WEBVTT 1 00:00:01.520 --> 00:00:05.200 Hello, and welcome, everybody. This is Joe from Stella Bright. They're all 2 00:00:05.200 --> 00:00:08.559 coming to you today again from a very, very hot Frankfurt 3 00:00:08.559 --> 00:00:12.275 area. We are expecting to crash the 30 degrees 4 00:00:12.495 --> 00:00:15.955 Celsius today. It's gonna be damn hot. 5 00:00:16.390 --> 00:00:19.369 Nonetheless, I do have Danco here with me. Hey. Welcome. 6 00:00:20.150 --> 00:00:20.650 Hello. 7 00:00:23.515 --> 00:00:27.275 Totally. My pleasure to have you here because you're a very 8 00:00:27.275 --> 00:00:30.494 interesting guy. You are a brain researcher 9 00:00:31.350 --> 00:00:35.084 turned entrepreneur and working on machine Menninger. 10 00:00:35.110 --> 00:00:38.675 And we get through this interview and, at the end, have a few 11 00:00:38.675 --> 00:00:42.195 ideas how machine learning can profit from your 12 00:00:42.195 --> 00:00:45.175 research and what your very interestingly 13 00:00:45.555 --> 00:00:48.490 named Startupradio, robots Joe Manta, 14 00:00:48.950 --> 00:00:52.470 actually does. But first, I would like to, 15 00:00:52.710 --> 00:00:56.325 to thank our enabler. This recording is supported by 16 00:00:56.785 --> 00:00:59.684 HDI, Essentrade and Invest, and Enterprise Europe Network. 17 00:01:01.265 --> 00:01:04.864 These organizations have made tremendous contributions to helping start up 18 00:01:04.864 --> 00:01:08.690 businesses succeed and thrive, providing a range of 19 00:01:08.690 --> 00:01:12.310 services from helping to find grants to ongoing partnerships. 20 00:01:12.930 --> 00:01:16.455 By taking advantage of these resources, startup companies can 21 00:01:16.755 --> 00:01:20.355 network and develop innovative strategies for success on the 22 00:01:20.355 --> 00:01:23.720 international stage. The dedicated support of both 23 00:01:23.720 --> 00:01:27.240 entities is paramount in providing start up businesses with the 24 00:01:27.240 --> 00:01:31.080 tools for lasting success. Look for our dedicated sub podcast 25 00:01:31.080 --> 00:01:34.045 in partnership with them called Tech Startups Germany. 26 00:01:34.825 --> 00:01:38.265 Now that we do have this out of the way, as 27 00:01:38.265 --> 00:01:41.965 always, I've been looking through your CV, and I found it 28 00:01:42.390 --> 00:01:46.070 extremely interesting. You have been 29 00:01:46.070 --> 00:01:49.530 teaching at the Technical University Darmstadt 30 00:01:52.505 --> 00:01:56.024 statistics. That Yes. Nonetheless, it it it takes 31 00:01:56.024 --> 00:01:59.570 it it takes quite a special person to really love 32 00:01:59.570 --> 00:02:03.170 statistics and even teach that on university level. And we have to say, 33 00:02:03.170 --> 00:02:06.610 Technical University is one of the top technical 34 00:02:06.610 --> 00:02:10.415 universities here in Germany, maybe even in Europe. So, they 35 00:02:10.415 --> 00:02:14.015 don't take anyone as statistics teacher. Then she 36 00:02:14.015 --> 00:02:17.790 became a group leader, but the interesting twist here is 37 00:02:17.849 --> 00:02:21.690 in brain research for the Max Planck Institute, can you take 38 00:02:21.690 --> 00:02:25.390 us along your journey how you came from statistics 39 00:02:25.735 --> 00:02:29.334 to brain research? Because then it kind of makes 40 00:02:29.334 --> 00:02:32.935 sense how you ended up in machine learning and 41 00:02:32.935 --> 00:02:36.239 artificial intelligence. Yeah. Yeah. 42 00:02:37.099 --> 00:02:40.379 Very gladly. So thank you very much for having me here. 43 00:02:41.180 --> 00:02:44.915 It's it's a pleasure to be here. So the story is a little 44 00:02:44.915 --> 00:02:47.655 bit different. So I was teaching it at, 45 00:02:48.435 --> 00:02:52.070 at only as a side gig when I 46 00:02:52.070 --> 00:02:55.910 was already at the at the, Max Planck Institute. So 47 00:02:55.910 --> 00:02:58.330 it was happening in parallel. Right? 