📝 Guest post: It's Time to Use Semi-Supervised Learning for Your CV models*
Was this email forwarded to you? Sign up here In this article, Masterful AI’s team suggests that instead of throwing more training data at a deep learning model, one should consider semi-supervised learning (SSL) to unlock the information in unlabeled data. IntroPreviously, we showed that throwing more training data at a deep learning model has rapidly diminishing returns. If doubling your labeling budget won’t move the needle, what to do next? Try SSLSemi-supervised learning (SSL) means learning from both labeled and unlabeled data. First, make sure you are getting the most out of your labeled data. Try a bigger model architecture and tune your regularization hyperparameters. (plug alert: executing on these two steps is hard, and the Masterful platform can help you). Once you have a big enough model architecture and optimal regularization hyperparameters, the limiting factor is now information. An even bigger, more regularized model won’t deliver better results until you train with more information. No more labeling budget but need more information? But wait – it seems like we are stuck between a rock and a hard place. There’s no more labeling budget and yet the model needs information. How do we resolve this? The key insight: labeling is not your only source of information... unlabeled data also has information! Semi-supervised learning is the key to unlocking the information in unlabeled data. SSL is great because there is usually a lot more unlabeled data than labeled, especially once you deploy into production. Avoiding labeling also means avoiding the time, cost, and effort of labeling. Is SSL Good Enough?SSL has been an academic topic for decades. But until about 18 months ago, it did not outperform traditional techniques for CV on standard benchmarks. All that changed with a series of papers published in 2020 and 2021, including Unsupervised Data Augmentation, Noisy Student Training, SimCLR, and Barlow Twins. Today, SSL techniques are responsible for training the most accurate convolutional neural networks (convnets). And transformers are also primarily trained using SSL techniques, in case someday transformers replace convnets as the workhorse CV architecture. How SSL WorksAn Algorithmic View There are a lot of SSL algorithms, but most of the recent approaches loosely share these attributes:
The special problem can be making the pair of outputs consistent with each other. Or using the pair of outputs to solve a pretextual problem, like contrasting between pairs of images, that either do come from the same source image or don't. Sometimes the final output of the model is used, and sometimes a feature embedding. Some algorithms place additional layers between the features and the loss function, while others feed the outputs to the loss function directly. And different techniques work better for low-shot data vs high cardinalities. Most techniques require two training phases, and sometimes the weights of the two models are shared while in other approaches, one model slowly receives weights from the other. Here are a few great walkthroughs by Spyros Gidaris of Valeo.ai and Thang Luong of Google. An Intuitive View An intuitive view of how these algorithms work focuses on clustering the feature embeddings. In concrete terms, the feature embedding is often the output of the penultimate layer of a convnet before the final linear/dense/logistic layer. For algorithms that directly train consistency, differently noised views of the same image must generate similar embeddings. If the feature embeddings are clustered together in the high-dimensional feature embedding space, the feature extractor has learned a useful representation of the data. If the noising function is able to move one image into the feature embedding space of another image, then it's also true that two different images now generate similar feature embeddings. This suggests that the noising function's goal isn't to be confusing, but rather, to transform a single image enough to collide with the feature embedding of other images in the same class, but not so far as to push it to collide with the feature embeddings of images from different classes. Indeed, when projecting the feature embeddings of one SSL algorithm, we see well-clustered feature embeddings. One place to start: research reposIf you want to try these approaches, VISSL from Meta AI and Tensorflow Similarity are two solid repos to start with for PyTorch and Tensorflow respectively. We've worked with both and they are awesome! But like any research repo, they are focused on experimentation, not production. You really have to understand the papers behind them to understand the code, they may not be robust on production datasets, and many hyperparameters will require manual guessing and checking. If you are looking for a productized implementation, consider Masterful. Three ways to access SSL via the Masterful platformThe Masterful platform for training CV models offers three ways to train with SSL.
