Aprende The Relationship Between Likelihood and Loss Functions

Understanding how likelihood relates to loss functions is a central concept in machine learning. When you train a model, you often want to find parameters that make the observed data most probable under your model — this is called maximizing the likelihood. However, most optimization algorithms are designed to minimize a function, not maximize it. To bridge this, you minimize the negative log-likelihood instead. The negative log-likelihood is what becomes your loss function. This means that minimizing loss is mathematically equivalent to maximizing likelihood, and the specific form of the loss depends on the probability distribution you assume for your data.


              123456789101112131415161718192021222324252627
            
import numpy as np

# Gaussian (Normal) negative log-likelihood for regression
def gaussian_nll(y_true, y_pred, sigma):
    n = len(y_true)
    # Negative log-likelihood for Gaussian (ignoring constant terms)
    nll = 0.5 * np.sum(((y_true - y_pred) / sigma) ** 2) + n * np.log(sigma)
    return nll

# Bernoulli negative log-likelihood (cross-entropy) for binary classification
def bernoulli_nll(y_true, y_pred_prob):
    # Clip probabilities to avoid log(0)
    eps = 1e-15
    y_pred_prob = np.clip(y_pred_prob, eps, 1 - eps)
    nll = -np.sum(y_true * np.log(y_pred_prob) + (1 - y_true) * np.log(1 - y_pred_prob))
    return nll

# Example data
y_true_reg = np.array([2.0, 3.0, 4.5])
y_pred_reg = np.array([2.2, 2.8, 4.4])
sigma = 1.0

y_true_clf = np.array([1, 0, 1, 1])
y_pred_prob_clf = np.array([0.9, 0.2, 0.8, 0.7])

print("Gaussian negative log-likelihood:", gaussian_nll(y_true_reg, y_pred_reg, sigma))
print("Bernoulli negative log-likelihood:", bernoulli_nll(y_true_clf, y_pred_prob_clf))

This principle — that minimizing loss is the same as maximizing likelihood — shapes the loss functions used in machine learning. In regression tasks, you typically assume your data follow a Gaussian distribution, which leads to the mean squared error loss. In classification, assuming a Bernoulli or multinomial distribution leads to cross-entropy loss. These losses are not arbitrary; they directly emerge from the statistical properties of the underlying probability distributions. By grounding your loss functions in likelihood, you ensure your models are statistically consistent with the data-generating process, which is key for both interpretability and performance.

¿Todo estuvo claro?

¡Gracias por tus comentarios!

Sección 4. Capítulo 1

Pregunte a AI

Pregunte lo que quiera o pruebe una de las preguntas sugeridas para comenzar nuestra charla

Desliza para mostrar el menú


              123456789101112131415161718192021222324252627
            
import numpy as np

# Gaussian (Normal) negative log-likelihood for regression
def gaussian_nll(y_true, y_pred, sigma):
    n = len(y_true)
    # Negative log-likelihood for Gaussian (ignoring constant terms)
    nll = 0.5 * np.sum(((y_true - y_pred) / sigma) ** 2) + n * np.log(sigma)
    return nll

# Bernoulli negative log-likelihood (cross-entropy) for binary classification
def bernoulli_nll(y_true, y_pred_prob):
    # Clip probabilities to avoid log(0)
    eps = 1e-15
    y_pred_prob = np.clip(y_pred_prob, eps, 1 - eps)
    nll = -np.sum(y_true * np.log(y_pred_prob) + (1 - y_true) * np.log(1 - y_pred_prob))
    return nll

# Example data
y_true_reg = np.array([2.0, 3.0, 4.5])
y_pred_reg = np.array([2.2, 2.8, 4.4])
sigma = 1.0

y_true_clf = np.array([1, 0, 1, 1])
y_pred_prob_clf = np.array([0.9, 0.2, 0.8, 0.7])

print("Gaussian negative log-likelihood:", gaussian_nll(y_true_reg, y_pred_reg, sigma))
print("Bernoulli negative log-likelihood:", bernoulli_nll(y_true_clf, y_pred_prob_clf))

¿Todo estuvo claro?

¡Gracias por tus comentarios!

Sección 4. Capítulo 1