WEBVTT 00:00:00.000 --> 00:00:03.359 You know, usually when we talk about artificial 00:00:03.359 --> 00:00:06.059 intelligence getting smarter, there's this pretty 00:00:06.059 --> 00:00:08.380 terrifying tradeoff right at the center of it. 00:00:08.619 --> 00:00:11.330 Oh, absolutely. The data tradeoff. Right. To 00:00:11.330 --> 00:00:13.669 get a brilliant AI, you have to feed it just 00:00:13.669 --> 00:00:16.329 an absolute ocean of data. And we're talking 00:00:16.329 --> 00:00:18.969 about your private text messages, your deeply 00:00:18.969 --> 00:00:22.370 personal health records. The exact GPS coordinates 00:00:22.370 --> 00:00:24.690 of where you took a photo. Exactly. The exact 00:00:24.690 --> 00:00:27.370 locations. It's all getting sucked up. And that's 00:00:27.370 --> 00:00:29.429 been the fundamental assumption in computer science 00:00:29.429 --> 00:00:32.859 for, well, really for the last decade. The idea 00:00:32.859 --> 00:00:36.420 has always been that machine intelligence requires 00:00:36.420 --> 00:00:39.820 this massive centralized stockpile of personal 00:00:39.820 --> 00:00:42.640 information. It's a completely data gathering 00:00:42.640 --> 00:00:45.219 paradigm. You basically surrender your privacy 00:00:45.219 --> 00:00:48.560 to a massive server farm and in exchange you 00:00:48.560 --> 00:00:50.899 get a smart voice assistant or a really good 00:00:50.899 --> 00:00:53.340 recommendation algorithm. But what if that assumption 00:00:53.340 --> 00:00:55.939 is just completely wrong? Because today we are 00:00:55.939 --> 00:00:58.320 doing a deep dive into a comprehensive set of 00:00:58.320 --> 00:01:01.000 sources on this massive paradigm shift. It's 00:01:01.000 --> 00:01:03.679 called federated learning. Yeah. And it's fascinating. 00:01:04.019 --> 00:01:07.079 It really is. And our mission for this deep dive 00:01:07.079 --> 00:01:11.280 is to help you understand the brilliant, almost 00:01:11.280 --> 00:01:14.760 paradoxical way that AI is now learning from 00:01:14.760 --> 00:01:17.079 our most sensitive private information without 00:01:17.079 --> 00:01:19.060 ever actually looking at it. Which is, I mean, 00:01:19.120 --> 00:01:21.200 it truly turns the architecture of the internet 00:01:21.340 --> 00:01:24.500 upside down. We are moving away from that centralized 00:01:24.500 --> 00:01:27.000 data gathering model and shifting toward what 00:01:27.000 --> 00:01:29.400 we call decentralized intelligence. Decentralized 00:01:29.400 --> 00:01:31.060 intelligence. I like that. Yeah, it completely 00:01:31.060 --> 00:01:34.019 changes the power dynamic between you, the user, 00:01:34.200 --> 00:01:36.359 and the massive tech company. OK, so let's unpack 00:01:36.359 --> 00:01:39.760 this. Because to really appreciate why federated 00:01:39.760 --> 00:01:41.760 learning, or FL, as we'll probably end up calling 00:01:41.760 --> 00:01:44.340 it, why it's so revolutionary, we need to look 00:01:44.340 --> 00:01:46.680 at how traditional machine learning usually works, 00:01:46.700 --> 00:01:49.120 and frankly, why it's a bit of a privacy nightmare. 00:01:49.180 --> 00:01:53.359 Right. Distributed Learning is basically a bunch 00:01:53.359 --> 00:01:56.599 of powerful, multi -billion dollar data centers. 00:01:56.840 --> 00:01:59.040 They're all connected by these incredibly fast, 00:01:59.299 --> 00:02:01.400 dedicated fiber optic networks. Best hardware 00:02:01.400 --> 00:02:04.019 money can buy. Exactly. They take a massive, 00:02:04.280 --> 00:02:06.379 perfectly organized data set, they chop it up, 00:02:06.540 --> 00:02:08.419 crunch the numbers, and boom, you build an AI 00:02:08.419 --> 00:02:11.319 model. But federated learning doesn't use those 00:02:11.319 --> 00:02:14.199 pristine data centers, does it? Not at all. Actually, 00:02:14.219 --> 00:02:16.860 it relies on the absolute chaos of the real world. 00:02:16.919 --> 00:02:20.919 It's chaos. Total chaos. It uses messy, unpredictable 00:02:20.919 --> 00:02:24.060 devices. We're talking about, like, your five 00:02:24.060 --> 00:02:26.879 -year -old smartphone running on a spotty coffee 00:02:26.879 --> 00:02:29.819 shop Wi -Fi connection with a battery that's 00:02:29.819 --> 00:02:32.360 hovering at, you know, maybe 2%. Wait, I mean, 00:02:32.360 --> 00:02:34.379 that sounds like a terrible foundation for building 00:02:34.379 --> 00:02:37.319 a supercomputer. How does a dying phone on bad 00:02:37.319 --> 00:02:39.960 Wi -Fi compete with a... data center. Well, because 00:02:39.960 --> 00:02:42.759 the secret isn't in the hardware at all. It's 00:02:42.759 --> 00:02:44.620 in the architecture of what is actually being 00:02:44.620 --> 00:02:46.780 transmitted over that connection. OK. See, in 00:02:46.780 --> 00:02:48.860 a traditional centralized model, you send your 00:02:48.860 --> 00:02:51.500 raw data, like your actual text messages, up 00:02:51.500 --> 00:02:54.180 to the server for it to learn how you type. Which 00:02:54.180 --> 00:02:56.659 is terrifying. Right. But in federated learning, 00:02:57.139 --> 00:02:59.419 that raw data never leaves your device. Ever. 00:02:59.500 --> 00:03:02.960 It leaves the phone. Never. Instead, the central 00:03:02.960 --> 00:03:05.800 server sends a tiny sort of blank slate version 00:03:05.800 --> 00:03:08.039 of the machine learning model down to your phone. 