Richard Sutton announced the availability of his student Fernando Hernandez Garcia’s PhD thesis, “Selective Reinitialization Algorithms for Preventing Plasticity Loss in Artificial Neural Networks.” In the abstract Sutton shared, Hernandez Garcia describes plasticity loss as the tendency of neural-network systems to become progressively worse at learning from new data.
Where plasticity breaks down
The thesis reports that plasticity degraded during continual learning across fully connected networks, convolutional networks, residual networks and vision transformers. According to the abstract, normalization, residual connections and regularization did not prevent the problem in the systems studied.
Hernandez Garcia’s experiments also found that plasticity loss was not unavoidable. Continual backpropagation periodically resets selected units, turning initialization into an ongoing process rather than a one-time step. The thesis generalizes that idea into selective unit reinitialization, which ranks units by utility, chooses which ones to reset and assigns them new values.
Units and weights offer different trade-offs
The thesis also proposes selective weight reinitialization. Hernandez Garcia reports that resetting units caused less disruption to network outputs, but required architecture-specific engineering because the definition of a unit varies. Resetting individual weights could be applied to arbitrary network architectures, but could reduce training stability.
According to the abstract, L2 regularization stabilized the weight-level method at the cost of another hyperparameter. Both the unit- and weight-level approaches maintained plasticity across the systems tested.
Comments