External Learning & Competitions

2024 OUTTA AI Bootcamp — Deep Learning Basic

I revisited the lecture files and my saved notebook copies, then connected data handling, CNNs, RNNs, BERT, and the three projects.

Provider: OUTTA
Dates: –2024-08-31
Duration: —
Format: Under verification

What I studied

deep-learning
image-classification
nlp
audio-classification

How I reorganized the course

I reviewed all 16 PDFs in the archive: 530 pages covering preprocessing, Dataset and DataLoader, regression, gradient descent, neural networks, CNNs, language models, embeddings, RNNs, BERT, GPT, transfer learning, and project guidance. I grouped the material into data and regression, neural networks and CNNs, language representation, sequence models, and the three projects.

Saved notebook outputs

Fifteen saved notebook copies retained 48 figures. Some contain my code changes, but not every copy differs from its starter version. I did not claim a fresh end-to-end run of every cell; the plots below are outputs saved in the notebooks.

Ten sigmoid basis functions from the Universal Approximator notebook — Changing weights and biases shifts the position and slope of each sigmoid basis.

Twenty-five CNN misclassifications with predictions and true labels — I inspected failure samples before changing preprocessing or augmentation.

RNN training accuracy rising while test accuracy stays near one half — The curve shows memorization without a matching gain on the test split.

P1, P2, and P3

I kept the original notebooks and built separate reruns. P1 compares HOG with an RBF-SVM, P2 groups duplicate descriptions before TF-IDF and LinearSVC training, and P3 combines MFCC, chroma, and spectral summaries with an RBF-SVM. The result pages explain the split and its limitations.

One old practice notebook contained a plaintext API key. I excluded it and changed the public example to read secrets from environment variables. Course PDFs and datasets are not redistributed.