00:03:08.460 --> 00:03:11.180 Oh, wow. So the AI comes to me rather than me 00:03:11.180 --> 00:03:13.879 having to send my data to the AI. Precisely. 00:03:14.180 --> 00:03:16.939 Your phone trains that localized model using 00:03:16.939 --> 00:03:19.840 your private data right there on your device's 00:03:19.840 --> 00:03:22.729 own internal processor. Then, instead of sending 00:03:22.729 --> 00:03:25.349 your data back, your phone only sends back the 00:03:25.349 --> 00:03:27.409 mathematical parameters. Just the math. Just 00:03:27.409 --> 00:03:29.629 the math. We're talking about abstract numbers, 00:03:29.930 --> 00:03:32.590 weights, and biases, and then the central server 00:03:32.590 --> 00:03:35.150 averages out all those mathematical updates from 00:03:35.150 --> 00:03:37.949 millions of phones to create a smarter global 00:03:37.949 --> 00:03:41.110 model. Okay, so think about it like this. Traditional 00:03:41.110 --> 00:03:44.110 AI is like a teacher demanding you hand in your 00:03:44.110 --> 00:03:47.250 deeply personal handwritten diary just so they 00:03:47.250 --> 00:03:49.830 can learn the nuances of human emotion. That's 00:03:49.830 --> 00:03:52.389 a great way to put it. Right. But federated learning 00:03:52.389 --> 00:03:54.509 is like the teacher giving you a blank worksheet. 00:03:54.990 --> 00:03:56.789 So you fill out the worksheet using the private 00:03:56.789 --> 00:03:59.229 experiences in your diary, but you only hand 00:03:59.229 --> 00:04:01.229 back the high level summary of the lessons you 00:04:01.229 --> 00:04:03.849 learned. And the diary itself. The actual diary 00:04:03.849 --> 00:04:06.280 stays locked inside your desk forever. That is 00:04:06.280 --> 00:04:08.759 an excellent analogy, really, because it highlights 00:04:08.759 --> 00:04:11.960 why this fundamentally serves the core tenets 00:04:11.960 --> 00:04:15.520 of data minimization. If the server only receives 00:04:15.520 --> 00:04:18.759 a mathematical summary. Just numbers. Yeah, literally 00:04:18.759 --> 00:04:20.600 just a series of decimal points representing 00:04:20.600 --> 00:04:23.100 how much a neural network node should care about 00:04:23.100 --> 00:04:25.600 a certain variable. It literally cannot read 00:04:25.600 --> 00:04:28.160 your texts or view your photos. Yeah. You maintain 00:04:28.160 --> 00:04:31.170 total physical ownership of your raw data. But 00:04:31.170 --> 00:04:32.949 here's where it gets really interesting, though. 00:04:33.269 --> 00:04:35.790 Because if the data never leaves the user's device, 00:04:36.089 --> 00:04:38.329 we run into a massive mathematical headache. 00:04:38.550 --> 00:04:40.889 A massive one. Because if the central server 00:04:40.889 --> 00:04:43.649 can't organize all the data in one place, it 00:04:43.649 --> 00:04:45.430 means everyone's data is completely different. 00:04:45.670 --> 00:04:49.029 The sources refer to this as the non -IID. problem, 00:04:49.250 --> 00:04:51.889 right? Non -independent and identically distributed 00:04:51.889 --> 00:04:55.110 data. Right. Non -IID. And to understand why 00:04:55.110 --> 00:04:57.509 that's such a nightmare for an AI, you have to 00:04:57.509 --> 00:05:00.009 look at what data centers normally do. In a data 00:05:00.009 --> 00:05:02.730 center, engineers spend months ensuring the data 00:05:02.730 --> 00:05:05.790 is perfectly balanced. Exactly. Cleanly labeled 00:05:05.790 --> 00:05:08.149 and perfectly identical in its distribution. 00:05:08.459 --> 00:05:12.279 But human data is naturally chaotic. When you 00:05:12.279 --> 00:05:15.180 rely on phones and laptops, you are just inviting 00:05:15.180 --> 00:05:18.180 all that chaos straight into the training process. 00:05:18.660 --> 00:05:21.279 It's so chaotic. And the research breaks this 00:05:21.279 --> 00:05:23.959 down into a few distinct categories of chaos, 00:05:24.019 --> 00:05:26.139 which I found fascinating. First, let's talk 00:05:26.139 --> 00:05:28.439 about the data itself changing. Right. So the 00:05:28.439 --> 00:05:31.379 first major hurdle is what we call covariate 00:05:31.379 --> 00:05:34.600 shift. Covariate shift. Yeah. This happens when 00:05:34.600 --> 00:05:37.899 the underlying features of the data change. even 00:05:37.899 --> 00:05:39.639 though the core meaning doesn't. Think about 00:05:39.639 --> 00:05:42.259 handwriting. Every single person listening to 00:05:42.259 --> 00:05:44.860 this right now. writes the letter A differently. 00:05:45.139 --> 00:05:47.319 Oh, sure. Different slants, different loops. 00:05:47.579 --> 00:05:49.800 Different stroke widths, different pressure on 00:05:49.800 --> 00:05:52.339 the screen if they're using a stylus. The AI 00:05:52.339 --> 00:05:54.139 has to learn that all these wildly different 00:05:54.139 --> 00:05:56.939 visual inputs mean the exact same letter. Right. 00:05:57.040 --> 00:05:59.100 And then you have something like prior probability 00:05:59.100 --> 00:06:01.740 shift, which is more about the sheer volume of 00:06:01.740 --> 00:06:03.720 certain data types based on simply where you 00:06:03.720 --> 00:06:05.680 are in the world. Regional differences, basically. 