EIP2 - Session1

By gaganyatri

EIP2 : External Internship Program is a hands-on machine learning course being taught by The Inkers .

The first session was pretty spot on, presented by Rohan Shravan, major takeaway being that ML and CV is progressing at a rapid pace and many concepts have become absolute due to being replaced by efficient algorithms.

Interesting new items heard during the Session

  1. The Bored Cat Experiment
  2. Speech to Image - this was research project 5 years ago.Can be proud of my idea now :P.
  3. Listening to convolution as feature vectors : Video.
  4. gitxiv.com - Papers which also publish source code for validation.
  5. fast.ai - For gathering latest information.
  6. distill.pub - opensource ML papers : apart from arxiv.org.
  7. The logic behind top-5 , top-1 errors as accuracy in ML papers.

Assignment Questions for Session 1

Write articles on the topics mentioned below between 50-100 words in a markdown format .

  1. Convolution
  2. Filters/Kernels
  3. Epochs
  4. 1x1 convolution
  5. 3x3 convolution
  6. Feature Maps
  7. Feature Engineering (older computer vision concept)

Bonus points if you write on any of these:

  1. Activation Function
  2. How to create an account on GitHub and upload a sample project
  3. Receptive Field.
  4. 10 examples of use of MathJax in Markdown

Assignment answers

Name : Sachin S Shetty

Batch : 4

  1. Convolution

It is the basic multiplication of matrices , ie. image matrix of size n x m into the kernel/filter matrix of size p X q . The result would be a convoluted matrix which would have either 1. increase the dimension of the image, for expanding the feature vectors 2. Decrease the dimension for analysis.

  1. Filters/Kernels

    The kernel/filters are one part of the matrix multiplication for convolution . These matrices are generated in random based for required distribution, ie. Gaussian. The randomness allows the variety of images/features to be clubbed together by backtracking.

  2. Epochs

    Epoch is the counter for running the entire convolution & backtracking operation with a reduced set of images. All the images from the training set are not taken completely in a single run, the images are taken in batches to train and improve the accuracy. In each epoch, multiple operations take place . The accuracy is calculated at end of each epoch and then experiment can either continue or else stop based on accuracy .

  3. 1x1 convolution

    Reduces the dimension of the feature map after conolution operation .

  4. 3x3 convolution

    Increase the dimensions of the feature maps after the convolution operation .

  5. Feature Maps

    Feature maps are the result of convolution operation, i.e the multiplication of feature matrix(images) with the filter/kernels.

  6. Feature Engineering (older computer vision concept)

    The earlier standard for solving computer vision problems before the advent of ML/AI, which in part was helped due to cloud computing was Feature Engineering.

    Feature Engineering techniques like hough transform for detecting lines/circles, sobel filter are still prevalent for basic CV tasks.

    These technique have the feature vectors/calculation hardcoded .ie. the values have been calculated and fixed to solve a particular problem/domain .

Bonus points if you write on any of these:

  1. Activation Function

    The activation functions helps to reduce the range of the matrix values/signal values. The prevailing standard function used in ML/AI is ReLu : Rectified Linear Unit, it rejects the negative values and passes 0 instead.
    few other functions used earlier were

    1. sigmoid
    2. tanh
  2. How to create an account on GitHub and upload a sample project. Link to have a personal website is sample_blog.

  3. Receptive Field.

  4. 10 examples of use of MathJax in Markdown