stability–plasticity dilemma

The trade-off, in any adaptive system that learns from a changing stream of experience, between plasticity — the capacity to acquire new knowledge quickly — and stability — the retention of existing knowledge against interference. In neural networks with shared weightsweights, n.: The learned numbers inside a neural network that set how strongly each part influences the next. The finished set of weights is, in effect, the trained model itself. the two pull against each other: high learning rates and unconstrained updates favor plasticity but cause catastrophic forgettingcatastrophic forgetting, n.: When a neural network trained on something new gets suddenly and badly worse at what it could already do, because the new training rewrites the internal settings the old skills relied on., while strong constraints favor stability but can freeze learning (loss of plasticityloss of plasticity, n.: The gradual loss of a neural network's ability to learn new things after long training on changing tasks, as parts of the network go inactive or become stuck.). continual learningcontinual learning, n.: The ability of an AI to keep learning new things over time from a stream of experience, without forgetting what it already knew and without being retrained from scratch. methods are commonly classified by how they strike the balance: replay, importance-weighted regularization such as elastic weight consolidationelastic weight consolidation, n.: A way to help a neural network learn a new task without forgetting an old one, by making the settings that mattered most for the old task harder to change, as if tied down with springs., parameter isolation, or architectures with fast- and slow-learning components inspired by the complementary roles of hippocampus and neocortex. The dilemma also frames fine-tuningfine-tuning, n.: Taking an AI model that has already learned a lot and training it a little more on a smaller, focused set of examples so it gets better at one job or behaves a certain way. choices for large models, such as using low-rank adaptationlow-rank adaptation, n.: A cheap way to customize a big AI model: instead of changing all its billions of numbers, you train a small add-on of a few million numbers that slightly adjusts how it behaves. or a KL penaltyKL penalty, n.: A cost added during training that grows as an AI model drifts away from its starting version, so it improves at the goal without forgetting how to write normally or gaming its scorer. to limit drift.

Freezing the lower layers bought stability but left too little plasticity to learn the new dialect.