00:06:06.019 --> 00:06:09.379 Exactly. Datasets are incredibly regional. So 00:06:09.379 --> 00:06:11.920 if an AI is trying to learn what an animal looks 00:06:11.920 --> 00:06:14.899 like and it's training locally on devices, the 00:06:14.899 --> 00:06:17.720 photos of animals on a phone in, say, Sydney, 00:06:17.759 --> 00:06:21.279 Australia, are going to be overwhelmingly kangaroos 00:06:21.279 --> 00:06:24.180 and spiders. Right. But a phone in Toronto is 00:06:24.180 --> 00:06:26.560 going to be full of raccoons and squirrels. Exactly. 00:06:26.720 --> 00:06:28.839 And then the complexity just multiplies when 00:06:28.839 --> 00:06:30.779 we get into what's called concept drift. Concept 00:06:30.779 --> 00:06:32.920 drift, yeah. This is when the label remains the 00:06:32.920 --> 00:06:35.990 exact same. but the visual or structural features 00:06:35.990 --> 00:06:39.290 of the concept change entirely based on the environment. 00:06:39.670 --> 00:06:42.009 Give me an example of that. So for example, if 00:06:42.009 --> 00:06:44.129 you're training a computer vision model for a 00:06:44.129 --> 00:06:47.180 self -driving car, A photo of a stop sign looks 00:06:47.180 --> 00:06:49.379 completely different to a camera on a bright, 00:06:49.480 --> 00:06:52.620 sunny afternoon in Arizona compared to, say, 00:06:53.139 --> 00:06:55.300 a blizzard at midnight in Minnesota. Oh, wow. 00:06:55.459 --> 00:06:57.420 Yeah, same label, totally different image. Right. 00:06:57.759 --> 00:06:59.819 And then the opposite of that is concept shift. 00:07:00.199 --> 00:07:02.379 Same exact features, entirely different labels. 00:07:02.500 --> 00:07:04.680 Think about human language. Language is the perfect 00:07:04.680 --> 00:07:07.600 example here. Yeah, like the exact same string 00:07:07.600 --> 00:07:09.879 of text. Let's say someone tapes, oh, brilliant 00:07:09.879 --> 00:07:12.399 idea, genius. That might be labeled as totally 00:07:12.399 --> 00:07:15.420 sarcastic by a user in one demographic. But completely 00:07:15.420 --> 00:07:18.560 genuine by another. Exactly. So how does the 00:07:18.560 --> 00:07:21.399 global AI not just become totally confused by 00:07:21.399 --> 00:07:24.040 all this? I mean, it's like trying to write a 00:07:24.040 --> 00:07:26.459 universal dictionary when one town speaks in 00:07:26.459 --> 00:07:28.899 rigid formal English and the neighboring town 00:07:28.899 --> 00:07:31.800 only communicates in modern internet slang. Well, 00:07:31.819 --> 00:07:33.680 the way you prevent the global AI from having 00:07:33.680 --> 00:07:36.600 a total breakdown is through a specialized framework 00:07:36.600 --> 00:07:39.819 called Heterogeneous Federated Learning, or HETERFL 00:07:39.819 --> 00:07:42.839 for short. HETEROFL. Yeah, because it's important 00:07:42.839 --> 00:07:45.160 to remember it's not just the data that's non 00:07:45.160 --> 00:07:48.579 -ID. The hardware itself is non -ID. Right, the 00:07:48.579 --> 00:07:51.639 dying phone versus the new laptop. Exactly. An 00:07:51.639 --> 00:07:54.199 old first -generation smart thermostat simply 00:07:54.199 --> 00:07:56.899 cannot process neural network layers at the same 00:07:56.899 --> 00:08:00.139 speed or depth as a brand new smartphone. So 00:08:00.139 --> 00:08:02.120 HeteroFL is a structural design that basically 00:08:02.120 --> 00:08:04.720 allows the system to train these vastly varying 00:08:04.720 --> 00:08:08.120 local models with completely chaotic data while 00:08:08.120 --> 00:08:09.939 still managing to fold them all together into 00:08:09.939 --> 00:08:13.279 a single highly accurate global model. It's like 00:08:13.279 --> 00:08:15.819 it generalizes the world's knowledge without 00:08:15.819 --> 00:08:18.860 forcing every single device into a one size fits 00:08:18.860 --> 00:08:21.379 all computational box. That's exactly it. OK, 00:08:21.379 --> 00:08:23.500 so I understand the goal here, but let's talk 00:08:23.500 --> 00:08:26.220 about the actual execution, because how do we 00:08:26.220 --> 00:08:29.420 actually tame all this chaotic, unbalanced data 00:08:29.420 --> 00:08:32.340 across millions of devices without, like, breaking 00:08:32.340 --> 00:08:35.759 the Internet? We obviously need a highly structured 00:08:35.759 --> 00:08:38.200 iterative process. Oh, absolutely. It requires 00:08:38.200 --> 00:08:41.559 this incredibly delicate orchestration by the 00:08:41.559 --> 00:08:43.889 central server. We call this process a federated 00:08:43.889 --> 00:08:45.970 learning round. A round, right. And it starts 00:08:45.970 --> 00:08:48.149 with server initialization. Correct. The central 00:08:48.149 --> 00:08:50.889 server picks a baseline model, sets all the parameters, 00:08:51.009 --> 00:08:53.210 and gets it ready for deployment. But it doesn't 00:08:53.210 --> 00:08:56.009 just blast this out to every single device on 00:08:56.009 --> 00:08:58.389 Earth simultaneously, does it? No, no. If it 00:08:58.389 --> 00:09:01.250 did that, the sheer volume of data being downloaded 00:09:01.250 --> 00:09:04.370 and uploaded would crash global communication 00:09:04.370 --> 00:09:06.409 networks entirely. It's just a complete meltdown. 