Similarity Learning lacks a framework. So we built one | Hacker News Reader

Similarity Learning lacks a framework. So we built one

blog.qdrant.tech

59 pointsgenerall24 comments

Not a full featured framework, but pytorch-metric-learning has data loaders, lossess, etc. to facilitate similarity learning: https://github.com/KevinMusgrave/pytorch-metric-learning

Disclaimer: I've made some contributions to it.

Seconding this. PML is high quality, active, and well documented.

PML is a great collection of implementations, but not the best framework. Also you can use PML with Quaterion: https://github.com/qdrant/quaterion/blob/master/examples/tra...

Found the wiki article more useful in describing what Similarity Learning and Metric Learning are: https://en.wikipedia.org/wiki/Similarity_learning

Great article. I've been working in and around this space since 2014, and I think similarity learning, vector search, and embedding management will be a core part of future applications that leverage ML.

I recently built a similarity search application that recommends new Pinterest users channels to follow based on liked images using Milvus (https://github.com/milvus-io/milvus) as a backend. Similarity learning is a huge part of it, and I'm glad more and more tools like Quaterion are being released to help make this kind of tech ubiquitous.

Is it somehow connected to Qdrant similarity search engine? Is there a default integration for it?

The idea is to fine-tune model, which can be used to produce embeddings required for Qdrant. Our design approach is to make things as modular as possible, so you can use framework and engine independently. But we are working on integrating components as well

I’m familiar with metric learning within the Mahalanobis family for kNN oriented applications . I’m not getting what use cases this framework targets? Is it custom image search type stuff which may benefit from fine tuning?

What is a realistic minimum viable dataset for an approach like this? When is it not advisable? How does it compare to other more basic approaches?

The main idea is to train a deep learning model to encode a high-dimensional sample to a low-dimensional vector in a latent space. Then it can be used in various downstream tasks such as KNN applications, semantic search, multimodal retrieval, recommendation systems, anomaly detection etc. It's not limited to the image domain --it can be also audios, texts, videos, or more specific entities such as authors, soccer players, songs etc. The size of the dataset can be thought of being similar to other deep learning methods, but you can make a choice among various similarity learning methods based on the size of your dataset or according to whether it's labeled or not. A common approach is (1) to train a base model by using a self-supervised method with a bulk amount of unlabeled data and (2) to finetune it on a more specific domain or task with a smaller labeled dataset. If you can start with a pretrained model such as ResNet or BERT, you can skip the first step.

I’d be surprised if it’d be useful for something like cars or soccer players, or really anything that may not have a continuous mapping. I guess more generally whenever the underlying “true” similarity function is not differentiable — categorical data springs to mind (cars, football players…).

I could see it making sense for complex unstructured data — Qdrant seems to point in that direction.

Some loss functions such as ArcFace loss and CosFace loss enforce the encoder model to organize their latent space in such a way that categories are placed with an angular margin from one another. Thus the model implicitly learns a continuous distance function.

Fun fact, one of the examples in Quaterion is for similar cars search.

If you find this topic and want to discover more, we collected a bunch of resources that might be helpful. https://github.com/qdrant/awesome-metric-learning

To answer my own question:

https://qdrant.tech/

Very cool. Can you comment on how this compares with tensorflow similarity? https://blog.tensorflow.org/2021/09/introducing-tensorflow-s...

I'm one of the authors of Quaterion and also a contributor of TF Similarity. First of all, Quaterion is based on PyTorch. From a more technical perspective, TF Similarity is currently stronger on the self-supervised methods for images and lacks some more generic losses while Quaterion puts a bigger emphasis on finetuning models from any domain efficiently for practical purposes. For example, Quaterion has a intermediate caching mechanism that accelerates finetuning considerably. Overall, They both do a good job for their own use cases.

I realise this is an overly-broad question, but any insight into what's the state-of-art in Similarity Learning for article-type text?

More specifically, I'm interested in deriving distances between writing style, arguing style, etc.

It is definitely possible to do, if you have a proper training set. You would need to somehow give model a signal, that you are interesting in e.g. arguing style specifically, and not the topic of the text.

There's study here: http://cs230.stanford.edu/projects_spring_2021/reports/57.pd...

Basically, you can collect text from different authors, then you can use authors names as labels to train a similarity learning with it. My suggestion would be finetune a Transformer model with a specific head and an ArcFace loss.

There is one https://github.com/jina-ai/finetuner pretty well-designed and also gives SOTA performance from its docs

Unfortunately, it's no longer open source, but requires using their cloud.

"From 0.5.0, Finetuner computing is hosted on Jina Cloud. THe last local version is 0.4.1, one can install it via pip or check out git tags/releases here."

but does being "opensource" naturally make a software good in quality & performance? In the end, people try to solve a problem right?

anyway, congrats on the release :P and I upvoted! so we are good.

Of course, solving a problem is the most important thing at the end of the day. However, there are some data privacy constraints you may sometimes need to fulfil, and sending the data to an external cloud, managed by 3rd party is not an option.

Moreover, fine-tuning might be just one of the applications of neural networks in the organization, and you may already have some pipelines built to train them, so it should be also unified.

And more importantly, Jina's finetuner gives you some pretrained models to choose from, while Quaterion is PyTorch Lightning based, so you can easily integrate it if you already use PyTorch and have the flexibility to fine-tune any custom network as well.

Starting from 0.5.0 finetuner is no longer an open-source.

> From 0.5.0, Finetuner computing is hosted on Jina Cloud. The last local version is 0.4.1, one can install it via pip or check out git tags/releases here.

But there are some cool ideas implemented there as well, I encourage you to try both!

The title is written in a clickbait format. So I had to point it out.

Reply on news.ycombinator.com