Tensor.Rocks 🔥

✅

Achieve the same prediction quality with 50% less data labeling costs

We are making an active learning platform, so you can label less data while achieving the same prediction quality.

What is Active Learning?

Active learning is a way to reduce data labeling costs during the model training stage. To do so one should label more unusual, diverse data. We do this iteratively.

First, we need the pool of unlabelled data. It can be 🖼️ images, 📑 texts, 🎵 audio, 🎬 video and practically any type of data.

Then, a seed dataset is created. We sample a small portion (up to 100 examples) of the most diverse data points and send them to labelers.

After that, the active learning begins:

Model is trained on labeled data.
⬇️

A trained model is used to score the pool of unlabelled samples to find a batch of the most informative ones.
⬇️

The batch is sent to labelers.
⬇️

A labeled batch is added to the labeled data pool.
🔃

Repeat from step 1 until the desired quality is achieved.

In practice, it allows us to label 50% of data to achieve 99.9% of the SOTA quality.

Features

Integration with popular data labeling services
- Amazon MechanicalTurk
- Yandex.Toloka

All the SOTA models - available on the platform, including:
- 🤗 Transformers
- 🔥 PyTorch Image Models

Interface to bring your own model into the active learning loop.

Our own labeling interface, so your people can do the labeling.

On-premise solution for most valuable datasets.

Follow the development