From 8144c0b4547d5f7f98d6a168b5f49bd6b488ffdb Mon Sep 17 00:00:00 2001 From: Yujin Chung Date: Fri, 11 Dec 2020 12:56:57 +0900 Subject: [PATCH] Figure name fix [EN]14-1.md (#701) Issue #461 --- docs/en/week14/14-1.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/docs/en/week14/14-1.md b/docs/en/week14/14-1.md index ec05e24f4..421e164f3 100644 --- a/docs/en/week14/14-1.md +++ b/docs/en/week14/14-1.md @@ -79,7 +79,7 @@ Let’s assume that $Z_1$, $Z_2$, and $Y_1$ are discrete binary variables, and $ However, we notice that for a given $X$, $E_a$ only has two possible input configurations: $Z_1 = 0$ and $Z_1 = 1$. Similarly, $E_b$ and $E_c$ only have 4 possible input configurations, and $E_d$ has 6. Hence, there is no need for more than $2 + 4 + 4 + 6 = 16$ single factor evaluations. -Hence, we can pre compute the 16 factor values, and put them on the arcs in a trellis as shown in Figure 19 (bottom). +Hence, we can pre compute the 16 factor values, and put them on the arcs in a trellis as shown in Figure 5 (bottom). The nodes in each column represent the possible values of a single variable. Each edge is weighted by the output energy of the factor for the corresponding values of its input variables. With this representation, a single path from the start node to the end node represents one possible configuration of all the variables. The sum of the weights along a path is equal to the total energy for the corresponding configuration. Therefore, the inference problem can be reduced to searching for the **shortest path** in this graph. This can be performed using a dynamic programming method such as the Viterbi algorithm, or the A\* algorithm. The cost is proportional to the number of edges (16), which is exponentially smaller than the number of paths in general. @@ -98,7 +98,7 @@ However, it only applies to factor graphs that are bipartite trees (with no loop Figure 6. -The factor graph shown in Figure 20 is a log domain factor graph for linear structured models ("simple energy-based factor graphs" we are talking about) +The factor graph shown in Figure 6 is a log domain factor graph for linear structured models ("simple energy-based factor graphs" we are talking about) Each factor is a linear function of the trainable parameters. It depends on the input $X$ and on a pair of individual labels $(Y_m, Y_n)$. In general, each factor could depend on more than two individual labels, but we will limit the discussion to pairwise factors to simplify the notation: