วีดีโอ

Thank you, Serine! I'm glad that you liked the video :)

Thank you for letting me know, makes me glad to hear that!

You are correct! Essentially, your approach is the same as what I do in the video, just based more on intuition and less on mathematical rigor. During my daily work as an engineer, my solutions look much more like what you're writing in this comment and less like what it looks like in this video ^^ However, as you point out, intuition only takes you so far, and it can be misleading. Many times, I convince myself of a solution at the white board, and then when I go to implement the resulting algorithm in real life, it doesn't perform as expected. Usually, because I made some mistake in the intuitive solution. Then, you have to go through the slog of looking at every detail and figuring out where your intuition failed :P

That's awesome, Rashawn! Finding a new way to solve the problem really reinforces that you understood it.

If the die was a normal die (all six sides show up with equal probability), then the answer 1/2 would be correct! However, note that the die is not fair. This die is loaded such that the outcomes "one", "two", and "three" turn up twice as often as the remaining three outcomes. I describe the problem at 0:37 Intuitively, since there are more odd numbers than even numbers in the set {1,2,3}, we should expect even numbers to show up less often than 50% of the time.

​@@henrikhellstrom1241 oh yes now i understand, i missed that point i guess. thank you for replying to me.

Hello Taartin, thank you for your comment! First question: The server is able to broadcast the new model because the weighted sum is exactly equal to the new model, i.e., the equation that you see inside of the "cloud" at 4:47 is how the new global model is generated. This model update step comes directly from the original description of Federated Learning. You can find that description in "Communication-Efficient Learning of Deep Networks from Decentralized Data" by McMahan et al. Second question: The device will take the broadcasted global model w^t and then compute a local update w_k^t by running a training algorithm. The most common learning algorithms are variants of the stochastic gradient descent algorithm. These algorithms are extremely popular, so a search for "gradient descent" will give you many good sources on the topic.

@@henrikhellstrom1241 I did not realize that the weighted model was for each parameter. I thought it was one sum for all parameters which did not make sense to me. Thank you very much for your answer:)

​@@Taartin Ah, I see! Yes, perhaps I should have been more clear about that. :) To clarify for the future: the idea is that each parameter in the global weight vector is computed as a separate function over-the-air. In other words, if you have a 100,000 dimensional model and K users, you compute 100,000 functions. In an orthogonal communication scheme, you would communicate 100,000*K parameters and then compute 100,000 functions in the CPU at the server.

Thank you for your kind comment, Salah!

Hello, k j! That's right, the man will have a win probability of approximately 68%! However, note that when the man wins, you are only losing 1 dollar, and when you win you are gaining 3 dollars. So even though the man wins 68% of all games, you are still making a profit on average!

Will do!!