shortened commentary on models

app.py

@@ -471,16 +471,17 @@ with gr.Blocks(theme=theme) as demo:
                     Models
                     ------
                     ''')
+                with gr.Row():
                     gr.Markdown('''
                     **1. 	Baseline or Non-DA Network**: This network lacks DA components and only includes feature extractor and task-specific component Training only involves biased source domain data (with train, validation and if available, test splits). The network is trained to minimize the loss objective, which is specific to the task 
                     ''')
                 with gr.Row():
                     gr.Markdown('''
-                    **2. 	DA (without EWC) Network**: This network includes DA specific components like domain classifier and gradient reversal layer in addition to the components of Non-DA (baseline) network. It uses two datasets, i.e., demographic-biased source dataset (same as baseline network with train, validation, and, if available, test splits) and demographic-neutral target dataset, which have less bias / no bias. The purpose of domain classifier network is to understand which data sample came from which of the two datasets (source or target).  To achieve unsupervised DA, gradient reversal layer flips the errors (gradient) propagated during backpropagation. This forces the feature extractor to prioritize features that are independent of the data source (i.e., domain-invariant) but still informative for the task at hand (i.e., task-specific). The parameters (weights and biases) of feature extractor and task-specific component are initialized using trained parameters from baseline network and further training is progressed. 
+                    **2. 	DA (without EWC) Network**: This network includes DA specific components like domain classifier and gradient reversal layer in addition to the components of Non-DA (baseline) network. It uses two datasets, i.e., demographic-biased source dataset (same as baseline network with train, validation, and, if available, test splits) and demographic-neutral target dataset, which have less bias / no bias. 
                     ''')
                 with gr.Row(): 
                     gr.Markdown('''
-                    **3. 	DA (with EWC) Network**: This network is similar to DA (without EWC) network, but enforces a regularization constraint on the parameters of feature extractor and task-specific component using EWC algorithm while training. Like in DA (without EWC) network, the parameters (weights and biases) of feature extractor and task-specific component are initialized using trained parameters from baseline network and further training is progressed. Training involves both source and target domain data.
+                    **3. 	DA (with EWC) Network**: This network is similar to DA (without EWC) network, but enforces a regularization constraint on the parameters of feature extractor and task-specific component using EWC algorithm while training. 
                     ''')
             with gr.Column():
                 with gr.Row():