Deep Learning Experiments: the slides are on Drive, not in the repository

The repository holds a table, not the material

The top level is four entries: a gitignore, a licence, a readme, and one requirements file, plus a versions directory. Everything a student needs is somewhere else. Twenty-one of the twenty-two slide decks in the theory table are Google Drive share links, and so are all the teaching guides. Exactly one slides link points into the repository itself, the state-space-model paper inside the Mamba row, which lives under a year directory as a PDF. So the repository versions the code and indexes the notes, while the notes themselves sit in someone's cloud drive. A fork does not carry them, a pull request cannot review them, and a drive link can be made private without anyone noticing.

The heading says 2026 and the newest code directory says 2025

The readme's second heading is a version label with two words under it, revised and expanded. The code is partitioned by year, and the years are readable straight out of the notebook paths. The multilayer perceptron, convolutional, recurrent and transformer demos are all under 2023. Mamba, segmentation and the language-model work are under 2024. Normalizing flows and diffusion are under 2025. The newest directory is therefore one year behind the heading, and the repository was pushed in September 2026, so the label is current while the most recent material is not. There is no changelog in the tree and no history beyond the heading itself, so there is no record of what the revision changed.

Detection is pinned to a compute stack the file contradicts

The requirements file lists fifteen packages with no version on any of them, starting with the compute stack. Underneath, a commented block installs a system audio library and then walks through building a detection framework from a git clone, naming a specific version of it in a comment. The commented wheel command points at an index built for CUDA 11.1 with torch 1.8.0. That is the opposite end of the range from whatever the unpinned torch on the first line resolves to today, and the two cannot coexist in one environment. So the detection and segmentation tracks do not install alongside the rest of the file, and the only version information anywhere in it describes a stack from several years back.

Eleven notebooks are named after the oldest dataset in the field

The theory table covers the current curriculum, ending in normalising flows, flow matching, diffusion and language models. Look at the executable filenames rather than the topic names and the picture changes. The perceptron, convolutional, recurrent and transformer demos, both Mamba demos, the flow demo, the variational autoencoder, its conditional variant, the unconditional GAN and the conditional GAN are all files whose names end in a handwritten-digit recognition dataset. Two more, the autoencoder and the diffusion one, do not. So the architecture coverage is modern and the evidence for it is one small grayscale benchmark used everywhere. That is a defensible teaching choice and it is also what makes the notebooks run quickly; it is just not visible from the topic column.

Eight of twenty-two topics have teaching notes

The theory table has four columns and the middle two are slides and a teaching guide. Reading across the rows, the teaching guide is filled in eight times: supervised learning, datasets and dataloaders, the perceptron, the convolutional network, the recurrent network, the transformer, the introduction to language models, and agents. The other fourteen have slides and nothing else. The gap lands awkwardly on the two newest topics. Mamba has two notebooks and a slides deck but no teaching guide, and the agents row has both a guide and a deck but its code link leaves the organisation entirely, pointing at a repository under a different account. So the most recent architecture is undocumented in prose and the most recent topic is documented elsewhere.

The language-model section is the only one that separates training from fine-tuning

Most rows hand you a notebook and stop. The language-model rows do something more careful. They ship a script and a notebook side by side, with the script named for training from scratch and the notebook for validation, then a second row repeating the pair with the names swapped to a fine-tuned variant. So a student can see the difference between a model trained from scratch and one adapted, with separate entry points for each, and the validation file is a notebook while the training file is a plain script. That row is also the only one that links a results document, a markdown file in the repository holding training and validation results.

One release tag, named after something that is not here

The repository has a single release. It is tagged models, with the human-readable version in the tag body reading as a zero-point-zero-one build, and it was published in April 2022. That is four years before the last push, and the name refers to artefacts the repository does not appear to contain, since the visible tree holds notebooks and documentation rather than trained weights. There is no build script, no training entry point at the root and no model artefact path in the readme, so a reader arriving through the releases page has no way to tell what that tag was pointing at or whether it still exists.

The description promises videos and the video column is empty

The repository describes itself as videos, notes and experiments to understand deep learning. The notes and the experiments are there. The videos are not: the practice table has a video column and every visible cell in it is a dash, from the development-environment row through Python, the array library, the two einsum notations and beyond. The theory table has no video column at all. So one of the three things the description claims is delivered by a column that is uniformly blank. The remainder of the practice table is a sensible progression, and giving einsum and einops their own separate rows is a deliberate choice worth copying.

Editorial conclusion

This repository suits a lecturer assembling a syllabus from parts, because the topic list is unusually current for a curriculum and each row separates slides, teaching notes and code. It does not suit anyone who needs the material to be self-contained, because the teaching content lives in shared Drive links and the code that works with it is versioned by year under a different label. Check three things before building a course on it. Check that every Drive link still resolves, since a revoked share breaks a row with no error. Check which year directory each notebook is actually in, because the heading is not the answer. And check the detection and segmentation requirements separately, because their install instructions are pinned to a compute stack the rest of the file contradicts.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. README
4. Releases
5. roatienza/Deep-Learning-Experiments on GitHub