Launch HN: Lightly (YC S21): Label only the data which improves your ML model
If you wonder what data labeling looks like for images then think about these captchas that want you to tag images in the web containing objects such as a bus or person. When we were working on training machine learning (ML) models from scratch, we often had to do this labeling ourselves. But there was always far too much data for us to be able to label all of it. We talked with more than 250 ML teams ranging from small groups of 2-3 people to large teams at Apple and Google, and they all face the same problem: they have too much data to label.
Not only that, but there wouldn’t be a lot of value in labeling everything. For example, if you have billions of images, it's a waste of time to get humans to label every one of them, because most of those labels wouldn't add useful information to the model you’re hoping to train. Most of the images are probably similar enough to other images that have already been labeled and they have nothing new to tell your model. Spending more labeling effort on those would be a bit like labeling the same image over and over again—quite wasteful.
As soon as your ML model surpasses the initial prototype stage, you’re most interested in the edge cases in your dataset — the ones that represent rare events. For example, a few days ago, there was a Twitter thread about failure cases for Tesla vehicles. One Tesla has mistaken a yellow moon for a yellow traffic light: https://twitter.com/JordanTeslaTech/status/14184133078625853.... Another edge case is a truck full of traffic lights: https://twitter.com/haltakov/status/1400797882891091970. Finding and labeling such rare cases is key to having a robust system that will work in difficult situations.
Rather than labeling everything, a better approach is to first discard all the redundant images and keep only the ones that it's worth spending time/money to label. Let's call those "interesting" images. If you could spend labeling effort only on the "interesting" images, you'd get the same value for a fraction of the cost.
Many ML companies in a more advanced stage have had to tackle this problem. One approach is to pay people to go through the images and discard the "boring" (nothing-new-to-tell-me) images, leaving the "interesting" (worth-spending-resources-to-label) ones. That can save you money if it's on average cheaper to answer the question "boring or interesting?" about an image than it is to label it. This solution scales as long as you have an increasing human labeling workforce every year. However, ML data doubles every year on average, and therefore the labeling capacity would need to double too.
Much better than that — the holy grail — would be for a computer to do the work of discarding the "boring" images. Compared to paying humans to do it, you'd get the "interesting" subset of your billion images almost for free. You would have much less work to do (or money to spend) on labeling, and you'd get just as good a model after training. You could split the savings with whoever knew how to make a computer do this for you, and you'd both come out ahead. That’s basically our intention with Lightly.
My co-founder Matt and I worked on many machine learning projects ourselves, where we also had to manage tooling and annotation budgets. Dealing with data in a production environment is different from academia. In academia, we have well-balanced and manually curated datasets. It is, as some of you know, a huge pain. The solution of the problem boils down to working with unlabeled data.
Luckily, in recent years, a new subfield of deep learning has emerged called self-supervised learning. It’s a technique to train models to understand data without any labels. In natural language processing (NLP), modern models like BERT or GPT all rely on it. In computer vision, we have had a similar breakthrough in the last year with models such as SimCLR or MoCo. Back in 2020, we started experimenting with self-supervised learning to better understand unlabeled data and improve our software. However, there was no easy-to-use framework available to work with the latest models. To solve that problem, we built our own framework to make the power of self-supervised learning easily accessible. Since we want to foster research in this domain and grow a bigger community around this topic we decided to open-source the framework in fall 2020 (https://github.com/lightly-ai/lightly). It is now used by universities and research labs all over the world. We realized that the ability to understand and visualize unlabeled data is also valuable to other ML teams and decided to offer our solution as a SaaS platform. The platform builds on the open-source framework and helps you work with the most valuable data.
Here are some examples where Lightly can help you:
Analyze the quality and diversity of your datasets. Our platform can also use metadata or labels if available. Uncover class distributions, dataset gaps, and representation biases before labeling to save time and money. You can do it manually or automated through our data selection algorithms which ensure the most diverse subset of your dataset is chosen. (https://docs.lightly.ai/).
