YZ Sözlüğü
Yapay Zekanın tam sözlüğü
Generative Modeling
Ability of a DBN to learn the underlying probability distribution of training data, allowing it to generate new plausible data instances that have never been observed.
Contrastive Divergence (CD)
Approximation algorithm for the log-likelihood gradient used in training RBMs, which uses a short Markov chain (often just one step) to estimate model expectations.
Fine-Tuning
Optional supervised training phase that follows the unsupervised pre-training of a DBN, where weights are globally adjusted to optimize a specific task like classification.
Persistent Markov Chain (PCD)
Improvement of Contrastive Divergence where the state of the Markov chain used for sampling is not reset between weight updates, accelerating convergence and improving generative model quality.
Mode Collapse
Phenomenon in generative models, including DBNs, where the model learns to generate only a limited subset of the modes of the real data distribution, producing samples lacking diversity.
Greedy Layer-Wise Pre-training
Training strategy for DBNs where each RBM is trained sequentially, using the output of the previous hidden layer as input for the next one, simplifying the overall optimization of the deep model.
Sampling from the Learned Distribution
Process of generating new data by a DBN by initializing visible units randomly and performing alternating Gibbs sampling across layers until convergence, producing a sample from the modeled distribution.
Partition Function (Z)
Normalization constant in the probability function of a DBN, obtained by summing the exponential of the negative energy over all possible configurations, whose exact calculation is the main training difficulty.