48 00:02:59.515 --> 00:03:01.695 But how I came to the brain research, 49 00:03:03.275 --> 00:03:07.060 is is a story Okay. That's hard Let me get this straight. You 50 00:03:07.220 --> 00:03:10.980 you've been, at the same time, a BRAID researcher and a 51 00:03:10.980 --> 00:03:14.760 statistics teacher? Yeah. Yeah. Yeah. Yeah. Yeah. 52 00:03:15.254 --> 00:03:19.095 Let's do a more. One can do that. I think I think about brain 53 00:03:19.095 --> 00:03:22.694 research is always about biology. But, apparently, you can 54 00:03:22.694 --> 00:03:26.330 also apply pretty much statistics. Sorry to interrupt you, but it's just 55 00:03:26.489 --> 00:03:30.330 Oh, yeah. It's quite uncommon. There 56 00:03:30.330 --> 00:03:34.165 is a lot of statistics in brain research. Trust me. 57 00:03:34.225 --> 00:03:37.925 Like, when you get all this data, when you record all these brain signals, 58 00:03:38.625 --> 00:03:41.550 you have to know what to do with that. Right? You have, like, bunch of 59 00:03:41.550 --> 00:03:45.230 numbers in your computer, and you have to analyze them. You have to 60 00:03:45.230 --> 00:03:48.050 understand them. There's lots of statistics. Believe me. 61 00:03:49.775 --> 00:03:53.474 I was I was totally fine with all the statistics I've been taught 62 00:03:53.694 --> 00:03:57.375 in Germany and the US. Menninger high five through professor Owen from Midwestern 63 00:03:57.375 --> 00:04:00.830 State, by the way. He he was the only one who made it really 64 00:04:00.830 --> 00:04:04.210 entertaining. Disclaimer, I've never attended your lessons. 65 00:04:05.550 --> 00:04:09.085 Don't worry. And, also, statistics 66 00:04:09.385 --> 00:04:13.225 is in in many ways, statistics is understood as 67 00:04:13.225 --> 00:04:17.009 as discipline that preceded machine learning. Right? So 68 00:04:17.009 --> 00:04:20.229 machine learning somehow built on top of statistical 69 00:04:20.769 --> 00:04:24.610 research, statistical analysis, inventions in in 70 00:04:24.610 --> 00:04:28.095 in statistics as well. Right? But that wasn't 71 00:04:28.095 --> 00:04:31.555 really my story. Right? My story was 72 00:04:32.335 --> 00:04:35.870 that I was first interested in artificial intelligence, 73 00:04:37.130 --> 00:04:40.670 like, back back then. Realized 74 00:04:40.730 --> 00:04:44.444 that artificial intelligence is struggling with doesn't have a 75 00:04:44.444 --> 00:04:47.805 good solution. There is no good way of of 76 00:04:47.805 --> 00:04:51.420 emulating brain, human brain in a computer. 77 00:04:51.880 --> 00:04:55.240 And that's what triggered me to to try to understand the 78 00:04:55.240 --> 00:04:58.794 brain. So I started with studying psychology, getting 79 00:04:58.794 --> 00:05:02.235 PhD in cognitive psychology. So kind of first 80 00:05:02.235 --> 00:05:05.375 understanding the brain at the level of behavior. 81 00:05:06.060 --> 00:05:09.840 And then go then say, okay. Now I understand that, kind of. 82 00:05:10.460 --> 00:05:14.240 Now let's go into the brain. Then I went to to to, 83 00:05:14.940 --> 00:05:18.384 brain research. That's where I needed a lot of statistics. 84 00:05:20.044 --> 00:05:23.565 And then after that, as 85 00:05:23.930 --> 00:05:27.770 I don't know if you mentioned, but I then I went to industry after 86 00:05:27.770 --> 00:05:30.430 some time at at the Max Planck Institute. 87 00:05:31.370 --> 00:05:34.925 I I would be curious. How was the interview 88 00:05:35.545 --> 00:05:39.225 with HR there? Because, for everybody, we 89 00:05:39.225 --> 00:05:43.040 we do have a quite a vast audience in in Europe 90 00:05:43.040 --> 00:05:46.740 where Max Planck Institution should be pretty well known. 91 00:05:47.280 --> 00:05:51.044 But for for all the thousands of people listening from 92 00:05:51.044 --> 00:05:54.725 abroad, Max Planck Institutes are one of the 93 00:05:54.725 --> 00:05:58.290 premier basic research institutions in Germany. 