Maker

📌

Hi Product Hunt! Let me share a word from my co-founder Isaac: For the past year we’ve been working to get rid of the last barrier to entry for building custom computer vision algorithms: training data. Today, we’re showing the world our progress! Three years ago I joined an autonomous robotics company as an ML engineer. I was shocked to discover that at most companies the hardest part of ML isn’t designing neural networks; turns out most people just use the open source model that scores the highest on relevant benchmarks. The hardest part was gathering enough data to train that network to work for our use case! I’m really passionate about the math of ML, so I looked into using the knowledge stored in foundation models to directly construct a model with the behavior I had in mind instead of gathering data and training one from scratch. CLIP (the model powering Stable Diffusion and DALL-E) has a lot of knowledge about the world and people use it to build zero shot image classifiers with just text. But zero shot methods don’t give you much control over the classifier. I built a method that could take a list of positive and negative examples and use CLIP to build a classifier that satisfied those constraints, and DirectAI was born! Decades ago, a company having a database was a big deal, but now you can create one by just specifying a schema! We’re working really hard to bring that level of control and ease of use to computer vision. We have a long way to go and we’re excited for the journey. Looking forward to hearing your thoughts/feedback and growing with community :) thank you!

Report

3yr ago

Chestr 🛍

This is epic, great job Ben and Isaac!!! Can't wait to see the range of companies that will make sure of this epic technology. 🚀

Report

3yr ago

this is awesome!!

Report

3yr ago

Very excited about this!

Report

3yr ago

Computer Vision with DirectAI

Object Detection without Training Data

Object Detection without Training Data