Name of Project: alsats: active learning (for a few) sats
Proposal in one sentence:
alsats is an open source project that aims to reduce the investment needed to create minimum viable datasets for supervised machine learning.
Description of the project and what problem is it solving:
alsats provides the ability to do inexpensive, intelligent training-as-a-service as mentioned in my post Idea Legos for Web3XAI. It combines Active Learning for guided labeling of training data for supervised machine learning with micropayments using the Lightning Network.
Labeling of datasets in supervised learning is often manual, unguided/indiscriminate and consequently expensive.
alsats implements Active Learning workflows via REST API endpoints for simultaneous, guided labeling of data and training of models. Only examples most in need of labeling (based on prediction uncertainty) get labeled and are used to train learners. This process of guided labeling reduces the number of labeled examples needed (and consequently, labeling costs) for comparable model predictive performance.
When coupled with the ability to pay in milisats for per-sample training on Lightning, alsats allows data scientists to focus on reducing their labeling investment with the following benefits
- No commitment/subscription/terms - label when you want, as much as you need
- Pay-per-label pricing - pay for as little as 1 train+label iteration in milisats.
- Full data security - alsats does not store training data.
- No need for user registration and user data-sharing - proof of Lightning payment is proof of authorized access.
Grant Deliverables:
Grant Deliverable 1 - alsats hosted on AWS EC2 server with publicly accessible API endpoints.
Grant Deliverable 2 - Streamlit/Gradio/equivalent app that allows invited data scientists to train + label with alsats.
Squad: antaraxia.eth/@antaraxia_kk