94 00:05:58.430 --> 00:06:01.950 So you simply can't can't tell. I go there. You have to do 95 00:06:01.950 --> 00:06:05.470 your your undergraduate studies, your graduate studies, your 96 00:06:05.470 --> 00:06:06.734 PhD, and then do some impressive work that they will consider you. I do believe 97 00:06:06.734 --> 00:06:07.375 the vast majority of the 98 00:06:11.855 --> 00:06:15.315 noble arts in Germany is associated with the Max Planck Institutes. 99 00:06:15.560 --> 00:06:18.940 Oh, no. And then you you you go there to to to 100 00:06:19.400 --> 00:06:22.760 study towards HR. And what did she ask 101 00:06:22.760 --> 00:06:26.315 you? There is no HR in Max Planck. Yes. It's you. 102 00:06:26.315 --> 00:06:29.995 Believe it or not. It's actually no way. When you went 103 00:06:29.995 --> 00:06:33.215 into industry, I I was curious. How did the interviews go? 104 00:06:33.750 --> 00:06:36.810 In the, in the industry? Yep. 105 00:06:38.230 --> 00:06:41.915 Well, quite well. You know, it took having coming 106 00:06:41.915 --> 00:06:45.215 from Max Planck Institute opens a lot of doors. 107 00:06:45.275 --> 00:06:48.635 Right? And nobody asked any more question whether I'm 108 00:06:48.635 --> 00:06:52.440 qualified to do statistics or machine learning. 109 00:06:53.140 --> 00:06:56.660 This is kinda assumed that you know. The questions that HR 110 00:06:56.660 --> 00:07:00.445 would in the industry ask was, like, how much do you know 111 00:07:00.445 --> 00:07:03.884 about industry specific things? Like, do you have 112 00:07:03.884 --> 00:07:07.505 experience with customers and and things like that, which I didn't. Right? 113 00:07:07.599 --> 00:07:11.199 So I had to learn these type of things. But the 114 00:07:11.199 --> 00:07:14.720 the the competence, they don't question. They just 115 00:07:14.720 --> 00:07:18.491 respect. So that that's how I would 116 00:07:18.907 --> 00:07:22.475 should also give we should 117 00:07:22.475 --> 00:07:26.270 also give a mental shout out to, the the the place you 118 00:07:26.270 --> 00:07:30.030 did your PhD because it's not too far away from where I 119 00:07:30.030 --> 00:07:33.835 studied. They have a very well maintained rivalry with the 120 00:07:33.835 --> 00:07:37.134 university. I attended its University of Oklahoma. 121 00:07:37.755 --> 00:07:41.270 Did you, by the way, eat a lot of Oklahoma 122 00:07:41.490 --> 00:07:44.630 onion burgers? Sorry, guys. The heat is really getting to me here. 123 00:07:45.889 --> 00:07:49.655 No. Well, I did eat some burgers, but 124 00:07:49.655 --> 00:07:53.415 I'm not great fan of burgers, I have 125 00:07:53.415 --> 00:07:56.155 to admit. So, yeah, occasionally, but, 126 00:07:58.439 --> 00:08:02.039 nah. Not too many. Not too 127 00:08:02.039 --> 00:08:05.400 many. Great. So we've been at the place we 128 00:08:05.400 --> 00:08:08.435 shifted into industry roles. 129 00:08:09.535 --> 00:08:13.215 What did you do there? Because I do believe, they would 130 00:08:13.215 --> 00:08:16.790 be they would be very happy to have you in any 131 00:08:16.790 --> 00:08:20.410 team, working on machine learning. 132 00:08:21.705 --> 00:08:25.545 Yeah. Exactly. As a machine learning is one 133 00:08:25.545 --> 00:08:29.245 big topic, and the other big topic is statistics 134 00:08:29.465 --> 00:08:33.250 in general, data analysis. This is what also what industry needs, what 135 00:08:33.250 --> 00:08:36.610 companies need. And, this is 136 00:08:36.610 --> 00:08:40.215 what I could help with, quite a bit. Right? 137 00:08:40.215 --> 00:08:43.815 So, you know, any anything that has to do with lots of 138 00:08:43.815 --> 00:08:46.235 data, trying to understand the data, systematize, 139 00:08:48.110 --> 00:08:51.870 get insight into data. This is where I could help quite a bit. But also 140 00:08:51.870 --> 00:08:55.650 then, of course, machine learning coming up with with, ways 141 00:08:56.035 --> 00:08:59.475 of machine using the data, basically. That's what machine learning 142 00:08:59.475 --> 00:09:03.075 is. Taking lots of data, compressing the 143 00:09:03.075 --> 00:09:06.780 knowledge that this data contain into a machine so that the 144 00:09:06.780 --> 00:09:09.600 machine could make some predictions 145 00:09:10.460 --> 00:09:14.080 or decisions and things like that. Right? So these two things. Right? 