Les modèles basés sur l'énergie (EBMs) sont un type de modèle probabiliste utilisé en apprentissage automatique and intelligence artificielle to represent complex distributions over data. They operate on the principle that each configuration of the model corresponds to an energy value, with lower energy indicating a more likely or favorable configuration. The key idea is to learn an energy function that assigns a scalar value to each possible data point, which can then be used to derive probabilities through normalization.
En termes mathématiques, un EBM définit une probability distribution by associating an energy value, denoted as E(x), to each data point x. The probability of a particular data point is calculated using the Boltzmann distribution, which is expressed as:
P(x) = exp(-E(x)) / Z
Ici, Z est la constante de normalisation connue sous le nom de fonction de partition, which ensures that the probabilities sum to one across all configurations. Learning in EBMs typically involves optimizing the energy function, often using techniques like contrastive divergence or other sampling methods.
Il a été démontré que les EBMs sont efficaces dans diverses applications, notamment génération d'image, denoising, and as generative models for apprentissage non supervisé. They can capture complex relationships in the data, making them a powerful tool in the field of deep learning and beyond.