00:09:06.629 --> 00:09:08.830 Yeah. So this brings us to the second step, which 00:09:08.830 --> 00:09:12.370 is client selection. The server only picks a 00:09:12.370 --> 00:09:15.809 tiny, tiny fraction of available nodes. Maybe 00:09:15.809 --> 00:09:17.769 just a few thousand phones that happen to be 00:09:17.769 --> 00:09:19.970 plugged into a charger and connected to Wi -Fi 00:09:19.970 --> 00:09:22.730 at like two in the morning. Which is brilliant, 00:09:22.929 --> 00:09:24.990 right? Because it's completely invisible to the 00:09:24.990 --> 00:09:26.889 user. You're just sleeping and your phone is 00:09:26.889 --> 00:09:28.549 quietly doing its homework in the background. 00:09:28.950 --> 00:09:31.889 Exactly. Then comes the configuration step. where 00:09:31.889 --> 00:09:34.389 the server tells your device exactly how to train 00:09:34.389 --> 00:09:37.509 on the data it holds. And after your phone crunches 00:09:37.509 --> 00:09:40.009 all those numbers locally, we hit the reporting 00:09:40.009 --> 00:09:42.129 phase. Right, where the math gets sent back. 00:09:42.190 --> 00:09:44.330 Yeah, your device sends that mathematical summary, 00:09:44.389 --> 00:09:46.309 those adjusted weights we talked about, back 00:09:46.309 --> 00:09:49.710 up to the server. And finally, termination, where 00:09:49.710 --> 00:09:52.090 the server aggregates all the updates from those 00:09:52.090 --> 00:09:55.690 thousands of phones and finalizes the new, slightly 00:09:55.690 --> 00:09:58.049 smarter global model. But the real magic here, 00:09:58.049 --> 00:10:00.350 the thing that actually makes this whole round 00:10:00.350 --> 00:10:03.419 possible, Is the math doing the aggregating? 00:10:03.679 --> 00:10:06.820 Let's talk about the math. Yeah, so early on, 00:10:07.080 --> 00:10:09.480 the field relied heavily on an algorithm called 00:10:09.480 --> 00:10:12.039 FedAVS, that stands for Federated Averaging. 00:10:12.639 --> 00:10:15.360 See, in standard machine learning, models communicate 00:10:15.360 --> 00:10:17.960 by sending complex gradients. And gradients are? 00:10:18.100 --> 00:10:20.360 They're essentially the detailed step -by -step 00:10:20.360 --> 00:10:22.840 mathematical directions of exactly how the AI 00:10:22.840 --> 00:10:25.620 should change its mind. But sending those gradients 00:10:25.620 --> 00:10:28.100 constantly back and forth takes up way too much 00:10:28.100 --> 00:10:30.899 bandwidth. So FedAv changes the communication 00:10:30.899 --> 00:10:33.539 style completely? Exactly. FedAvstree allows 00:10:33.539 --> 00:10:35.940 the devices to run multiple batches of training 00:10:35.940 --> 00:10:38.440 locally, over and over, figuring out the best 00:10:38.440 --> 00:10:41.179 answer on their own first. Then, instead of sending 00:10:41.179 --> 00:10:42.980 the step -by -step directions, they just send 00:10:42.980 --> 00:10:45.600 the final destination, the final weights. Oh, 00:10:45.600 --> 00:10:47.980 I see. And the server simply averages those final 00:10:47.980 --> 00:10:50.500 weights together. It saves just a monumental 00:10:50.500 --> 00:10:52.740 amount of communication bandwidth. But wait, 00:10:52.940 --> 00:10:55.379 think about this. If you just blindly average 00:10:55.379 --> 00:10:57.379 everything together, and we already established 00:10:57.379 --> 00:11:00.460 that people's data is chaotic and wild, wouldn't 00:11:00.460 --> 00:11:03.159 some outlier device totally skew the average? 00:11:03.500 --> 00:11:05.799 Like, if my phone's data is heavily skewed by 00:11:05.799 --> 00:11:08.600 me constantly making weird typos, wouldn't my 00:11:08.600 --> 00:11:10.960 phone pull the entire global model off a cliff? 00:11:11.179 --> 00:11:12.960 That is a great point. And that was actually 00:11:12.960 --> 00:11:16.100 a massive vulnerability early on, which is why 00:11:16.100 --> 00:11:18.580 researchers had to develop algorithms like FedProx. 00:11:18.779 --> 00:11:21.320 FedProx. Yeah, FedProx introduces a mathematical 00:11:21.450 --> 00:11:23.490 constraint called a proximal term. OK, wait, 00:11:23.570 --> 00:11:26.450 let me stop you there. What exactly is a proximal 00:11:26.450 --> 00:11:28.730 term in plain English? OK, think of it like a 00:11:28.730 --> 00:11:30.809 mathematical rubber band. A rubber band. Yeah, 00:11:30.809 --> 00:11:33.490 it basically tethers your local model to the 00:11:33.490 --> 00:11:36.590 global baseline. Your phone is completely allowed 00:11:36.590 --> 00:11:39.669 to learn from your weird chaotic data, your typos, 00:11:39.850 --> 00:11:43.190 whatever. But if its resulting mathematical weights 00:11:43.190 --> 00:11:46.169 start drifting too wildly far away from what 00:11:46.169 --> 00:11:48.429 the rest of the world is learning, the rubber 00:11:48.429 --> 00:11:51.250 band snaps it back. Oh, wow. So it literally 00:11:51.250 --> 00:11:53.429 penalizes the local model for getting too far 00:11:53.429 --> 00:11:56.289 out of line. Precisely. It ensures the global 00:11:56.289 --> 00:11:59.509 model learns the subtle nuances of your data 00:11:59.509 --> 00:12:02.870 without letting your specific weird quirks hijack 00:12:02.870 --> 00:12:04.929 the entire system. That's amazing. I also really 00:12:04.929 --> 00:12:07.009 love the lottery ticket hypothesis mentioned 00:12:07.009 --> 00:12:09.649 in the research. Oh, via the SUBFED -AVG algorithm. 