We were discussing stochastic gradient method for text mining.
The stochastic gradient descent also hoped to avoid the batch processing and repeated iterations of the batch that came to refine the gradient.  The overall cost function is written with a term that measures how the hypothesis is holding on the current datapoint. To fit the next data point, the parameter is modified. As it scans the data points, it may wander off but eventually reach near the optimum. The iteration over the scanning is repeated so that the dependence of the sequence of data points is eliminated. For this reason, the datapoints are initially shuffled which goes along the lines of a bag of word vectors rather than the sequence of words in a narrative. Generally, the overall iterations of all data points are restricted to a small number, say 10, and the algorithm makes improvements to gradients from one datapoint to another without having to wait for all of the batch.  
The steps to take in order to utilize this method for text mining, we first perform feature extraction of the text. Then we use the stochastic gradient descent method as a classifier to learn the distinct features and Finally we can measure the performance using the F1 score. Again, vector, classifier and evaluator become the three-stage processing in this case.

We explore regions of interest detection using neural net. Here we begin with a feature map. A residual learning network can take the initial sample as input and output feature maps where each feature map is specialized in the detection of topic. If the model allows backpropagation for training and inference, we could simultaneously detect topics which are location independent as well as the position of occurrence.   

With the feature map,  there are two fully connected layers formed one for box-regression and another for box-classification.

The bounding boxes are the proposals. each box from the feature map is evaluated once each for regression and classification.
The classifier detects the topic and the regressor adjusts the coordinates of the bounding box.
We mentioned that the text is a bag of words and we don't have the raster data which is typical with image data. The notion here is merely the replacement of a box with a bag so that we can propose different clusters. If a particular bag emphasises one and only one cluster, then it is said to have detected a topic.The noise is avoided may even form its own cluster.