Kaneel Senevirathne

Hi! Thank you for visiting my website. My name is Kaneel. I am a machine learning engineer at the Clinical Trial Coordinating Center at Johns Hopkins University. I am originally from Sri Lanka and currently I am living in Baltimore, Maryland.

For my master's I studied how the brain integrate different senses such as proprioception and vision in stroke patients. I used computational models to describe integration of these senses post-stroke. After studying computational models of the brain, I realized how elegant the brain is at solving problems and how useful the models of the brain are to advance our current knowledge in artificial intelligence (AI). Moving forward, my goal is to learn more about the brain and use the knowledge to advance AI systems. For example, use human multisensory integration models to advance multisensory integration in AI systems. (i.e integration of vision and audition in artificial systems) I am also very interested in applying AI and ML to solve problems in healthcare, autonomy, finance etc. Furthermore, I also enjoy participating in data science competitions on Kaggle and working as a freelancer on Upwork to expand my knowledge.

Apart from AI and cognitive sciences, I enjoy reading & studying about financial markets, autonomous trading (click here to check out my ML-Quant research work) and value investing. During my free time, I like to play chess, workout, cycle and travel.

If you like to check my CV/resume, click here. If you like to vist my github account click here

Research Interests:
Cognitive science, Deep Learning, Reinforcement Learning, Machine Learning, Cognitive Neuroscience, Multisensory Integration, Medical Image Processing

Below I have listed all the projects I have worked on. Each project has a summary, the link to it's github account (if available) and a link to the blog which has a detailed description of the project. You can also go to the blog tab to see these articles.

Stock trading algorithm

Here I used machine learning(ML) to create a trading stratergy. The model uses data from past 30 days and predicts the probability of a stock going up on a given day.

Github repository
Blog article

Multimodal Video Classification

Here I use principals from computational models of multisensory integration of the brain for gender classification. I am using a sample of videos from the VoxCeleb dataset. I have used a CNN model for video classification , a CNN model (using mel spectrograms) for audio classification. Then I used the trained models to create a multimodal network. You can find the multimodal network and comparison of the three models here. Abiding to principals from multisensory integration, the multimodal network produces an optimal prediction.

Github repository
Blog article

Brain Tumor Radiogenomic Classification

MTMT promoter methylation has shown to be a strong predictor of responsiveness to chemotherapy in Glioblastoma cancer patients. Here we have brain scans of patients with and without the MGMT promoter methylation. In this project I used the tensorflow functional API to create a convolutional neural network model with 4 inputs (four scan types) to predict the presence of MGMT in the brain. I also wont a silver medal in my first Kaggle competition!!

Kaggle notebook - Image processing & data visualization

Training agents to play tic-tac-toe using Reinforcement Learning

Here we use Reinforcement Learning (RL) to teach an agent to play tic-tac-toe. The RL agent is playing games with another RL agent and a random agent. After multiple iterations, the agent learn to master the game from past experience. The results show that the RL agent has a significant advantage over the random agent. Also when playing against another RL agent both agents learn to draw games in order to maximize their rewards.

Github repository
Blog article

Simulations of cognitive models from papers. Click here to go to the github repository

Audio-Visual Integration

In this jupyter notebook, I have recreated the audio-visual integration model from the paper "The Ventriloquist Effect Results from Near-Optimal Biomodal Integration". The paper uses a Bayesian model to describe how humans optimally integrate vision and audition.

Github repository

Visuoproprioceptive Integration

In this jupyter notebook, I have created a similar simulation to the one in the paper "Integration of Proprioceptive and Visual Position: An Experimentally Supported Model" The paper uses a 2 dimensional Bayesian model to describe how humans optimally integrate vision and proprioception. Interestingly, the paper shows that the brain integrate the two modalities based on the directional precision of each sense.

Github repository

Multi-rate Model of Motor Adaptation

In this jupyter notebook, I have recreated a the model from the paper "Interacting Adaptive Processes with Different Timescales Underlie Short-Term Motor Learning". In this paper the authors describe how the brain uses two distinct processes for motor adaptation. They further describe that one process reacts fasts to adaptation but retains less whereas the other process reacts slow but retains more.

Github repository

Here are some Machine Learning & Deep Learning Tutorial Notebooks I have created. Click here to visit the Github repository. ML-Quant repository contains a collection of notebooks demonstrating various ML techniques for researching quantative trading strategies.

Gaussian Mixture Model

This notebook shows how to train a Gaussian Mixture Model using Tensorflow Probability.

Jupyter notebook

Probabilistic Logistic Regression

This notebook shows how to create a Probabilistic Logistic Regression model using Tensorflow Probability.

Jupyter notebook

Probabilistic Linear Regression

This notebook shows how to create a Probabilistic Linear Regression model using Tensorflow Probability.

Jupyter notebook

Logistic Regression using gradient descent and maximum likelihood estimation

This tutorial shows how to create a Logistic Regression model from scratch. We are using two different approaches. In one of them we are using gradient descent and in the second method we are using maximum likelihood estimation.

Jupyter notebook

Object classification and localization using tensorflow

This tutorial shows how to create a deep learning model using tensorflow to classify and localize an object in images. Here we are synthesizing example images from the classic game "Where's Waldo?" to detect where Waldo is in an image.

Jupyter notebook
Blog article

Hello world of Deep Learning

This tutorial shows how to use a deep learning technique to estimate the paramters of a quadratic function. Here we use a quadratic function to synthesize data and use backpropagation to learn the parameters a, b and c of the function.

Jupyter notebook
Blog article

Projects, Opinions & Tutorials Click here to visit my blog.

Forecasting Stock Market Trends: A Mixture Model Approach with TensorFlow Probability
30th October 2023

How to create an automated stock recommendation system on GCP
1st January 2023

How I'm Using Machine Learning to Trade in the Stock Market
2nd February 2022

Combining principals from Cognitive Psychology with Deep Learning — Part 1
3rd November 2021

Finding "Waldo" using a simple Convolutional Neural Network
31st August 2021

The "hello, world" of Deep Learning
20th August 2021

Teaching Agents to play tic-tac-toe using Reinforcement Learning
11th August 2021

Preprints

Jacob Johnson, Kaneel Senevirathne, and Lawrence Ngo, “Development and Validation of Suspicious Liver Lesion Deep Learning Algorithm for Computed Tomography,” MedRxiv,[Preprint] 12.15.2021, https://doi.org/10.1101/2021.12.15.21267873

Jacob Johnson, Kaneel Senevirathne, and Lawrence Ngo, “Development of a Natural Language Processing Algorithm for the Classification of Suspicious Liver Lesions from Radiology Reports,” MedRxiv,[Preprint] 12.15.2021, https://doi.org/10.1101/2021.12.15.21267875

Welcome to my website!