude an unlocated resolution.

Preliminary source check; not an expert-verified open status or an independent audit of proposed proofs. The compilation date is not a literature-review cutoff.

Sources and further reading

  1. Joudaki et al. — Barriers for Learning in an Evolving World: Mathematical Understanding of Loss of Plasticity ↗partial theory · source checked 2026-11-20

    Primary abstract checked: stable trapping manifolds from saturated/cloned units explain a simplicity–adaptability tension. Mathematical plasticity-loss mechanisms already exist; finite-width tracking under drifting tasks is a separate target.

  2. Elsayed and Mahmood — Addressing Loss of Plasticity and Catastrophic Forgetting in Continual Learning ↗related algorithm · source checked 2026-11-20

    Primary search abstract checked in the lab review: utility-weighted perturbations protect useful units and rejuvenate others. Empirical gains do not establish the drifting-teacher regret frontier below.

  3. Ineffable Intelligence — mission, beliefs and A Note from Dave ↗research agenda · source checked 2026-11-20

    Primary site checked: experience-driven, continual learning without reliance on human data is the stated agenda. The site supplies no corresponding theorem; rewards, environment access and architectures are still inductive assumptions.