Once you have a labeled dataset and trained your model, our active-learning algorithms allow you to gradually select the next data to be added to your training set. Only label the best data for model training until you reach your target accuracy. https://docs.lightly.ai/getting_started/active_learning.html
Check it out yourself with this quick demo video https://www.youtube.com/watch?v=38kwv0xEIz4
Lightly integrates with an API directly into your pipeline, is available on-prem, and can process up to 100M samples within hours.
We're excited we get to show Lightly to you all. Thank you for reading! Please let us know your thoughts and questions in the comments.
https://github.com/capitalone/DataProfiler
In this space, Prodigy really dominates:
https://prodi.gy/
We actually built our own internal system which integrates and can export the labels (does predictive labeling, etc). Of course, we only focused on text data at the moment.
All that being said, this is going to become a crowded space and highly competitive. Plus, once the data is labeled companies often drop their labelers. I would recommend ensuring some consistent use. Potentially, hosting their models off-prem or something to lock companies in.
Prodi.gy is great but focuses heavily on NLP and speeding up the labeling process. Our goal is really to help you reduce what you label before you use any labeling tool.
We are working with labeling tools as well as providers to streamline the workflow.
Haven't used Prodi.gy so don't know how its active learning algo works. May you share the difference?
However, with Lightly you can additionally make sure you only select images that are visually different from each other. And you always get visual feedback on the selected data in our web platform. The additional control and feedback mechanism allow you to have a more focused workflow.
Can using Lightly introduce additional bias in the model, since only a select few of inputs are being labeled? This may be a concern for publicity purposes.
By the way, I thought ETH spinoff requirements were incompatible with YC requirements - nice to see it can be made to work.
>How do models trained with Lightly compare with other approaches wrt adversarial robustness?
We have no benchmark available. Both approaches can be combined. You can use Lightly to pick a diverse subset, label it and then during training/ evaluating the model check for adversarial robustness and re-iterate.
>Can using Lightly introduce additional bias in the model, since only a select few of inputs are being labeled? This may be a concern for publicity purposes.
If we remove bias we automatically introduce bias. BUT we want the introduced bias to be controlled and known.
Bias typically comes from the way we collect data. For example, more data is being collected during the day than during night for autonomous driving. We also have more data collected during sunny weather than rain or snow. We also have more data from cities like San Fransisco than cities like New Mexico. Most of our datasets are biased.
> By the way, I thought ETH spinoff requirements were incompatible with YC requirements - nice to see it can be made to work.
From what we know we are the first ETH spin-off who is part of the YC program. we hope they don't abandon us.
Are there any libraries to facilitate a workflow like this?
For example: A simple workflow for a classification task would be to train a self-supervised model on the whole dataset and find samples with a different annotation than their nearest neighbors. These can be identified quickly either in a colored scatter plot or by simply measuring disagreement.
https://github.com/modAL-python/modAL
It is built on scikit-learn which is also evident from the suggested workflow. Lightly on the other hand was specifically built for deep learning applications supporting active learning for classification but also object detection and semantic segmentation.
modAL provides uncertainty-based active learning. However, it has been shown that uncertainty-based AL fails at batch-wise AL for vision datasets and CNNs, see https://arxiv.org/abs/1708.00489. Furthermore it only works with an initially trained model and thus labeled dataset. Lightly offers self-supervised learning to learn high dimensional embeddings through its open-source package https://github.com/lightly-ai/lightly. They can be used through our API to choose a diverse subset. Optionally, this sampling can be combined with uncertainty-based AL.
But how could the real-world accuracy be computed? Is a separate dataset needed for that purpose?
However, in practice we see that it is very hard to collect a good dataset. There is a great twitter thread from Abubakar(CEO Gradio) about this topic: https://twitter.com/abidlabs/status/1423067498862219267
What we ideally want, is that the data we collect for training, validating, and testing our model represents the environment the trained model will operate in. However, collecting this data can be super difficult. E.g. for autonomous driving systems that would mean collecting data from every corner of the world, at every time of the day, during all kinds of weather conditions, in all 4 seasons...
However, this is a really difficult task. You will very likely end up having more data collected from one city (e.g. because your fleet is much bigger in city A than city B)