C4W4L02 One Shot Learning

Рет қаралды 104,184

Күн бұрын

Пікірлер: 32

@atulsain6170 Жыл бұрын

He uploaded this video 5 years ago. 😅 I am learning machine learning for more than a year and I learned this technique today.

@donghunpark379 5 жыл бұрын

In summary, rather than training a classifier that can classify a database by label, training a 'similarity function' that can 'distinguish' between different images makes (recognition)system free to the number of databases. Good Idea.

@lampham7874 5 ай бұрын

I have watched a few videos about few shot learning before this on but this is the most understandable.

@mahmoudfathy2074 4 жыл бұрын

I like the fact that I can not see them as the same person while the algorithm will learn better than me 🤣 Danielle's looking great in the second photo though 😁

@SalteRage 3 жыл бұрын

Isn't this the description of similarity (or metric) learning? How is one-shot learning different?

@fratcetinkaya8538 2 жыл бұрын

Does those similarity function work on another things, daily prodoucts etc. for example..

@adamlee9347 6 жыл бұрын

Danielle.. lol her second photo looks so much better

@tarat.techhh 4 жыл бұрын

lol true dat

@mofiro6758 3 жыл бұрын

Doesn't it need to retrain the network when adding a new person to the database? If so, what if we add like 10 people or 100 new people?

@petarulev6977 2 жыл бұрын

No. You would need to do it if you had a classification task (i.e you would need that your final dense layer has x+1 output probabilities after you add 1 more person on top of the database with x people). Not a good idea, since each class would have to see at least a couple of examples. On the other hand, you can see this as regression - have just a normal dense layer without activation that outputs a distance. I am speculating this distance is just the probability of that the pair is of the same person, but im not sure.

@Vinay1272 Жыл бұрын

The network still needs to learn the facial structure of the new person, right? In fact, the network needs to learn the facial structure of all the people in the database do calculate the differences. Isn't it?@@petarulev6977

@MrMikael1337 3 жыл бұрын

I understand that a similarity function could discriminate the new class from previous classes. But still, doesnt this require that the model has a very good "understanding" (i.e been trained on a lot of samples) of the previous classes? Otherwise, it wont understand what a face even is.

@osiris1102 3 жыл бұрын

You can train the network on many face images from the internet then it would be ready to tell if the two images are of the same person or not.

@sandipansarkar9211 3 жыл бұрын

nice explanation

@thiliniyatanwala2349 5 жыл бұрын

Thank you for sharing the knowledge ..Can you please give me some idea how zero shot/one shot learning can be used to apply in the area edge computing ? will edge computing be benefited from zero shot learning ?

@debarunkumer2019 4 жыл бұрын

How do we determine the threshold value for every data input ?

@MuhannadGhazal 4 жыл бұрын

the function will return the face distance and you will check this number with your own threshold. if it's below 0.4 then it's the same person, if above then it's a different person.

@trident8638 6 жыл бұрын

how to implement this using python and tensorflow?

@ammarazlan2919 5 жыл бұрын

Could you recommend an established model (like alexnet for image recognition) for time series forecasting?

@RinkiKumari-us4ej 4 жыл бұрын

he is not going to reply you😂😂

@MrAcenit 4 жыл бұрын

Isnt this just knn classification?

@Jononor 3 жыл бұрын

With kNN the distance metric is a standard function such as euclidean/cosine/manhattan distance, where as here the distance function is learned from data (using a neural network). This makes this approach work much better for complex high-dimensional data, such as images, audio etc.

@tsunghan_yu 6 жыл бұрын

is it like nearest neighbors?

@CyborgGaming99 5 жыл бұрын

I can see why you would think that, but no, it is not. The basis of KNN is that it measures Euclidean Distance of one point in comparison to others, and here we do NO such thing. Here, you just measure the similarity of two compared images, no distance between them, or nothing like that which you would find in KNN or some clustering algorithms

@ta6847 5 жыл бұрын

It depends on how the difference function is implemented. You can imagine representing each face as some embedding or encoding, and then using some distance metric to determine similarity.

@ta6847 5 жыл бұрын

Apparently that's exactly what's described here: kzbin.info/www/bejne/bJvJqGuDqrCqpqs Again, not nearest neighbor exactly, but definitely the same flavor.