146 00:09:14.645 --> 00:09:18.265 I Startupradio, basically, as a data scientist. 147 00:09:18.405 --> 00:09:22.025 That's what the job is, described 148 00:09:22.325 --> 00:09:22.740 today. 149 00:09:27.140 --> 00:09:30.600 I see. And then you got into 150 00:09:31.825 --> 00:09:35.445 several positions. I see research fellow, head of AI, 151 00:09:37.745 --> 00:09:41.285 senior AI consultant, a a lot of different positions 152 00:09:43.170 --> 00:09:46.550 in start ups, in tech companies, also in research 153 00:09:46.770 --> 00:09:50.470 institutes, for example, Frankfurt Institute for Advanced Studies, 154 00:09:51.725 --> 00:09:54.305 all gearing toward, 155 00:09:55.165 --> 00:09:57.585 different areas, but always, 156 00:09:58.845 --> 00:10:02.200 kind of put together by either 157 00:10:02.339 --> 00:10:06.100 either psychology, statistics, or both of 158 00:10:06.100 --> 00:10:09.204 them. And then you started 159 00:10:09.904 --> 00:10:13.745 robots Joe Menninger back in 2016 here in 160 00:10:13.745 --> 00:10:17.500 lovely Frankfurt. But first, how did you get the name? 161 00:10:20.120 --> 00:10:23.785 I was looking for for a cool name. And I there were a 162 00:10:23.785 --> 00:10:27.485 couple of companies that that I I ran into, 163 00:10:28.265 --> 00:10:31.225 and I thought, they have a cool name. I would like to have a company 164 00:10:31.225 --> 00:10:35.010 with such a cool man. Right? One of them was what was 165 00:10:35.010 --> 00:10:37.990 it called? Incredible Machines. 166 00:10:38.904 --> 00:10:41.945 I was like, ah, that's a nice thing. And and, 167 00:10:43.145 --> 00:10:46.530 couple of others, I can't remember now. So I was, like, thinking. I was I 168 00:10:46.530 --> 00:10:50.150 switched on my creative mind for a while, and it took me 169 00:10:50.610 --> 00:10:53.910 a month or so to come up with something 170 00:10:55.330 --> 00:10:58.884 interesting. But, also, you know, the domain had to be available. That's 171 00:10:58.884 --> 00:11:01.464 also another challenge that you have. Right? 172 00:11:03.440 --> 00:11:06.420 And then when I came up with it, I was very happy. 173 00:11:07.760 --> 00:11:11.514 Interestingly, the first person I I told 174 00:11:11.514 --> 00:11:15.214 this idea to didn't get it. It's like, what? I don't get 175 00:11:15.514 --> 00:11:19.194 it. So not everyone finds the name interesting. You do. 176 00:11:19.194 --> 00:11:22.690 Many people do. Maybe, I don't know, 50%, 177 00:11:23.390 --> 00:11:27.070 60%. But other people just some people 178 00:11:27.070 --> 00:11:30.845 just don't get it. So the the the idea. Joe, 179 00:11:30.845 --> 00:11:33.185 you know, it's a, it's a 5050 180 00:11:34.605 --> 00:11:38.370 job. And what do you 181 00:11:38.370 --> 00:11:41.110 guys are actually doing there? 182 00:11:42.450 --> 00:11:46.245 So, we have 183 00:11:46.245 --> 00:11:49.925 developed a new machine learning algorithm, which 184 00:11:49.925 --> 00:11:53.279 was inspired by my understanding 185 00:11:53.579 --> 00:11:56.959 of how the brain works. So if you 186 00:11:58.540 --> 00:12:02.365 compare the neural networks in computers and and neural 187 00:12:02.365 --> 00:12:05.665 networks in in in the brains, human or animal, 188 00:12:05.964 --> 00:12:09.420 there's one big difference. Right? The brains in, 189 00:12:11.560 --> 00:12:15.000 in computer, once they they are in 190 00:12:15.000 --> 00:12:18.845 production, once you do once you train your neural network and you put 191 00:12:18.845 --> 00:12:22.305 it in production, you don't train it anymore. It's not allowed. 