edit pass: cognitive-services-face-articles-batch1

Author: paulth1

articles/cognitive-services/Face/Face-API-How-to-Topics/how-to-use-large-scale.md

@@ -201,13 +201,13 @@ and [Identification](https://westus.dev.cognitive.microsoft.com/docs/services/56
 
 To better utilize the large-scale feature, we recommend the following strategies.
 
-## Step 3.1: Customize time interval
+### Step 3.1: Customize time interval
 
 As is shown in `TrainLargeFaceList()`, there's a time interval in milliseconds to delay the infinite training status checking process. For LargeFaceList with more faces, using a larger interval reduces the call counts and cost. Customize the time interval according to the expected capacity of the LargeFaceList.
 
 The same strategy also applies to LargePersonGroup. For example, when you train a LargePersonGroup with 1 million persons, `timeIntervalInMilliseconds` might be 60,000, which is a 1-minute interval.
 
-## Step 3.2: Small-scale buffer
+### Step 3.2: Small-scale buffer
 
 Persons or faces in a LargePersonGroup or a LargeFaceList are searchable only after being trained. In a dynamic scenario, new persons or faces are constantly added and must be immediately searchable, yet training might take longer than desired. 
 
@@ -222,7 +222,7 @@ An example workflow:
 1. When the buffer collection size increases to a threshold or at a system idle time, create a new buffer collection. Trigger the Train operation on the master collection.
 1. Delete the old buffer collection after the Train operation finishes on the master collection.
 
-## Step 3.3: Standalone training
+### Step 3.3: Standalone training
 
 If a relatively long latency is acceptable, it isn't necessary to trigger the Train operation right after you add new data. Instead, the Train operation can be split from the main logic and triggered regularly. This strategy is suitable for dynamic scenarios with acceptable latency. It can be applied to static scenarios to further reduce the Train frequency.