Andy Bosyi: Few-shot learning as a trade-off between software development and data science
0 likes21 views
The document details the evolution and applications of few-shot learning techniques in AI, highlighting milestones from the internet in 1995 to current methods such as transfer learning and fine-tuning using various models like LLMs and diffusion. It covers diverse implementations including document classification, image detection, anomaly detection, and recommendation systems, alongside developments in zero-shot and one-shot learning. The content emphasizes practical applications in data-driven industries, such as document classification, price prediction, and IoT command centers.
Andy Bosyi: Few-shot learning as a trade-off between software development and data science
- 1. mindcraft.ai Few Shots Learning History: 1995 - Internet 2000 - Software 2005 - Web 2010 - Machine Learning 2016 - Deep Learning 2018 - Transformers 2020 - Large Models 2022 - Transfer Learning sudeep.co
- 2. mindcraft.ai Transfer Learning in LLM - what caused impact on few shot learning - nocode - text instruction - text embeddings - data translation - federated learning medium.com
- 3. mindcraft.ai Fine Tuning - image classification - object detection - semantic, identity segmentation - style transfer - text classification, NER coco dataset
- 4. mindcraft.ai Dataset Generation - generate or augment problem-specific data using LLM, diffusion etc - select a model from HuggingFace - fine tune it with the dataset
- 5. mindcraft.ai Zero Shot Learning - it is not unsupervised - train on some classes and then predict on a new one - multimodal (image and text embeddings) - human language prompt in a Large Model
- 6. mindcraft.ai One Shot Learning - template matching - clustering and finding closest centroid - triple loss and face detection - human language prompt + example medium.com/@crimy
- 7. mindcraft.ai Few Shot Learning - recommendation systems - prototypical networks - chat models - prompt + examples - LLM fine tuning towardsdatascience.com
- 8. mindcraft.ai Reinforcement Learning - translate images into language tokens - using autoregressive Transformer to learn world - 2 hours playing games to train - outperformed human in 10 of 26 games ICLR 2023, Transformers are Sample-Efficient World Models
- 9. mindcraft.ai Document Classification - text classification task in ~80 categories - multiple languages - used BERT and Ada embeddings - dataset augmentation with GPT3.5 - simple NN for the classification task - planning to add fine tuning
- 10. mindcraft.ai Anomaly detection - check fin declarations - zero shot learning approach - catches only obvious things - requires historical and snapshot clustering open.ai
- 11. mindcraft.ai Custom Assistant - replacing categorization bot with a human language one - collected dataset of ~2k Q&A - fine tuned open.ai Davinci - spent $300 on open.ai - such system can only gently suggest, not decide
- 12. mindcraft.ai NER - NEs are collected by pattern search (RegEx) - validated if possible - used GPT3 for context matching - alternatively using GPT3.5 for direct search open.ai
- 13. mindcraft.ai Document Scope - factorize document using few shot learning GPT3.5 - create template by using clustering on Ada embeddings - assign names with zero-shot learning on GPT3.5 - check document portions against template open.ai
- 14. mindcraft.ai Fixing UNSPSC - old rule-based system - collecting embeddings (BERT, Ada) - XGBoost for classification - 400+ classes, almost 80% accuracy
- 15. mindcraft.ai Address Normalization and Deduplication - few shot learning on GPT3.5 - similarity check with Levenshtein distance - moving to a model from HuggingFace - saved 2 months of work open.ai
- 16. mindcraft.ai Future of Our Job - Data preparation - Fine Tuning - Edge Models - Prompt Engineering promptbase.com
- 17. mindcraft.ai Das ist MindCraft Decision-making Engines for Data-driven Businesses, especially: - Document and Web pages Classification, Capturing (NLP, CNN, CV, NER) - Price Prediction (DNN, Regression, Prognosis) - Command Centers for IoT systems (RNN, Time Series, Anomaly Detection) - Computer Vision and Object Detection - Data Analysis and Generation