AI Dictionary of Terms

Convergence

The point during the training of a machine learning model when the loss function stops decreasing significantly and the model’s parameters stabilize, indicating it has found a minimum (local or global).

The Simple Version

The moment a student stops improving their test scores because they’ve mastered the material. In AI, it’s when the model’s errors stop going down, and further training won’t make it any smarter.

Detailed Explanation

During optimization, an algorithm (like Gradient Descent) iteratively updates weights to minimize a loss function. Convergence occurs when the gradient approaches zero, meaning the model is at the bottom of a “valley” in the loss landscape.

Key Characteristics

Business Context

Real-World Analogy

Walking down a mountain in thick fog. You take steps downhill (gradient descent). Convergence is when you finally reach a flat spot where every step you take in any direction starts going uphill again. You’ve reached the bottom of that specific valley.

Code Example

# Conceptual: Checking for convergence in a training loop
prev_loss = float('inf')
convergence_threshold = 0.0001

for epoch in range(1000):
    loss = train_one_epoch(model, data)
    
    # Check if the change in loss is smaller than our threshold
    if abs(prev_loss - loss) < convergence_threshold:
        print(f"Model converged at epoch {epoch}!")
        break
        
    prev_loss = loss

Common Misconceptions

Sources & Further Reading