Good luck on your journey with SSL! Join Masterful AI slack anytime you want to talk SSL! And to try Masterful, just run pip install masterful to install our product and try it out. *This post was written by Yaoshiang Ho, co-founder and head of product at Masterful AI, and originally posted here. We thank Masterful AI for their ongoing support of TheSequence.You’re on the free list for TheSequence Scope and TheSequence Chat. For the full experience, become a paying subscriber to TheSequence Edge. Trusted by thousands of subscribers from the leading AI labs and universities. |
Older messages
🔬 Edge#190: Continuous Model Observability With Superwise
Thursday, May 12, 2022
Introducing to you the platforms that deal with the ML challenges
🚰 Edge#189: What is Pipeline Parallelism?
Tuesday, May 10, 2022
In this issue: we discuss pipeline parallelism; we explore PipeDream, an important Microsoft Research initiative to scale deep learning architectures; we overview BigDL, Intel's open-source library
👄 A New Open Source Massive Language Model
Sunday, May 8, 2022
Weekly news digest curated by the industry insiders
📝 Guest post: Active Learning 101: A Complete Guide to Higher Quality Data* (part 2)
Friday, May 6, 2022
In this article, Superb AI's team explains the benefits of building an active learning flow for your computer vision project
🧙🏻♂️ Edge#188: Inside Merlin, the Platform Powering Machine Learning at Shopify
Thursday, May 5, 2022
The eCommerce giant published some details about the platform powering its ML workflows
You Might Also Like
ASP.NET Core News - 05/10/2024
Friday, May 10, 2024
View this email in your browser Get ready for this weeks best blog posts about ASP.NET Core! This newsletter is sponsored by elmah.io - the most advanced, yet so simple to set up, error logging and
Chrome Zero-Day Alert — Update Your Browser to Patch New Vulnerability
Friday, May 10, 2024
THN Daily Updates Newsletter cover PowerShell 7 Workshop ($35.99 Value) FREE for a Limited Time Discover the capabilities of PowerShell 7 for your everyday tasks with this carefully paced tutorial that
Post from Syncfusion Blogs on 05/10/2024
Friday, May 10, 2024
New blogs from Syncfusion Performance Optimization in React Pivot Table with Data Compression By Manikandan Murugesan This blog explains the performance optimization in React Pivot Table using Data
Hacker Newsletter #699
Friday, May 10, 2024
If you're thinking without writing, you only think you're thinking. //Leslie Lamport hackernewsletter Issue #699 // 2024-05-10 // View in your browser #Favorites Turbocharging Low-Code
OpenAI publisher pitch leaks 🤖, Xbox mobile game store 🎮 , dangerous protocols 🌎
Friday, May 10, 2024
OpenAI is pitching partnership opportunities to news publishers through its Preferred Publishers Program. It has been courting publishers since July Sign Up |Advertise|View Online TLDR Together With
Master iOS 17.5
Friday, May 10, 2024
Did you know that Apple has released 12 updates to iOS 17 since its release last fall? Apple has added a ton of major new features, including updates to Messages, Photos, and a brand new built-in app!
Press, Pause 🗜️
Friday, May 10, 2024
Apple doesn't understand why people like hydraulic presses. Here's a version for your browser. Hunting for the end of the long tail • May 09, 2024 Press, Pause Beyond misunderstanding its iPad
Data Science Weekly - Issue 546
Friday, May 10, 2024
Curated news, articles and jobs related to Data Science, AI, & Machine Learning ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏
Dell’s data breach
Thursday, May 9, 2024
Plus: Mistral AI is raising funds and Bumble's new strategy View this email online in your browser By Christine Hall Thursday, May 9, 2024 Welcome back to TechCrunch PM. Today I have for you a
💻 Issue 416 - The new disposable APIs in Javascript
Thursday, May 9, 2024
This week's Awesome JavaScript Weekly Read this email on the Web The Awesome JavaScript Weekly Issue » 416 Release Date May 09, 2024 Your weekly report of the most popular JavaScript news, articles