00:12:09.870 --> 00:12:12.909 Yes, SubFedAV. It basically figures out that 00:12:12.909 --> 00:12:15.409 neural networks are just massive, right? And 00:12:15.409 --> 00:12:17.850 most devices don't actually need the whole thing. 00:12:18.250 --> 00:12:20.909 So the system can personalize and sort of prune 00:12:20.909 --> 00:12:23.730 the network, essentially handing out smaller, 00:12:24.009 --> 00:12:26.470 highly targeted models to specific clients based 00:12:26.470 --> 00:12:28.909 on what they actually need. It's incredibly efficient. 00:12:29.210 --> 00:12:32.049 But if we are talking about algorithms, we absolutely 00:12:32.049 --> 00:12:35.309 have to talk about the huge 2024 breakthrough 00:12:35.309 --> 00:12:38.679 highlighted in our sources. Hi, FDCA. Yes, high 00:12:38.679 --> 00:12:42.259 IFFDCA, or for the full name, hybrid federated 00:12:42.259 --> 00:12:45.519 dual coordinate ascent. It is a monumental leap 00:12:45.519 --> 00:12:48.179 forward for the field. But to understand why 00:12:48.179 --> 00:12:50.120 it's such a big deal, you have to understand 00:12:50.120 --> 00:12:52.320 what a hybrid federated setting actually is. 00:12:52.320 --> 00:12:55.279 In the real world, clients rarely have a full, 00:12:55.500 --> 00:12:58.240 perfectly complete data set. So say one hospital 00:12:58.240 --> 00:13:00.399 might have patient ages and blood pressures, 00:13:00.620 --> 00:13:03.220 but no cholesterol records. Another hospital 00:13:03.220 --> 00:13:05.240 might have cholesterol and heart rate, but no 00:13:05.240 --> 00:13:07.909 ages. They only have random subsets. of features 00:13:07.909 --> 00:13:10.809 and samples. Older algorithms, like one called 00:13:10.809 --> 00:13:12.909 HIFEM, tried to solve this, but they were just 00:13:12.909 --> 00:13:15.250 a total nightmare to manage. Because they required 00:13:15.250 --> 00:13:18.779 tuning hyperparameters, right? Exactly. Hyperparameters 00:13:18.779 --> 00:13:22.259 are essentially the manual dials and knobs that 00:13:22.259 --> 00:13:24.860 human engineers have to physically set before 00:13:24.860 --> 00:13:27.879 the AI even runs. And if you set them wrong... 00:13:27.879 --> 00:13:31.919 The AI just fails to learn. It fails. HIFEM required 00:13:31.919 --> 00:13:35.019 engineers to manually balance four different 00:13:35.019 --> 00:13:38.240 highly sensitive dials just to get the local 00:13:38.240 --> 00:13:40.919 and global models to talk to each other properly. 00:13:41.120 --> 00:13:44.379 And HIFDCA fixes this. It completely revolutionizes 00:13:44.379 --> 00:13:48.309 it. HIFDCA solves what we call... convex problems. 00:13:48.409 --> 00:13:50.159 Which are what? They're basically mathematical 00:13:50.159 --> 00:13:53.299 puzzles that have one single perfect ultimate 00:13:53.299 --> 00:13:56.179 solution. And high FDCA reaches that perfect 00:13:56.179 --> 00:13:58.340 solution with an incredible convergence rate, 00:13:58.659 --> 00:14:01.000 meaning it zeros in on the exact right answer 00:14:01.000 --> 00:14:03.559 remarkably fast without wasting time bouncing 00:14:03.559 --> 00:14:05.679 around bad guesses. And what about the dials? 00:14:05.759 --> 00:14:08.279 That's the best part. It only requires the engineers 00:14:08.279 --> 00:14:10.940 to tune exactly one dial, just the number of 00:14:10.940 --> 00:14:12.659 inner iterations. Everything else is handled 00:14:12.659 --> 00:14:15.159 by the algorithm. That is wild. It's like a massive, 00:14:15.559 --> 00:14:19.179 impossibly complex group project. Imagine a thousand 00:14:19.179 --> 00:14:21.659 students are taking a massive final exam, but 00:14:21.659 --> 00:14:24.120 every student only read one random chapter of 00:14:24.120 --> 00:14:27.080 the textbook. And worse, some of them only read 00:14:27.080 --> 00:14:31.350 half the pages in that one chapter. HIA -IFDCA 00:14:31.350 --> 00:14:34.289 is basically the magical system that synchronizes 00:14:34.289 --> 00:14:37.389 all their partial fragmented notes so perfectly 00:14:37.389 --> 00:14:40.570 and so efficiently that the group as a whole 00:14:40.570 --> 00:14:43.850 still gets 100 % on the final exam. That's a 00:14:43.850 --> 00:14:45.690 really good way to visualize it. But let me push 00:14:45.690 --> 00:14:47.990 back on the physical reality of this for a second. 00:14:48.610 --> 00:14:51.049 What happens to this synchronized group project 00:14:51.049 --> 00:14:54.570 when someone's phone battery simply dies mid 00:14:54.570 --> 00:14:57.009 -update or, you know, they drive into a tunnel 00:14:57.009 --> 00:14:59.370 and totally drop off the network? See, that is... 00:14:59.309 --> 00:15:02.309 the exact bottleneck that plagued early federated 00:15:02.309 --> 00:15:05.909 learning. If the central server relies on synchronous 00:15:05.909 --> 00:15:08.570 updates, meaning every six there and waits for 00:15:08.570 --> 00:15:11.629 all 5 ,000 selected devices to report back before 00:15:11.629 --> 00:15:13.769 moving to the next round. The whole multi -million 00:15:13.769 --> 00:15:16.750 dollar system just grinds to a halt waiting for 00:15:16.750 --> 00:15:19.230 one guy's slow Wi -Fi in a basement somewhere. 