Projects
Spotify Music behavior Analysis
A recommender system tells about the similarity between the users listening to a particular kind of music and based on the ranking of their musical tastes, further songs are recommended in the next cycle. The latent features from user behavioral data helps to cluster similar types of user with K- means clustering. The Bayesian Matrix Factorization I have used to implement the recommender system.The accuracy is 93% . The left diagram is the predicted features for recommended songs ( beat_strength and bounciness ).
Tools Used : Python, R, Tableau, SAS.
For Deployment : AWS sagemaker, Flask. rest API
Spotify_Results -1
Music Clustering
Music Clustering
Music Clustering
This picture tells about the clusterization of particular songs based on their listening behavior by the user.
model performance
Music Clustering
Music Clustering
Accuracy, Precision and Recall of the recommender system model
Noise reduction
Feature engineering
Feature engineering
starting from the top songs depicted primarily filled with noise. After removing noise the melody became clear and seperated with fine grained 0 and 1 labels for classification.
Feature engineering
Feature engineering
Feature engineering
Principle component analysis tells about the probabilistic values of the feature importance.
Object -Detection in GTA-V
Goals and Implementation
- We wanted to see how action points behave when our model detects objects through the boundary box. Surprisingly it gives us some area of identification which could be named as signal or alerts. If our vehicle or other close proximited objects behave abnormally then our boundary box would give a crash alert.
- We wanted to train our model in such a way so that from those generated action points , it will be able to detect it's further spaces and generate commands for movements. Eventually we implemented the self dirving on this simulation.
- Tools Used : Python., Tensorflow, RCNN, Keras, theano, Pytorch
- Deployment: Flask, Rest APi, Docker.
Object detection and Self Driving on GTA-V
Boundary box creation
Boundary box creation
Boundary box creation
Object Detection through boundary box
Crash Alert
Boundary box creation
Boundary box creation
Close proximity of Object Gives crash alert.
Human detection
Boundary box creation
Neural network Training
position detection of pedestrians
Neural network Training
Self Driving on alex net
Neural network Training
The CNN trained on coco data and tested on kitty dataset with an accuracy of 92.45%
Self Driving on alex net
Self Driving on alex net
Self Driving on alex net
ALexnet Training on multiple frames captures with an accuracy of 96.62%.
Final Self Driving
Self Driving on alex net
Self Driving on alex net
After lane detection, I have implemented self-paced learning simulation based on the gradients generated from the frames
Boston City Crime prediction-results
Crime EDA
Feature Enigneering
Feature Enigneering
This EDA tells about at what time in the past 4 years different crime has happened over 1 week;
Feature Enigneering
Feature Enigneering
Feature Enigneering
Tells about which features are most important?
Shap plots for feature relevances
shap plot for positional occrence of crime
shap plot for positional occrence of crime
Shap plot tells about the feature relevance with the most occurred crime which is larceny
shap plot for positional occrence of crime
shap plot for positional occrence of crime
shap plot for positional occrence of crime
shap values of latitude and longitude and their correlation with larceny.
Boston City crime prediction results -2
Census Data analysis
Census Data analysis
Census Data analysis
We collected census data and tried to merge with zip codes
Model Performance
Census Data analysis
Census Data analysis
Model Accuracy on social compositonal Factors
Decision Tree with probabilistic decision Rule
Decision Tree with probabilistic decision Rule
Decision Tree with probabilistic decision Rule
Decision Tree Tells about the probability of occurrence of crime in weekdays vs weekends. prediction accuracy is 96%.
Support Vector machine
Decision Tree with probabilistic decision Rule
Decision Tree with probabilistic decision Rule
clustering of offenses based on the impact of the crime, and the number of occurrences of the crime.
Robust Physical Attack Against Road Sign Classifier
Abstract
Deep Learning is a subset of Machine Learning in artificial Intelligence(AI) with networks that are capable of learning from unstructured data. With this, we have seen Deep Neural Networks (DNNs) to be vulnerable to adversarial examples. These are inputs the attacker intentionally designs for machine learning models to cause them to make a mistake. Specially, classifier networks fell prey to perturbed examples and were mislead into targetted or non-targetted misclassification attacks. The effect of adversarial attack in physical world is different, however, as physical aspects of environment are brought into consideration.We have chosen Robust Physical Perturbation attack for physical world,termed as RP2 as foundation for our project. RP2 generates adversarial example for road sign images captured in practical driving scenario. Our proposed approach is to evaluate the robustness of physical world attack on road sign benchmark dataset and employ various defensive methods against the attack. Defending attacks lessens vulnerability of classifier and is useful to gain insights into future attacks of the same nature. We propose few novel strategies of defense as well as apply established defense methods to make it work for physical road sign perturbations. We analyze the effectiveness of methods, problems encountered and possible improvements.
Road Sign Classifier Sticker Detection results
Sticker detection
Sticker detection
Sticker detection
Unet Performances
Sticker detection
Sticker detection
Seamese net Image Retrieval
Image Retrieval with gradient masking method
Image Retrieval with gradient masking method
Image Retrieval with gradient masking method
Image Retrieval with gradient masking method
Image Retrieval with gradient masking method
