Uploaded by Hello World

Assignment1-30032020-083147pm

advertisement
Assignment 1
DSC 707 Deep Learning
MSDS-3A – Spring 2020
Task 1: Implementing Back Propagation (by Hand)
Back propagation is the core of training neural networks. The objective of back propagation is to
optimize the weights so that the network can learn how to correctly map arbitrary inputs to
outputs. You need to apply back propagation to the given neural network (by hand) and show
the weight updates resulting from the first iteration.
Input: [10 20];
Learning Rate: 0.2;
Target: [1 0]
Task 2:
2: Digit Recognition using MLP (Keras)
Digit recognition is a classical pattern classification problem. For this task, you will be working on
the CVL single digit (normalized) database. The database comprises 7000 training images, 7000
ASSIGNMENT#1 | DSC 707 Deep Learning
validation images and more than 21,000 test images. Refer to the readme file with the database
for more information.
Refer to file Assignment1.py which you need to update. The file contains functions to load all
images (as vectors) in a folder and return the corresponding labels. Design an MLP (using Keras
Sequential API), train the model and evaluate it on the test set. You may use any distribution of
training and test sets.
Prepare a comprehensive report with detailed analysis on how the performance varies with
respect to number of layers, activation functions, amount of training data, number of examples
per class etc.
Submission
Submission Requirements
Submit your assignment in the form of a technical report. The report must be formatted
according to IEEE two column conference paper template1. Attach the source code of your
implementation as an annexure to the report. You are free to seek help from any sources with
proper citation/acknowledgment.
Submission Date
Submit the printed copy of your assignment on 4th March 2020 by 1900.
Late submissions will NOT be accepted.
+++++++++++
1
https://www.ieee.org/conferences/publishing/templates.html
ASSIGNMENT#1 | DSC 707 Deep Learning
Download