CST 499 - Week 5

 Hey everyone,

This week, I made meaningful progress in both understanding deep learning concepts and building practical skills with the fastai workflow. I learned that modern deep learning makes tasks possible that were extremely difficult before 2015, such as distinguishing bird photos from forest photos with relatively little code and training time. I also developed a much clearer picture of what images are to computers: numerical pixel data that can be downloaded, resized, organized, cleaned, and fed into models.

I learned how to search for and download images, organize them into folders, resize them, detect broken files, and remove problematic images. I also learned how to use DataBlock and DataLoaders to prepare training data, how to display batches of images to verify the dataset, and how to fine-tune a pretrained model locally to classify images. On top of that, I deepened my understanding of key deep learning ideas such as models, weights, loss functions, gradients, learning rates, validation sets, and overfitting. I also explored the broader fastai ecosystem, including notebooks, Kaggle, Hugging Face Spaces, and Gradio, which made deep learning feel much more practical and real.

1. Project milestones I accomplished this week

This week, I personally accomplished several important milestones.

First, I learned how to collect and prepare image data for a computer vision project. That included downloading bird and forest images, creating folder structures for each category, resizing images, and identifying broken files for removal. This gave me hands-on experience with one of the most important parts of project development: data preparation and cleaning.

Second, I built my understanding of the fastai pipeline for image classification. I learned how to create a DataBlock, turn it into DataLoaders, and inspect image batches before training. This helped me understand how raw files are transformed into model-ready data.

Third, I learned how to build and fine-tune a pretrained vision model on a local machine. I now understand the role of pretrained models such as ResNet, what fine-tuning does, and how to use the trained model to make predictions on new images.

Fourth, I explored model interpretation tools such as confusion matrices, top losses, and data-cleaning workflows after training. I learned that training a model can actually help reveal problems in the dataset, which is an important milestone in developing a more realistic machine learning workflow.

Finally, I strengthened my general workflow skills by becoming more comfortable with Jupyter notebooks, fastai best practices, Kaggle, and the idea of quickly iterating on practical experiments rather than getting stuck on theory first.

2. My plan for next week

Next week, my plan is to build on this week’s foundation by moving from understanding the workflow to applying it more independently.

I plan to create a small but complete image classification project of my own, including collecting and cleaning a dataset, building the DataBlock, training a model, and evaluating its performance. I want to make sure I can go through the whole process without relying too heavily on step-by-step guidance.

I also plan to spend more time improving model quality by experimenting with resizing, augmentation, training epochs, and model interpretation tools such as confusion matrices and top losses. I want to better understand not just how to train a model, but how to improve one systematically.

In addition, I want to continue exploring deployment-related steps, especially exporting a trained model and turning it into a simple Gradio or Hugging Face demo. That would help me connect model training with a real application.

Finally, I plan to review the lecture questionnaires and course notes before studying each lesson so that I can learn more intentionally and retain the key ideas better.

3. Challenges I am currently facing and whether I need instructor assistance

The main challenge I am currently facing is not a lack of interest, but the sheer amount of material. There are many exciting topics, tools, and side paths, and it is easy to feel overwhelmed or want to understand everything at once. My challenge is to stay focused on the most important practical steps instead of getting lost in all the details.

Another challenge is developing confidence in choosing what matters most during project development. For example, I am still learning when to prioritize cleaning data, when to tweak the model, and when to stop iterating and move on. I understand these ideas conceptually, but I need more experience applying them in practice.

At this stage, I do not think I need urgent instructor assistance, but guidance on prioritization would be very helpful. In particular, it would help to know what level of project completeness or polish is expected right now, so I can focus my time effectively and not overcomplicate the work too early.