192 00:12:22.764 --> 00:12:26.540 And it's a really bad idea. Some people tried, It always 193 00:12:26.540 --> 00:12:29.980 ended up with a disaster. And, 194 00:12:30.940 --> 00:12:34.620 a famous example is when Google put one language model 195 00:12:34.620 --> 00:12:38.145 out and allowed it to learn while people interact with 196 00:12:38.685 --> 00:12:42.445 it. The model very quickly became racist and 197 00:12:42.445 --> 00:12:46.240 had all sorts of bad behavior because people provoked 198 00:12:46.240 --> 00:12:49.380 it. People knew it's gonna learn, so they actually intentionally 199 00:12:50.320 --> 00:12:53.325 tried to gave gave it bad 200 00:12:53.705 --> 00:12:56.605 bad, things because that's what human 201 00:12:57.145 --> 00:13:00.525 do. And and then you Because you operate. 202 00:13:01.850 --> 00:13:05.450 That's what's being human. You provoke everything. You try to 203 00:13:05.450 --> 00:13:08.890 break things. And, and the 204 00:13:08.890 --> 00:13:12.025 model was helpless against 205 00:13:12.405 --> 00:13:13.385 mean humans. 206 00:13:16.325 --> 00:13:20.029 And today, you don't do it. You don't do it anymore. Right? There's even 207 00:13:20.029 --> 00:13:23.569 legally legally forbidden for 208 00:13:23.949 --> 00:13:27.765 autonomous driving machines to have training 209 00:13:27.765 --> 00:13:31.125 while the car is driving for similar reasons. Now 210 00:13:31.125 --> 00:13:34.265 mathematically, there's I can explain to you statistically, 211 00:13:34.485 --> 00:13:34.985 mathematically 212 00:13:41.019 --> 00:13:44.605 Right? Because human brain is not like that. A 213 00:13:44.605 --> 00:13:48.205 human brain can learn. We learn as we 214 00:13:48.205 --> 00:13:51.649 go, and it's not a total disaster. We're not 215 00:13:51.649 --> 00:13:55.250 perfect, but it's not such a disaster. May 216 00:13:55.250 --> 00:13:58.769 I just say the bottom line is you 217 00:13:58.769 --> 00:14:02.365 develop a new machine learning algorithm that 218 00:14:02.365 --> 00:14:06.145 is training itself why it's already in production 219 00:14:06.685 --> 00:14:10.300 without going mental, being 220 00:14:10.300 --> 00:14:12.540 a total dick hat, or, 221 00:14:14.700 --> 00:14:17.340 putting out all crate crazy kind of stuff and, 222 00:14:18.355 --> 00:14:21.495 fantasizing, I do believe, is a call in some models. Right? 223 00:14:24.115 --> 00:14:27.800 Yeah. Yeah. Yeah. Yeah. You you said it right. You 224 00:14:27.800 --> 00:14:31.640 said it right. So, Okay. Can you tell us a little bit 225 00:14:31.640 --> 00:14:35.260 about this? Because I do believe a lot of AI 226 00:14:35.335 --> 00:14:38.855 and machine learning people, c level 227 00:14:38.855 --> 00:14:42.535 executives in different companies, the their ears are now 228 00:14:42.535 --> 00:14:46.350 really perking up here. Okay. Very good. Very 229 00:14:46.350 --> 00:14:48.610 good. They should. So, 230 00:14:49.790 --> 00:14:53.330 basically, the reason why human brain 231 00:14:53.390 --> 00:14:56.872 is not, Joe 232 00:14:56.975 --> 00:15:00.195 susceptible to manipulation and to random inputs 233 00:15:00.975 --> 00:15:04.680 is that we have internal control of what is relevant, 234 00:15:04.740 --> 00:15:08.500 what is not relevant. We can direct our attention towards learning. 235 00:15:08.500 --> 00:15:11.885 We can make decisions. What do we pay attention? We we know what is 236 00:15:12.045 --> 00:15:15.345 important, what is not important. We know where we wanna go. Right? 237 00:15:15.805 --> 00:15:19.360 The classical AI algorithms are not 238 00:15:19.440 --> 00:15:23.200 do not know anything. They are just completely open. Just teach me anything. 