Active Learning. The Secret of Training Models Without Labels.

A large part of the success of supervised machine learning systems is the existence of large quantities of labeled data. Unfortunately, in many cases, creating these labels is difficult, expensive, and time-consuming.
An obvious solution is to use machine learning to aid in the creation of the labels, but this presents a chicken and egg problem: how do we build a model to create labels before labeling our data to train that model?
Active Learning is one solution. A semi-supervised learning technique to build better-performing machine learning models using fewer training labels.
Paper mentioned in the video:
Active Learning Literature Survey. burrsettles.com/pub/settles.a...
🔔 Subscribe for more stories: www.youtube.com/@underfitted?...
📚 My 3 favorite Machine Learning books:
• Deep Learning With Python, Second Edition - amzn.to/3xA3bVI
• Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow - amzn.to/3BOX3LP
• Machine Learning with PyTorch and Scikit-Learn - amzn.to/3f7dAC8
Twitter: / svpino
Disclaimer: Some of the links included in this description are affiliate links where I'll earn a small commission if you purchase something. There's no cost to you.

[@tecbrain]

Fantástico vídeo. La verdad es que ahora voy a trabajar el código para entenderlo. Gracias por el trabajo que haces para ayudarnos.

[@hasanx8317]

Duplicated records in the data has a significant meaning. It means that this repeatedly appearing record in the past is probably going to repeatedly appear in the future, it a VIP records, and knowing how to handle it well means you succeeded in high percentage of your supposed to do. So having duplicate data should some how eventually make the model very accurate in predicting it's related lable, more accurate than unique records.

[@thecouchman2112]

Really helpful video, thanks. One small thing though, the sound effects on the title screens were a bit loud imo :)

[@underfitted]

Noted! Thanks for the feedback!

[@underfitted]

GOOD ONE!

[@emeebritto]

yaa... >.

[@miguelduqueb7065]

Nice video! You can also use a similar approach to compare models and stay with the one that performs best. Here is how: A few years ago I was collecting data in the chemistry lab in order to fit some models. Each experiment took 1 day to complete, so I started with a simple factorial design, fitted all models to the initial data set, and then predicted the point of maximum divergence between all models. That point was used as the next experiment and models we refitted thereafter. This procedure was repeated several times. Computing uncertainty in your predictions is similar, but only with one model.

[@underfitted]

Thanks for sharing!

[@sahanakaweraniyagoda9866]

This is lit 🔥. Love this practical approach to Machine learning. Keep doing the amazing work 👏👏

[@underfitted]

Thanks! Much more coming!

[@fikriansyahadzaka6647]

Nice video! Could you also explain about semi-supervised learning? There are not many videos that clearly explain about the progress so far in semi-supervised learning, even though the topic become more popular nowadays

[@Param3021]

Another nice video! Learned a new concept - *Active Learning*

[@underfitted]

Glad to hear that!

[@maheshBasavaraju]

Loved the Idea of smart labelling. very cool

[@knutjagersberg381]

Love it, world class content! Also agree. A thought: Why not start with few shot or zero shot learning before active learning?

[@underfitted]

If you have a model capable of zero-shot, absolutely!

[@jubakala]

Thanks! This was exactly what I needed at the moment! (:

[@vidyachandran944]

Great content! Thank you :)

[@JoaquinRevello]

Excellent Video. This channel is going to be huge soon

[@jayantghadge4027]

This method to me seems a little bit like boosting. I might be wrong though, but boosting is what came to my mind after watching the video.

[@jainamshroff4998]

A Very good video!

[@mahendrakumargohil6384]

Excellent Information 👍👍

[@underfitted]

Glad it was helpful!

[@brunoras]

Super insightfull, I`m using this ideas right now!

[@123arskas]

If you've made it public (for smaller scale projects) please give the link to its repo. Thank you

[@underfitted]

Wonderful!

[@kemalariboga]

Great content!

[@underfitted]

Thanks!

[@roshanaryal7786]

Hi, Santiago! Love your content! Could you please make a video on how to start machine learning as a beginner with some programming experience. I've been doing web dev but want to transit into ML. I will appreciate your response 😊

[@underfitted]

It's coming soon!

[@erdi749]

I love your videos, nice and extremely informative! Just a quick comment: is it possible not to have those " bommmm!" soun?(: It make impossible to listen your videos in a car or with headphone. Thank you!

[@underfitted]

Thinks, Erdi! Yes, if you watch my last few videos, I’ve improved the audio, including removing that particular sound 😏

[@lorenzoleongutierrez7927]

Great explanation, thanks! Do you have some example of labeling services providing this approach?. greetings !

[@fobaogunkeye3551]

Lovely video Santiago! Quick question: How do we label the low confidence data that the model initially had a hard time predicting since we also didn't know what the label was in the first place. How do we know the label/class to use for that low confidence predicted data when we re-train ?

[@underfitted]

We will start by labeling some of the data manually. The goal is to seed the process to start generating automatic labels.

[@dimasveliz6745]

dynamic! Liked it more!

[@underfitted]

Cool, thanks!

[@juan.o.p.]

Very interesting

[@underfitted]

Glad you think so!

[@123arskas]

I've some queries. There's no proper practical application of it is it? Since the paper talks about methods proposed along with practical issues. Since your videos are straight to the point and you try to keep it simple, just wanna know if you've found practical implementation of it in Python etc. Do give a link to it in the description. Thank you

[@underfitted]

Yeah, I've personally used Active Learning multiple times. It's a very practical way to decide how to label a dataset.

[@Param3021]

1:03 - We need to Build a Model to Label the data we need, to Build a Model 🤯

[@underfitted]

Yup :)

[@sodipepaul9370]

Wow.

[@underfitted]

Wow indeed

[@CarlosBCU]

Hi, maybe a silly question but how you calculate the confidence after step 2?

[@underfitted]

Assuming you are using a classification model, for example, that will be the confidence (probability) returned by the model. More specifically, the softmax value corresponding to the highest predicted class.

[@CarlosBCU]

@@underfitted many thanks for your answer! What if we are running a regression?

[@modakad]

@@underfitted Answering CarlosBCU's question on confidence : I dont think your answer sufficiently clarifies the approach. Lets take an example. Suppose we have two classes, class 0, class 1. for observation A, softmax vector is [0.92,0.08] and for observation 2 its [0.60,0.40] {remember, Softmax gives a vector of values, which all add up to 1}. Which observation should we pick ? Not obs1. Obs2 is where the model has low confidence - as the model separates its predictions by only a magnitude of 0.2 (abs(0.6-0.4)) and in osb1, the separation is higher.

[@modakad]

@@CarlosBCU I think the answer would be - choose the observations with higher error (RMSE, MSE etc.)

[@modakad]

If you are using sigmoid loss function, then it would be trickier.