Offbeat Datasets

Here is a collection of datasets which I found to be a bit offbeat. It might be that the datasets may be well-known in the domains, but just that the curiosity factor impels me to collect this information.

Link to the file

Annoying terms in post modern machine learning literature

 A collection of extremely stupid, artificial(sic) and pretentious terms which have sprouted around the "AI" hype. One day, if i get time, i will attempt to write parody definitions for the terminology; as of now i merely list them below.

Ingest, Consume, Foundation Model(God only know what is "foundational" about it), Artificial General Intelligence(Really?! General intelligence?), "practitioners", "in production"(what does it produce is what i wonder), Foundry (all i can think of such foundries is producing gas), "digital twin", "stand up". "training"(you can train a monkey or a dog; you can even train a human to imitate others like apes but how can you "train" a mathematical model? one can optimise its parameters.)

Irritating fake pretend words, mostly encountered in LinkedIn:

"chuffed"(another annoying term appearing a lot in LinkedIn), "reach out", "update"(one person asking for job recommendations is then asking for an "update").

A very readable article on the Register on the AI/AGI bull

https://www.theregister.com/2024/05/18/distros_ai_code/

A simple tool for searching PDFs on your local storage

For many years i have been facing this simple problem. I download a lot of PDFs; usually they are papers, electronic academic books in the PDF format and the life. There may have been a few non academic content, such as fiction. Earlier I used to categorise them while saving them on the local storage. As an aside, this local storage is the hard disk but this can also be a network drive which is reachable by the local file commands (e.g. ls or file explorers). A lot of the time, it was or still is possible for me to recall which directory a particular paper had been saved into. As the number of papers or files keeps growing, it becomes difficult to track or remember. Therefore, it was on my mind for many months to write a small software which I could use to search the contents of the PDFs; hopefully quickly and efficiently. I did not want an exact and accurate search; if at all this was possible.

The repository here is the fruition of such efforts. This is a pure Python package. I have written an extensive README. Feel free to use it. For both positive, negative suggestions and suggestions for improvement, feel free to contact me at the e-mail ID mentioned in the README. 

Of course, this can be extended to search DOC files to; drop me a line if you want that functionality.

Stanford AI blog link on weekly supervised link

http://ai.stanford.edu/blog/weak-supervision/

 good read.

A very good read on supervised contrastive learning

Reference:

https://arxiv.org/abs/2004.11362

Supervised Contrastive Learning

Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, Dilip Krishnan

Contrastive learning applied to self-supervised representation learning has seen a resurgence in recent years, leading to state of the art performance in the unsupervised training of deep image models. Modern batch contrastive approaches subsume or significantly outperform traditional contrastive losses such as triplet, max-margin and the N-pairs loss. In this work, we extend the self-supervised batch contrastive approach to the fully-supervised setting, allowing us to effectively leverage label information. Clusters of points belonging to the same class are pulled together in embedding space, while simultaneously pushing apart clusters of samples from different classes. We analyze two possible versions of the supervised contrastive (SupCon) loss, identifying the best-performing formulation of the loss. On ResNet-200, we achieve top-1 accuracy of 81.4% on the ImageNet dataset, which is 0.8% above the best number reported for this architecture. We show consistent outperformance over cross-entropy on other datasets and two ResNet variants. The loss shows benefits for robustness to natural corruptions and is more stable to hyperparameter settings such as optimizers and data augmentations. Our loss function is simple to implement, and reference TensorFlow code is released at this https URL.

Confidence Intervals references

Here are couple of references for CI. Without getting too technical on the computations, these provide the motivation, "how to" and more importantly, the interpretation of the CI.

Part 1: https://pubmed.ncbi.nlm.nih.gov/9287892/

Part 2: https://pubmed.ncbi.nlm.nih.gov/9287893/

The key lines for me:

"The theoretical basis for the calculation of a CI includes the assumption that (he study can be repeated many times. Each time, different results would be obtained through the selection of a slightly different sample of patients from the population (sampling variability). Each trial would therefore also produce a different 95% CI. If the trial were performed 100 times, then, on average, 95 of the 95% CIs calculated would contain the true value (and 5 would not). In practice, however, we usually perform a study only once. Once we actually perform a trial, and calculate a single 95% Cl, the true value either lies within this confidence interval or it does not. Therefore, in referring to the particular results from a single study, it is not correct to state that there is a 95% chance or probability that the true value lies within the CI. It is correct to say that if the true value lies outside the 95% CI, the likelihood of obtaining the data observed in the study is 5% or less."

Emphasis is mine.

Lecture notes in Generalized linear models: Rodriguez

Link: https://grodri.github.io/glms/notes/

docker startup command

VOLS="-v /etc/passwd:/etc/passwd:ro -v /etc/group:/etc/group:ro " # add any other mounts you want

echo Exported paths $VOLS

docker run -it --user $(id -u):$(id -g) --rm --gpus all --ipc=host --net=host $VOLS projectmonai/monai:latest bash

Interpretable Machine Learning: Fundamental Principles and 10 Grand Challenges

Interesting title and abstract. It is a long read of 80 pages.

https://arxiv.org/abs/2103.11251