underpruning

Underpruning refers to a pruning approach or outcome in which too little pruning is performed, leaving a model, algorithm, or plant with more complexity than is optimal for the intended task or environment. It can arise from conservative pruning thresholds, insufficient data on the cost of complexity, or neglecting a pruning step. The term is used across fields such as machine learning and horticulture, but generally implies higher complexity than necessary.

In machine learning, pruning reduces unnecessary complexity. In decision trees, pruning removes branches that contribute little to predictive power. Underpruning occurs when pruning is too weak or applied too late, producing a large, highly detailed tree that captures noise in the training data and may overfit, reducing generalization and increasing evaluation variance. It can also increase training time and memory usage.

In neural network pruning, the goal is to remove weights or neurons to create a sparser, more efficient model. Underpruning retains a large portion of the original connections, leading to limited reductions in size or speed improvements and often higher inference cost without proportional gains in generalization.

In horticulture, pruning aims to remove growth to improve plant health, yield, and airflow. Underpruning leaves the canopy too dense, which can hinder light penetration, increase disease risk, reduce fruit quality, and complicate maintenance.

Detection relies on validation performance in ML or physical assessment of canopy structure in plants. Remedies include adjusting pruning criteria, applying regularization or cost-complexity pruning in ML, and adopting staged pruning schedules or density targets in horticulture. It contrasts with overpruning, which aggressively reduces complexity and can cause underfitting or poor plant structure.