00:15:19.429 --> 00:15:21.830 Exactly, which is completely unscalable. Totally. 00:15:21.970 --> 00:15:23.669 Right. So that is why the field is rapidly moving 00:15:23.669 --> 00:15:25.730 toward asynchronous updates and this brilliant 00:15:25.730 --> 00:15:28.370 technique called split learning. Split learning. 00:15:28.669 --> 00:15:32.039 Yeah. Neural networks are built in layers, kind 00:15:32.039 --> 00:15:34.779 of like a cake. Split learning doesn't wait for 00:15:34.779 --> 00:15:37.480 the device to bake the entire cake. As soon as 00:15:37.480 --> 00:15:39.639 the computations for the very first layer are 00:15:39.639 --> 00:15:42.200 finished, it transmits those weights back to 00:15:42.200 --> 00:15:43.759 the server immediately. Oh, so it just takes 00:15:43.759 --> 00:15:46.580 whatever mathematical progress is available exactly 00:15:46.580 --> 00:15:48.960 when it's available. Exactly. If your phone dies 00:15:48.960 --> 00:15:51.559 on layer three, the server still banks the progress 00:15:51.559 --> 00:15:53.720 you made on layers one and two, and the global 00:15:53.720 --> 00:15:56.139 round just continues without you. Okay, so what 00:15:56.139 --> 00:15:59.200 does this all mean? We have a system that protects 00:15:59.200 --> 00:16:02.320 raw data, overcomes the chaos of the real world, 00:16:02.960 --> 00:16:04.840 dynamically adjusts to dropping connections, 00:16:05.440 --> 00:16:08.309 and builds brilliant AI. I mean, it sounds like 00:16:08.309 --> 00:16:10.450 a perfect technological utopia. It really does. 00:16:10.730 --> 00:16:14.029 But every major tech leap has a dark side. And 00:16:14.029 --> 00:16:16.070 we have to look at the structural vulnerabilities 00:16:16.070 --> 00:16:18.470 of this architecture. Because the very thing 00:16:18.470 --> 00:16:20.370 protecting our privacy here, the fact that the 00:16:20.370 --> 00:16:22.970 data is utterly invisible to the central server, 00:16:23.330 --> 00:16:26.269 is the exact same thing making the AI incredibly 00:16:26.269 --> 00:16:28.250 vulnerable to attack. You're talking about the 00:16:28.250 --> 00:16:31.370 security mechanisms, specifically data poisoning 00:16:31.370 --> 00:16:34.559 and back doors. Exactly. Think about how a traditional 00:16:34.559 --> 00:16:37.340 data center works. If a hacker tries to feed 00:16:37.340 --> 00:16:40.759 an AI millions of pictures of stop signs, but 00:16:40.759 --> 00:16:43.000 they're maliciously labeled as speed limits. 00:16:43.360 --> 00:16:46.580 To try and crash self -driving cars. Right. The 00:16:46.580 --> 00:16:48.480 data center engineers can just look at the raw 00:16:48.480 --> 00:16:51.220 data, manually see the fake labels, and delete 00:16:51.220 --> 00:16:54.080 them. But in federated learning, the server only 00:16:54.080 --> 00:16:56.700 sees the mathematical summary. It cannot see 00:16:56.700 --> 00:16:59.159 the underlying photos. Which means malicious 00:16:59.159 --> 00:17:02.200 actors can seamlessly inject backdoors into the 00:17:02.200 --> 00:17:04.940 global model. How does that work? Well, a coordinated 00:17:04.940 --> 00:17:07.299 group of devices can intentionally train their 00:17:07.299 --> 00:17:10.140 local models on poisoned data, subtly tweaking 00:17:10.140 --> 00:17:12.700 those mathematical weights and biases. When they 00:17:12.700 --> 00:17:14.859 send their updates back to the server, the server 00:17:14.859 --> 00:17:17.339 just blindly averages them into the global brain. 00:17:17.690 --> 00:17:20.230 The malicious code is basically baked right into 00:17:20.230 --> 00:17:22.190 the math. So the AI might function perfectly 00:17:22.190 --> 00:17:25.549 like 99 % of the time? Right. But the attackers 00:17:25.549 --> 00:17:28.029 have installed a hidden trigger that causes the 00:17:28.029 --> 00:17:30.529 model to fail predictably under very specific 00:17:30.529 --> 00:17:33.569 conditions. Man, it's like wearing a thick blindfold 00:17:33.569 --> 00:17:36.529 to protect your visual identity. But because 00:17:36.529 --> 00:17:38.690 you're wearing a blindfold, you can't see if 00:17:38.690 --> 00:17:41.170 the person standing next to you is actively pouring 00:17:41.170 --> 00:17:44.970 poison into your drink. you are trading visibility 00:17:44.970 --> 00:17:48.009 for privacy, and you're losing security in the 00:17:48.009 --> 00:17:51.410 process. It's a huge trade -off. And beyond the 00:17:51.410 --> 00:17:53.650 technical security, this architecture raises 00:17:53.650 --> 00:17:56.650 incredibly difficult questions about governance 00:17:56.650 --> 00:17:59.269 and economics. I mean, it's an organizational 00:17:59.269 --> 00:18:02.230 nightmare. Because who actually owns the final 00:18:02.230 --> 00:18:04.730 model once it's built? Exactly. It's not just 00:18:04.730 --> 00:18:07.029 one company's data anymore. It's built by a consortium. 00:18:07.480 --> 00:18:10.380 So imagine a group of competing pharmaceutical 00:18:10.380 --> 00:18:13.000 companies or financial institutions, and they 00:18:13.000 --> 00:18:15.420 all agree to pool their mathematical updates 00:18:15.420 --> 00:18:18.299 to build a superior fraud detection AI. Sounds 00:18:18.299 --> 00:18:20.880 great in theory. Sure. But as more clients join 00:18:20.880 --> 00:18:22.720 the training process, the model gets smarter. 