ello world

Hey! First post is here, published using Astro.

I will use this site as a devlog on my journey getting deeper into AI/ML.

My first interaction with machine learning was sklearn, which still has a fantastic and well designed API. I used it primarily for classification on tabular data. Think of a large spreadsheet with many columns, and you need to predict a column value based on the rest.

I opened the door to NLP in around 2017, when I discovered Word2vec. Never managed to get off the ground until I discovered SpaCy where I managed to create some meaningful classification and named entity recognition. Still requried lots of labelling!

Then came BERT and Hugging Face 🤗 was only a neural coreference system, but soon released Write with Transformers and the rest is history!

Computer Vision was often more effective on some tasks. For example classification of a dense document as being a certain type of form (airworthiness approval forms) was more efficient with better accuracy as an image classification problem using AWS Rekognition at the time.

Applied similar technique to find page orientation problems in scanned documents:

Object detection was also useful with document images:

Today, large language models and moreso, Vision Language Models (VLMs) can do all of this with few-shot prompting (if not zero-shot). For efficiency, use it to label examples and distil a smaller model, optimize it with quantization and ONNX runtime and you have a beast!