Fernando Hernandez Garcia's PhD thesis tests ways to keep neural networks learning over time

Fernando Hernandez Garcia’s thesis examines plasticity loss, the tendency of continually trained neural networks to become worse at learning new data. The abstract reports that plasticity degraded across fully connected, convolutional, residual and vision-transformer systems. Selectively reinitializing units or weights maintained learning ability in the tested systems, though the two approaches trade off output stability against architecture-specific engineering.