Neural Structured Learning - Part 1: Framework overview

Рет қаралды 43,527

Күн бұрын

In this video series, we’re going to introduce a new learning framework to our developers called Neural Structured Learning. TensorFlow Senior Software Engineer Da-Cheng Juan kicks off this series with the framework overview.
Links:
Get started with Neural Structured Learning → goo.gle/2sOz6DE
Hands-on tutorial → goo.gle/37pAD27
Coding TensorFlow → goo.gle/Coding-TensorFlow
Subscribe to the TensorFlow channel → goo.gle/TensorFlow

Пікірлер: 31

@rohanmanchanda5250 2 жыл бұрын

Showed this to my three year old cousin. One year later, he's now a neuro-scientist (neural network and machine learning researcher). Awesome work Shang Chi!!

@robertcrowe7000 4 жыл бұрын

Great stuff, can't wait for part 2!

@bhuvaneshs.k638 4 жыл бұрын

This is interesting ....!! Thqs for this new approach

@oo7knight 4 жыл бұрын

Interesting !! look forward o following it more closely

@vagifmolla5978 4 жыл бұрын

quite interesting idea, will check that!

@jorgevaldez3478 4 жыл бұрын

Very interesting, but where do we get this structure from? Is something we should have or can we create it with other neural network?

@pradeepkumar-qo8lu 4 жыл бұрын

Is this analogy close enough to assume ?? King - man +woman in word2vec gives queen because of taking into account the 'neighbouring' vectors ie king, man, woman however here we basically pointed out the neighbours to be used, what is being proposed here in the video is to take into account the network structure (not sure if all layers or just the top few layers) and feed it along with images (pixel features) basically feeding in context + feature for the network to train/learn on

@christianstrickhausen8070 4 жыл бұрын

Hi Juan, possible to show this approach to Regression problems with structured assumptions over the data? For example: Lerning a regession on y= -5 + 2 x - 0.5 x**2 with tensorflow needs a lot of trainings but fits bad at the corners and is mostly looking like a spline solution. To use a system of structured assumption (list of formulas) can solve this problem mutch faster with less experiments. Do have an Idea how to nest this in your approach? Kind regards Christian

@fahemhamou6170 Жыл бұрын

تحياتي الخالصة شكرا جزيلا

@meifishK 4 жыл бұрын

Also, in the Training phase, we have both the sample image and its neighbors information as INPUT. However in the Testing phase, when the label of the input image is unknown, how are we able to have the neighbors information to be part of the input?

@jazlielee9995 4 жыл бұрын

Can we use Neural Structured Learning with NLP LSTM text classification?

@muniaak5864 4 жыл бұрын

is this approach the same as graph convolutional networks?

@axeldroid2453 4 жыл бұрын

It looks similar to siamese networks. Is there a major difference between them ?

@ChopLabalagun 4 жыл бұрын

is this part of TF2.0? of we can work on this on old releases.

@sanikamal 4 жыл бұрын

This is interesting

@Jos241988 4 жыл бұрын

Don’t really understand what is meant by structure here, the relation between samples?

@samrijijkot 4 жыл бұрын

As I understand it, yes it. Information on similarity of the content of the pictures. For example, two bulldogs would be close in that graph and a retriever would be a bit farther, but a cat would be very far...or mabe not even in that graph? That part Im not clear about

@pradeepkumar-qo8lu 4 жыл бұрын

Word vectors but for images

@AllanPichardo 4 жыл бұрын

Yes, you can define what the relationship means according to your specific problem. It could be semantic similarity between words, it could be the distance between embeddings of other similar or dissimilar images, things like that.

@robn2497 4 жыл бұрын

Imagine you assign values to image features (the structure is a multi dimensional position in space) After training many images you get a space or a map of where dogs are roughly. So as you can classify a dog by the map. You cannot have seen all dogs. But the map means you don't need too.

@junhuangho6937 4 жыл бұрын

is this related to graph neural networks?

@johnlao1469 4 жыл бұрын

Do you have a video where it also classify image by its color?

@rahuldeora5815 4 жыл бұрын

I am also looking for this. You find anything?

@meifishK 4 жыл бұрын

How do we define "neighbors" and how to construct the structured signals? If it is like the word2vec idea, would the neighbors simply be the images that appear together in a same article or similar idea? I assume it wouldn't be from human knowledge, because if so, then it is nothing more than adding more "labeled data" into the training. Can anyone comment?