Week 18 – Practicing with Kaggle Challenges

Dates: September 28 – October 4
Internship: AI/ML Intern at SynerSense Pvt. Ltd.
Mentor: Praveen Kulkarni Sir


Focus

This week was all about hands-on competition experience — using Kaggle challenges to test my skills in real-world, data-driven scenarios.
After weeks of theory, model tuning, and internal experimentation, the goal was to measure how well those learnings translate to open-ended, public datasets where creativity, efficiency, and robustness all matter.


Goals for the Week

  • Participate in Kaggle competitions across vision and NLP domains
  • Build, train, and submit models using real-world datasets
  • Improve leaderboard scores through feature engineering and tuning
  • Learn collaborative workflows using Kaggle notebooks and discussions
  • Analyze top-scoring kernels and learn from community approaches

Tasks Completed

Task Status Notes
Participated in Kaggle “Digit Recognizer” (MNIST) competition ✅ Completed Achieved >99% accuracy using CNN and data augmentation
Tried “House Prices – Advanced Regression Techniques” ✅ Completed Practiced feature engineering, one-hot encoding, and cross-validation
Participated in NLP “Disaster Tweets” classification task ✅ Completed Used BERT embeddings and fine-tuned for better F1-score
Explored top Kaggle notebooks and implemented improvements ✅ Completed Compared ensemble methods and model stacking
Documented findings and model versions ✅ Completed Stored results and metrics in GitHub for reproducibility

Key Learnings

  • Competitions mirror real-world pressure. Kaggle forces clear thinking under data and time constraints.
  • Feature engineering matters more than architecture sometimes. Clean data often beats complex models.
  • Community learning accelerates growth. Reading public notebooks provided valuable shortcuts and insights.
  • Evaluation metrics guide better models. Understanding RMSE, F1, and log-loss changed how I approached validation.

Challenges and Solutions

Challenge Solution
Handling large datasets with limited compute Used Kaggle’s GPU runtimes and reduced image sizes for faster training
Overfitting on training data Added dropout, regularization, and early stopping
Poor initial leaderboard ranking Iteratively tuned features and hyperparameters
Comparing models efficiently Created a local validation pipeline with consistent metrics

References


Goals for Next Week

  • Review all weeks’ learnings and results
  • Prepare final project summary and insights
  • Create a portfolio section showcasing competition solutions
  • Draft the final internship reflection

Screenshots (Optional)

Screenshots of Kaggle leaderboards, notebooks, and result plots from CNN and BERT models.


“Week 18 transformed learning into performance — every competition became a classroom, every submission a lesson.”