Scribd
Upload
124 views

E1213 PRNN: Assignment 1 - Basic Models: Prof. Prathosh A. P. Submission Deadline: 1st March 2022

This document outlines an assignment on implementing basic machine learning models for 5 tasks: binary classification, multi-class classification, bounding box regression, frame classification on audio data, and generative models. Students are asked to implement models like Bayes classifiers, KNN, linear models, and GMMs from scratch without using machine learning libraries. Results must be reported in a 4-page LaTeX report with graphs, interpretations, and observations. The deadline for submission is March 1st 2022.

Uploaded by

rishi gupta
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
124 views

E1213 PRNN: Assignment 1 - Basic Models: Prof. Prathosh A. P. Submission Deadline: 1st March 2022

This document outlines an assignment on implementing basic machine learning models for 5 tasks: binary classification, multi-class classification, bounding box regression, frame classification on audio data, and generative models. Students are asked to implement models like Bayes classifiers, KNN, linear models, and GMMs from scratch without using machine learning libraries. Results must be reported in a 4-page LaTeX report with graphs, interpretations, and observations. The deadline for submission is March 1st 2022.

Uploaded by

rishi gupta
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 3

E1213 PRNN: Assignment 1 - Basic Models

Prof. Prathosh A. P.
Submission deadline: 1st March 2022

1 Introduction
This assignment is about implementation of primitive ML algorithms for 5 dif-
ferent tasks. The tasks would remain fixed across all the assignments while the
algorithms and experiments to be conducted would vary.

2 Problem statements
We consider five distinct tasks for the assignments which are described below:

2.1 Binary Classification problem

Here we consider a 2 class classification problem on image data. The dataset is


PneumoniaMNIST which can be found here:
https://zenodo.org/record/5208230#.YfqjwFhBxHQ. It contains images from
chest X-rays belonging to two classes - Normal and Pneumonia. The metrics to
be computed are: classification accuracy, AUC and F1 score. Training is to be
done on the training split and test on the test split.

2.2 Multi-Class Classification Problem

In this task, we look at an 8-class classification problem on Blood Cell Micro-


scope images. The dataset is called the BloodMNIST which can be obtained
from this link - https://zenodo.org/record/5208230#.YfqjwFhBxHQ. The met-
rics are same as above.

2.3 Bounding box regression Problem

In this problem, we are given images of traffic signs and the task is to find out
the bounding boxes that encompass the sign in the image. The dataset is avail- We'll see about this
able here - https://www.kaggle.com/andrewmvd/road-sign-detection. There is later
an xml annotation file that has the co-ordinates of the boxes. For example
- <bndbox> <xmin>98</xmin> <ymin>62</ymin> <xmax>208</xmax>
<ymax>232</ymax> </bndbox>. The task is to take the input image and

1
regress over the co-ordinates (4) of the box. The metrics here are the mean
MSE, mean MAE and mean Intersection over Union (mIoU).

2.4 Frame classification on audio data


In this problem, the task is to classify every sample of a speech/audio signal. We
use the TIMIT dataset for this purpose - https://www.kaggle.com/mfekadu/darpa-
timit-acousticphonetic-continuous-speech. Consider one of the Directories (DRx)
each from the Train and Test sets for all your experiments. The task here is to
classify every sample of the utterance to be belonging to vowel or not vowels.
The ground truth information has to be generated from the .phn file that ac-
companies every .wav file. It lists the phonemes corresponding to time intervals
in the utterance. Eg - 0 3050 h# 3050 4559 sh 4559 5723 ix 5723 6642 hv 6642
8772 eh 8772 9190 dcl 9190 10337 jh 10337 11517 ih 11517 12500 dcl. Define
Vowels to be all phonemes that contain /a,e,i,o/ and u in them. The metrics
are average true positive, average true negative, average false positive and av-
erage false negative. For this problem, use short segments of speech signals (of
duration 10 to 40 ms) as data points. Either use RAW speech or features such
as MFCCs or LPCs may be used as input space.

2.5 Generative Models

In this module, we build generative models on Tinyimage net (https://www.kaggle.com/c/tiny-


imagenet/data). Use Frechet Inception Distance (FID) between 1000 generated
and real data as the metric for evaluation.

3 Models for Assignment 1


• Bayes’s Classifier with several Class Conditional Densities such as Gaus-
sian, GMMs (Have to code up EM) Rishi

• Bayes’s Classifier with different density estimates (ML, MAP and Parzen Mangal
Window and nearest neighbor estimates)
• K-Nearest Neighbor and Naive Bayes classifiers
• Linear Models - Linear Classification/Regression, Linear Models with Poly- Rohit
nomial Kernels, Logistic Classifier/Regression, Fisher LDM

• All the Linear Models with different Regularizers such as L1, L2 and
Elastic
• GMM for the Generative model part - Fit a GMM and sample more points Hiren
from it
Train(X,y) -> theta(parameters)
Predict(X_test, theta) -> y_pred

2
4 Expected Outcomes
• Submit a single python notebook

• You are not allowed to use any ML libraries such as Sklearn


• Try as many different hyper-parameters as possible
• Submit a 4-page report with your graphs/results and interpretations

• Reports SHOULD be in IEEE double column format and strictly prepared


in LaTeX
• Your evaluation depends on the quality of your code, experiments and
your observations of the results and the report that you submit
• Slightest detection of any sort of plagiarism (code/report) will immediately
lead to heavy penalty

You might also like