Category: Machine Learning

Summary: Testing whether temporary training-data deprivation harms neural networks more when it occurs early in training than when it occurs later.

Biological critical periods suggest that the timing of a disruption can matter as much as the disruption itself. This experiment asks whether artificial neural networks show a comparable effect when normal training data are temporarily replaced by degraded inputs such as noise, zeros, or shuffled labels.

The script trains a model, inserts a deprivation window at different epochs, then resumes standard training and compares final accuracy. By varying both the timing and the type of deprivation, it probes whether early damage causes lasting deficits that later normal experience cannot fully repair.

That makes the project a timing-sensitive learning study rather than a standard robustness benchmark. The central question is whether there is a privileged early window in which training experience has disproportionate long-term influence.

Method: Neural-network training with controlled deprivation windows inserted at different epochs and under multiple deprivation types.

What is measured: Final accuracy, timing sensitivity of deprivation, effect of deprivation type and duration, and evidence for an early critical period.