good video. re: secure aggregation. why do you bother with the 'buddy' system? wouldn't this work if each individual phone is sent and uses random values to secure it's data in trasit? what does the buddy thing add?
Training data is only from suitable clients. Once trained, the model parameters are updated by testing it with other clients. Yes, model parameters means only weights here.
very good morning mam, your explanation is very nice. mam, can we use federated learning on the internet of vehicles environment? can you suggest few use cases of this learning for vehicles?
Keyword here seems like other parameters learned by the model. Weights and biases in themselves are okay, but without the features associated with them would be useless? What's returned would need to be studied too by somebody. How is data privacy being ensured?
Doesn't this approach create a "bias bubble" around all the clients in the learning process? I think it does. If you acquire your information through this process you will be just like everyone else. Now you don't have to worry about being different.
First of all, thanks for an amazing video! I want to ask three questions to clarify my understanding: 1. You've said for model to be distributed user need to be suitable available, if not at the time would it wait until user becomes suitable available(go home at night and charge) then distribute model for training? or if at the time of typing(in our ex case) if user's device is not suitable available they can't learn at all? 2. @3:04 you said only updated model's weights, biases and other parameter leaves the model not the model itself. But right after you are saying "server gets locally trained models". Could you clarify? 3. You said near the beginning that benefit of federated learning is to improve UX by not Giving/Getting data to and from server because internet connectivity, network latency, and others can affect giving/getting data. However we get model from server then pass changed weights and biases to the server and this process is repeated, I'm not sure why giving/getting model from server doesn't face bottlenecks I've mentioned.
2. - "server gets locally trained models" mean their weights, biases and other parameters leave the device and reach the server. If someone asks me to show how the model looks in literal sense, I would show them the weights, biases and other parameters. 3. giving/getting model happens in the background and this process does not affect the user experience because it is happening in the background. 1. I don't quite understand the question.
@@satish9367 Great answers Satish! Also, the answer to 3. would be the fact that the above mentioned exchanges does not happen every now and then in real time. Anything happening in real time in the background would have hindered the process due to internet connectivity, network latency and other factors he mentioned. But this happens in a versioning kind of system I guess, where the trained parameters gets updated to the server once in a while and the improved model is updated to every client device as an improved version of the app.
well ! thanks for a great explaination. that was an amazing concept. looking forward for future advancements in federated learning. i would also like to do a contribution.
If the model needs to trained in a client it needs tensorflow or flower or at least the python framework to train. But the python framework is not installed (is it installed by Android itself by default when the rom is built??)
Great Video. Is it possible to train a pre-trained image recognition model on mobile device. I dont really need it to be federated but just to be trained and used on an ios and android for use on that device. However it would be nice if it was federated learning but dont really need it for my current use case. Any advice would be great. Newbie to ML
Great technology but suddenly I worry about my Phone's battery life.... May be detection if the iPhone is on Charging and we start the program to train the data. I think training Data take a lot resources... Although in this case just keyword (keyboard) input data learning wont take much.. But for bigger data like photo and video, it is different story
Get $300 and start running workloads for free → goo.gle/3tss1Dh
Nice video
Tesla Autopilot system. Camera data from each car can be used to train the model.
How could I not like a video presented with such enthusiasm? It is an excellent idea as well.
This piece is actually great, thank you Priyanka. Where and how do I start my federated learning journey please?
Cool bookshelf
Thanks!
I have started my research on federated learning , it's really cool@
You should also try voice modelling. Your voice and way of speaking is really good!!!!!
thank u, lots of love from Bangladesh
That is freaking well explained, thank you!
Thankyou mam wonderfull Introduction
That was comprehensive due to your style of instruction. Thanks!
Thank you for such am informative video. Easy to understand with such a simple words and great explanation.
good video. re: secure aggregation. why do you bother with the 'buddy' system? wouldn't this work if each individual phone is sent and uses random values to secure it's data in trasit? what does the buddy thing add?
Very well explained!! But, how to monitor the model performances, be sure that there is no bias,... Because you haven't data and thus a test dataset!
