Fix explnation in Section 2 about training cycles

[twinkarma]

Should be:

• Do inference of all data sample
• Find the loss
• Then back propagate

Instead of backprop for each individual sample as currently shown.