Figure 3From: Discovering and mapping chromatin states using a tree hidden Markov modelFree energy for approximate inference methods. Free energy for different inference methods are compared, with parameters learned using exact inference. The test dataset is restricted to a 2 MB region of chromosome 22 with only three cell types and K = 5. The approximate methods use the parameters (learned by the exact method) and only perform inference steps. Note that for the exact algorithm (clique), the free energy equals the negative log-likelihood.Back to article page