I have used t-sne to visualize a set of images which I have used for training a binary classifier. Let us assume that the binary classifier is trained to detect cat(1) vs. no-cat(0). I have used the final features (last hidden layer) of 64 values to run the t-sne algorithm. Following image shows the output:
The color shows the probability (red showing towards 1, blue showing towards 0) of the trained model and the shape shows if the input image is cat or not (triangle showing cat, circle showing not cat). So, red circles and blue triangles are false detection and missed detection respectively.
Now can we visualize the hyperplane that separates the 64-dimensional space in the t-sne output space (i.e., 2-D space)? From the image it looks like the red and the blue points can be separated by a crooked open curve. Can we plot as well? If we cannot plot it, is it correct way to think that a crooked curve is the true representation of the hyperplane?