Training data is only from suitable clients. Once trained, the model parameters are updated by testing it with other clients. Yes, model parameters means only weights here.
Great! thanks very much!
You are welcome!
very good morning mam,
your explanation is very nice. mam, can we use federated learning on the internet of vehicles environment?
can you suggest few use cases of this learning for vehicles?
Thank you!
You're welcome!
Keyword here seems like other parameters learned by the model. Weights and biases in themselves are okay, but without the features associated with them would be useless? What's returned would need to be studied too by somebody.
How is data privacy being ensured?
Hi, may I check can I use this to work like weglot for my website?
Doesn't this approach create a "bias bubble" around all the clients in the learning process? I think it does. If you acquire your information through this process you will be just like everyone else. Now you don't have to worry about being different.
First of all, thanks for an amazing video! I want to ask three questions to clarify my understanding:
1. You've said for model to be distributed user need to be suitable available, if not at the time would it wait until user becomes suitable available(go home at night and charge) then distribute model for training? or if at the time of typing(in our ex case) if user's device is not suitable available they can't learn at all?
2. @3:04 you said only updated model's weights, biases and other parameter leaves the model not the model itself. But right after you are saying "server gets locally trained models". Could you clarify?
3. You said near the beginning that benefit of federated learning is to improve UX by not Giving/Getting data to and from server because internet connectivity, network latency, and others can affect giving/getting data. However we get model from server then pass changed weights and biases to the server and this process is repeated, I'm not sure why giving/getting model from server doesn't face bottlenecks I've mentioned.
2. - "server gets locally trained models" mean their weights, biases and other parameters leave the device and reach the server. If someone asks me to show how the model looks in literal sense, I would show them the weights, biases and other parameters.
3. giving/getting model happens in the background and this process does not affect the user experience because it is happening in the background.
1. I don't quite understand the question.
@@satish9367 Great answers Satish! Also, the answer to 3. would be the fact that the above mentioned exchanges does not happen every now and then in real time. Anything happening in real time in the background would have hindered the process due to internet connectivity, network latency and other factors he mentioned. But this happens in a versioning kind of system I guess, where the trained parameters gets updated to the server once in a while and the improved model is updated to every client device as an improved version of the app.
please post video of working model for ferderated learning
well ! thanks for a great explaination. that was an amazing concept. looking forward for future advancements in federated learning. i would also like to do a contribution.
Is the concept of federated learning the same as swarm learning?
Nice GCP logo
Thanks! 😀
Hearing google metion "eliminating biases" is truly an oxymoron. Extreme Bias is one of their pillars..
If the model needs to trained in a client it needs tensorflow or flower or at least the python framework to train. But the python framework is not installed (is it installed by Android itself by default when the rom is built??)
thats so cool!
Great Video. Is it possible to train a pre-trained image recognition model on mobile device. I dont really need it to be federated but just to be trained and used on an ios and android for use on that device. However it would be nice if it was federated learning but dont really need it for my current use case. Any advice would be great. Newbie to ML
What about media, documents, files ect.?
Excelent!!!
would be cool if they told us we were training their models while using a keyboard
🙏
Great technology but suddenly I worry about my Phone's battery life.... May be detection if the iPhone is on Charging and we start the program to train the data.
I think training Data take a lot resources... Although in this case just keyword (keyboard) input data learning wont take much.. But for bigger data like photo and video, it is different story
I am trying to do this using Python 2 and Pawn. This requires unusual knowledge such as how LISP functions.
You're talking about the datasets. I understand data, but I haven't used data for language. I was going to buy GPU clusters soon for training models.
Why is FedAvg called Vanilla Federated Learning ? Is there any specific reason for this,can anyone plz explain ?
It's the first proposed and the most naive. It just takes weighted average of the weights.
Is this happening now, or is this future?
It has been in use for some years now actually. So this is currently in use. GBoard is just one of the many places where FL is used.
Sounds frightening to me.
If I must see you all this video when all I want is your explanation, at least stop moving your hands. Will you?
I like it.