00:18:23.119 --> 00:18:24.880 But eventually it hits a performance threshold. 00:18:24.940 --> 00:18:27.299 You get diminishing marginal utility. The first 00:18:27.299 --> 00:18:29.799 10 companies to join contribute massive leaps 00:18:29.799 --> 00:18:32.920 in intelligence. The 50th company to join, they 00:18:32.920 --> 00:18:34.619 barely move the needle. So you have a massive 00:18:34.619 --> 00:18:37.950 free rider problem. Precisely. If early participants 00:18:37.950 --> 00:18:41.170 contribute, say, 90 % of the mathematical value 00:18:41.170 --> 00:18:43.529 required to make the model brilliant, why should 00:18:43.529 --> 00:18:46.410 a latecomer who only contributed 1 % of the value 00:18:46.410 --> 00:18:49.829 get equal ownership, equal access, or equal commercial 00:18:49.829 --> 00:18:52.309 rewards? That's a really tough question. The 00:18:52.309 --> 00:18:54.650 asymmetric nature of who contributes what and 00:18:54.650 --> 00:18:57.490 exactly when they contribute it creates massive 00:18:57.490 --> 00:19:01.180 organizational tension. Resolving the consortium 00:19:01.180 --> 00:19:04.180 contracts is often much, much harder than resolving 00:19:04.180 --> 00:19:06.660 the actual algorithms. I believe it. So we've 00:19:06.660 --> 00:19:09.059 talked about the math, the algorithms, the vulnerabilities, 00:19:09.720 --> 00:19:11.700 but what does this actually look like when life 00:19:11.700 --> 00:19:15.000 is on the line? Where is federated learning happening 00:19:15.000 --> 00:19:17.299 in the real world right now? Let's look at healthcare. 00:19:17.519 --> 00:19:19.180 Oh, healthcare is a prime example. Yeah, the 00:19:19.180 --> 00:19:21.319 source material highlights a phenomenal open 00:19:21.319 --> 00:19:24.099 source platform called Medperf. Recently, there 00:19:24.099 --> 00:19:27.819 was this massive 20 institution global collaboration 00:19:27.819 --> 00:19:30.440 published in Nature Medicine. That study is a 00:19:30.440 --> 00:19:32.660 perfect distillation of why this technology matters 00:19:32.660 --> 00:19:35.740 so much. They use federated learning to predict 00:19:35.740 --> 00:19:38.619 the oxygen needs of COVID -19 patients. And just 00:19:38.619 --> 00:19:40.720 think about the stakes of that. During the height 00:19:40.720 --> 00:19:43.279 of the pandemic, predicting exactly which patients 00:19:43.279 --> 00:19:45.779 would crash and need a ventilator hours before 00:19:45.779 --> 00:19:48.079 it actually happened was literally the difference 00:19:48.079 --> 00:19:50.789 between life and death. Absolutely. But to build 00:19:50.789 --> 00:19:53.970 an AI capable of predicting that accurately across 00:19:53.970 --> 00:19:56.809 diverse populations, you need data from thousands 00:19:56.809 --> 00:19:59.809 of diverse patients. But medical privacy laws, 00:19:59.970 --> 00:20:02.589 rightfully so, make it totally illegal to just 00:20:02.589 --> 00:20:04.869 email patient health records across international 00:20:04.869 --> 00:20:08.509 borders. So they didn't. Exactly. Using the federated 00:20:08.509 --> 00:20:12.009 architecture, those 20 hospitals kept every single 00:20:12.009 --> 00:20:14.789 patient's private medical record locked tightly 00:20:14.789 --> 00:20:17.690 in their own local, heavily regulated databases. 00:20:18.439 --> 00:20:21.940 Never. The model traveled to the hospitals, learned 00:20:21.940 --> 00:20:24.079 the complex physiological markers indicating 00:20:24.079 --> 00:20:26.720 a crash in oxygen levels, and only transmitted 00:20:26.720 --> 00:20:28.559 the mathematical insights back to the central 00:20:28.559 --> 00:20:31.700 hub. They built a highly accurate, generalized 00:20:31.700 --> 00:20:35.039 AI that literally saved lives without ever moving 00:20:35.039 --> 00:20:37.309 a single private record. It's just incredible. 00:20:37.549 --> 00:20:40.089 And it's also fundamentally changing transportation. 00:20:40.390 --> 00:20:43.069 Think about autonomous vehicles. A self -driving 00:20:43.069 --> 00:20:45.789 car is essentially a rolling supercomputer constantly 00:20:45.789 --> 00:20:48.430 executing machine learning models from computer 00:20:48.430 --> 00:20:52.490 vision to pacing algorithms. Right. In a centralized 00:20:52.490 --> 00:20:55.309 system, if a car encounters a wildly bumpy road 00:20:55.309 --> 00:20:58.230 or an unexpected obstacle, it has to send that 00:20:58.230 --> 00:21:01.009 raw sensor data up to a cloud server, wait for 00:21:01.009 --> 00:21:03.349 the server to process it, and wait for a command 00:21:03.349 --> 00:21:05.730 to come back down. Which is entirely unacceptable. 00:21:05.710 --> 00:21:07.910 when you were traveling at 70 miles per hour. 00:21:08.309 --> 00:21:10.369 Waiting for a cloud server introduces latency. 