Day-8 of VIth Semester

First hour was as usual computer vision. Convolution Neural Networks — CNN were started. From the past few classes, whichever I had left unread and unwritten I will take them over now, no matter how long the writing goes. I am running out of time and all this just makes me realize more deeply that, though it’s just a week of classes, so much of syllabus has been taught and I am not upto date with it. Multiple tasks are there I have to prioritize well and complete them without excuse. Because of time constraints I will complete from notes whatever I have written and no matter what, I will prioritize this task and complete these for the subjects.

Convolutions are required because a feature vector that’s useful in one part of image maybe useful in other parts too — parameter sharing. Second reason being spatial connections — in each layer, output value depends only on small number of inputs. Odd size grid is preferred as it allows a centre. Generally output feature map has smaller dimensions than input but can be made of equal size using padding. Stride is a parameter that governs how many cells the filter is moved in the input to calculate next cell output. Padding has the following benefits as it allows processing of entire image with filter, which means allows to use a convolution layer without necessarily shrinking the height and width of image, which is necessary for building deeper networks. Multiple features can be detected using multiple filters. The output will have the same number of channels as the number of filters. 1x1 convolution is special case of depth wise CNN. This effect is to flatten or merge channels together which can save computation later in network. Some activation functions are added and pooling layer is the building block of CNNs. I shall discuss the vaguely discussed concepts in tomorrow’s master notes.

Second hour was security, AES algorithm was discussed. This was made to replace DES with improved security and standardized by NIST for its efficiency. Adding sub-bytes, shifting rows, shifting columns and then adding round keys are the major steps. Adding round keys is skipped in last round. Depending the key size, number of bytes are decided. For 128-bit key, 16 bytes are required so 16 rounds and 4x4 matrices. I shall discuss in depth about each of them in next note.

Data analytics was a 2 hour lecture, it was really interesting as mam made the best possible ways to build out thought process. My most favorite would be parapsychology one. I will discuss in next writing. I have lot of things to search about also, so I will make it tomorrow.

Yours truly