Sovradattamento (Overfitting)

Overfitting occurs in machine learning when a model learns the training data too well, capturing noise and random fluctuations rather than the underlying data distribution. Instead of learning general patterns that apply to new, unseen data, an overfitted model effectively memorizes the specific examples in the training set. This results in excellent performance on the training data but poor generalization to real-world scenarios. It is often described as "high variance," meaning the model's predictions vary significantly depending on the specific dataset used for training.

Perché si verifica l'overfitting

The primary cause of overfitting is excessive model complexity relative to the amount of available data. If a neural network is too large—meaning it has too many layers or parameters—it can easily memorize the training examples. Other contributing factors include:

• Insufficient Training Data: Small datasets may contain spurious correlations that do not exist in the broader population. Models trained on limited data are prone to learning these accidental patterns.
• Data Noise and Outliers: High levels of noise or unrepresentative outliers in the training data can mislead the model, causing it to adjust its internal parameters to fit anomalies rather than the true signal.
• Extended Training Duration: Training for too many epochs allows the model to continue refining its weights until it fits the noise in the training set. This is often monitored using validation data.

Overfitting vs. Underfitting

It is important to distinguish overfitting from underfitting. While overfitting involves learning too much detail (including noise), underfitting occurs when a model is too simple to capture the underlying structure of the data at all. An underfitted model performs poorly on both the training data and new data, often resulting in high bias. Balancing these two extremes is known as the bias-variance tradeoff.

Preventing Overfitting

Engineers use several techniques to mitigate overfitting and improve model robustness:

• Regularization: Techniques like L1/L2 regularization or adding dropout layers introduce penalties or randomness during training, preventing the model from becoming overly reliant on specific features.
• Early Stopping: Monitoring the loss function on a validation set allows training to be halted once performance on unseen data stops improving, even if training accuracy continues to rise.
• Data Augmentation: artificially increasing the size and diversity of the training set using data augmentation makes it harder for the model to memorize exact images.
• Cross-Validation: Using techniques like k-fold cross-validation ensures the model is tested on different subsets of data, providing a more reliable estimate of its performance.

Esempi reali

Overfitting can have serious consequences when deploying AI in production environments:

• Medical Diagnosis: In AI in healthcare, a model trained to detect skin cancer might overfit to the lighting conditions or ruler markings present in the training images. When deployed in a clinic with different lighting or equipment, the model may fail to correctly identify malignant lesions because it relied on irrelevant background cues.
• Financial Forecasting: A stock price prediction model might overfit to historical market trends that were driven by a specific, non-repeatable event (like a one-time economic crisis). Such a model would likely fail to predict future stock movements accurately because it memorized past anomalies rather than learning fundamental market dynamics.

Code Example: Early Stopping with YOLO26

Using the Ultralytics Platform or local training scripts, you can prevent overfitting by setting an early stopping patience. This stops training if the validation fitness doesn't improve for a set number of epochs.

from ultralytics import YOLO

# Load the YOLO26 model (latest generation)
model = YOLO("yolo26n.pt")

# Train with early stopping enabled (patience=50 epochs)
# If validation metrics don't improve for 50 epochs, training stops.
results = model.train(data="coco8.yaml", epochs=100, patience=50)

Concetti correlati

• Generalization: The ability of a model to adapt and perform well on new, previously unseen data, which is the opposite of overfitting.
• Cross-Validation: A technique for assessing how the results of a statistical analysis will generalize to an independent data set.
• Regularization: Methods used to reduce errors by fitting a function appropriately on the given training set and avoid overfitting.