00:21:10.829 --> 00:21:13.549 And in an autonomous vehicle, a half second of 00:21:13.549 --> 00:21:16.910 latency is a literal fatal safety risk. It's 00:21:16.910 --> 00:21:19.769 a crash. Right. So federated learning shifts 00:21:19.769 --> 00:21:23.029 the computation to the edge. The car adapts immediately 00:21:23.029 --> 00:21:25.369 using its local model, learns how to navigate 00:21:25.369 --> 00:21:27.990 the obstacle in real time, and then simply shares 00:21:27.990 --> 00:21:30.269 the mathematical insight with the rest of the 00:21:30.269 --> 00:21:32.890 global fleet later. It completely eliminates 00:21:32.890 --> 00:21:35.190 the dangerous delay of transferring raw video 00:21:35.190 --> 00:21:39.289 data. And it's making massive waves in biometrics 00:21:39.289 --> 00:21:43.700 and what we call Industry 4 .0. Imagine smart 00:21:43.700 --> 00:21:46.000 manufacturing factories that want to improve 00:21:46.000 --> 00:21:49.220 their automated safety protocols. Fiercely competitive 00:21:49.220 --> 00:21:51.799 companies absolutely refuse to share their raw 00:21:51.799 --> 00:21:54.660 factory data because it contains proprietary 00:21:54.660 --> 00:21:57.460 trade secrets. Obviously. So federated learning 00:21:57.460 --> 00:21:59.619 lets them pool their safety insights without 00:21:59.619 --> 00:22:02.039 ever exposing their manufacturing processes. 00:22:02.559 --> 00:22:04.740 Or, you know, just look at your own phone. You 00:22:04.740 --> 00:22:07.160 are likely participating in federated learning 00:22:07.160 --> 00:22:09.160 right this second. Oh, for sure. Every time your 00:22:09.160 --> 00:22:11.299 phone's keyboard accurately predicts your next 00:22:11.299 --> 00:22:14.359 word, based on your unique typing habits or your 00:22:14.359 --> 00:22:16.960 camera recognizes your face to unlock the screen, 00:22:17.440 --> 00:22:19.440 you are acting as a local node. You're part of 00:22:19.440 --> 00:22:21.359 the system. Yeah, the system is keeping your 00:22:21.359 --> 00:22:23.400 fingerprints and facial templates strictly on 00:22:23.400 --> 00:22:26.119 your device, mathematically updating its accuracy 00:22:26.119 --> 00:22:28.500 without ever sending a picture of your face to 00:22:28.500 --> 00:22:31.759 a server farm in California. It is really incredible. 00:22:32.140 --> 00:22:35.769 So let's just recap our journey today. We started 00:22:35.769 --> 00:22:38.750 with the chaotic reality of the real world, how 00:22:38.750 --> 00:22:42.630 non -IID data like messy handwriting, wild regional 00:22:42.630 --> 00:22:45.109 photo differences, and varying computational 00:22:45.109 --> 00:22:48.109 power makes decentralized learning an absolute 00:22:48.109 --> 00:22:49.990 mathematical nightmare. The total nightmare, 00:22:50.049 --> 00:22:52.630 yeah. We explored how clever frameworks orchestrate 00:22:52.630 --> 00:22:55.490 the communication and how algorithms like FedAV, 00:22:55.869 --> 00:22:58.109 the mathematical tethers of FedProx, and the 00:22:58.109 --> 00:23:00.829 single -dial brilliance of high FDCA actually 00:23:00.829 --> 00:23:03.809 overcome that chaos. The heavy lifters. Exactly. 00:23:04.299 --> 00:23:06.619 We navigated the tricky, invisible waters of 00:23:06.619 --> 00:23:08.859 data poisoning and consortium -free writers, 00:23:09.099 --> 00:23:11.480 and we saw how this technology is actively predicting 00:23:11.480 --> 00:23:14.359 oxygen needs in hospitals and removing latency 00:23:14.359 --> 00:23:17.099 from our roads, all while keeping our metaphorical 00:23:17.099 --> 00:23:19.619 diaries locked safely in our desks. It really 00:23:19.619 --> 00:23:21.819 is one of the most elegant solutions to a modern 00:23:21.819 --> 00:23:23.720 crisis that computer science has produced in 00:23:23.720 --> 00:23:26.579 a decade. It allows fierce rivals, whether they 00:23:26.579 --> 00:23:29.059 are competing hospital networks, rival automakers, 00:23:29.079 --> 00:23:31.819 or massive tech giants, to collaborate on a shared 00:23:31.819 --> 00:23:33.819 brain without ever having to surrender. their 00:23:33.819 --> 00:23:36.319 proprietary crown jewels. It's diplomacy through 00:23:36.319 --> 00:23:39.119 math. It really is. But, you know, looking at 00:23:39.119 --> 00:23:42.079 all of this leaves me with one final, perhaps 00:23:42.079 --> 00:23:45.470 unsettling thought to consider. Oh. What's that? 00:23:45.690 --> 00:23:47.690 Well, if our phones, our cars, and our hospitals 00:23:47.690 --> 00:23:49.829 are constantly and silently collaborating in 00:23:49.829 --> 00:23:52.890 the background to build a massive global intelligence, 00:23:53.569 --> 00:23:56.069 an intelligence that is so decentralized that 00:23:56.069 --> 00:23:59.009 no single human can fully see its data and no 00:23:59.009 --> 00:24:01.430 single corporation can fully own its origins. 00:24:01.589 --> 00:24:04.269 Okay. Are we transitioning away from AI as a 00:24:04.269 --> 00:24:07.740 discrete software product? Are we instead slowly 00:24:07.740 --> 00:24:10.200 building a collective digital subconscious for 00:24:10.200 --> 00:24:13.319 the human race? Wow. A global digital subconscious 00:24:13.319 --> 00:24:16.000 built entirely from the invisible math of a billion 00:24:16.000 --> 00:24:18.259 private lives. That is definitely something to 00:24:18.259 --> 00:24:20.839 ponder the next time your phone seamlessly autocompletes 00:24:20.839 --> 00:24:22.680 your sentence. Thanks for diving deep with us.