July 2, 2025
Day 27 - Laying the Foundation: Model Refinement and Presentation Kickoff
What I Learned
Today marked another productive day in my summer internship journey. The main focus of the day was twofold: continuing the core technical work on our AI ECG diagnostic project and beginning preparations for our final group presentation. In the morning, I focused on refining the 1D CNN + Transformer model we’re training on the frequency-domain ECG data. After identifying a few inconsistencies in dataset alignment yesterday, I double-checked the sample sizes for each input (signal array and positional embedding) and ensured that the data loader now correctly handles both modalities. I also started incorporating a learning rate scheduler and early stopping callback into the model’s training loop to optimize performance and reduce overfitting. These changes should help improve the model’s convergence and stability in the next few training rounds. One key highlight from today was how we divided responsibilities among team members for the final presentation. I’ll be responsible for presenting the frequency-domain processing and model architecture, while other teammates will handle the time-domain CNN and the spectrogram-based 2D CNN. We agreed to meet again tomorrow to start creating the slides and rehearse our parts.
Blockers
No Blockers
Reflection
Today was a day of reflection, progress, and forward-thinking in my internship journey. As I sat down to continue work on our ECG classification project, I realized how much I’ve grown in just a few weeks — not only technically but also in how I approach problems and collaborate with others. The first part of the day was spent debugging and refining the 1D CNN + Transformer model we’re training on the frequency-domain ECG data. At first, I ran into a frustrating shape mismatch error, which reminded me how easily small inconsistencies in input dimensions can derail training. After carefully going through the code and sample sizes, I finally caught the issue — the positional embeddings didn’t align with the signal input. Fixing it felt like a small win, but an important one.Looking back, today wasn’t just about fixing bugs or planning slides. It was a reminder that real progress happens when we combine deep technical work with clear communication. As the presentation approaches, I’m excited — and a little nervous — but grateful to be part of something meaningful.