239 00:15:23.200 --> 00:15:27.014 I'll take anything. Just give me data. That's practical. For some 240 00:15:27.014 --> 00:15:30.855 purposes, it's a disaster if you have very small number of data 241 00:15:30.855 --> 00:15:33.355 points and if these data points are 242 00:15:34.260 --> 00:15:38.020 somehow not what you wanna learn. Right? So we have 243 00:15:38.020 --> 00:15:41.395 developed a way to guide deep 244 00:15:41.395 --> 00:15:45.175 learning, to guide neural networks. Right? So there is another layer 245 00:15:45.795 --> 00:15:49.420 of parameters, another layer of knowledge, which which 246 00:15:49.420 --> 00:15:52.780 knows what is allowed, what is not allowed, what could be done, 247 00:15:52.780 --> 00:15:56.560 what what what shouldn't be done. Right? 248 00:15:57.535 --> 00:16:01.375 Basically, each parameter in a neural network has its 249 00:16:01.375 --> 00:16:05.215 own parameter telling telling it whether it should be 250 00:16:05.455 --> 00:16:09.280 how much it should be allowed to learn. Right? And the made most 251 00:16:09.280 --> 00:16:12.900 part of the network is not allowed to learn, only some parts. 252 00:16:12.960 --> 00:16:14.740 Right? So 253 00:16:16.665 --> 00:16:20.204 what we have in addition to that, we have an algorithm that 254 00:16:21.065 --> 00:16:24.899 teaches this guiding algorithm. Right? So there are 2 levels of 255 00:16:24.899 --> 00:16:28.339 learning, basically. 1st, you have a phase where you learn how to 256 00:16:28.339 --> 00:16:32.075 learn, how to how to direct your 257 00:16:32.075 --> 00:16:35.675 learning. And after that, you can go in production and 258 00:16:35.675 --> 00:16:38.815 learn safely and know where you are going and being 259 00:16:39.035 --> 00:16:42.850 being directed and not being open to any kind of 260 00:16:42.850 --> 00:16:46.070 random input that you that you receive. Right? 261 00:16:47.170 --> 00:16:50.715 So If if if I put it in my very, 262 00:16:50.715 --> 00:16:54.235 very layman's term, you put a parent in there. 263 00:16:54.235 --> 00:16:57.870 Because I do have a toddler here at home, and if 264 00:16:57.870 --> 00:17:01.490 he's not supervised, he'll do all crazy 265 00:17:01.870 --> 00:17:05.505 kind of stuff. Not to mention, 266 00:17:05.505 --> 00:17:09.105 he has, like, a little shopping cart, in which he 267 00:17:09.105 --> 00:17:12.740 sat backwards brushing his teeth or stuff like this, and this is 268 00:17:12.740 --> 00:17:16.500 just the tip-off the iceberg. So, basically, you develop kind 269 00:17:16.500 --> 00:17:20.040 of a a parental algorithms for 270 00:17:20.385 --> 00:17:24.224 toddler AI. Is it something like that? It's it's 271 00:17:24.305 --> 00:17:27.984 yeah. It's whatever is left in in us after we 272 00:17:27.984 --> 00:17:31.540 have been parented. Right? So, like, you 273 00:17:31.540 --> 00:17:35.380 know, you gotta get parented, but then after that, 274 00:17:35.380 --> 00:17:39.140 you internalize these rules. And you're like, you don't do this. No. No. That's a 275 00:17:39.140 --> 00:17:42.184 no no. Right? So these internalized things 276 00:17:42.885 --> 00:17:46.405 we give we give to the network. Right? Yes. 277 00:17:46.405 --> 00:17:50.110 That's that's correct. Right? I Another another 278 00:17:50.110 --> 00:17:53.790 way before we proceed, would it be possible that you 279 00:17:54.510 --> 00:17:57.865 because a lot of people this is, this is not necessarily 280 00:17:58.165 --> 00:18:01.925 like a coding deep tech podcast per se. We try to 281 00:18:01.925 --> 00:18:05.650 keep it on a very Bonnie level. Could you, before we 282 00:18:05.650 --> 00:18:09.250 proceed, kind of split a little 283 00:18:09.250 --> 00:18:12.790 bit for us what you mean by talk about AI, 284 00:18:13.395 --> 00:18:16.915 machine learning, and deep learning because they're not necessarily all the 285 00:18:16.915 --> 00:18:19.495 same even though people use them interchangeably. 286 00:18:20.835 --> 00:18:24.580 Yeah. Okay. So machine learning 287 00:18:24.580 --> 00:18:28.020 is a discipline on how to make machines learn 288 00:18:28.020 --> 00:18:31.735 something. Right? It's a very general discipline. There's lots of different 289 00:18:31.735 --> 00:18:35.195 algorithms, lots of different ways to make machine learning. 290 00:18:35.815 --> 00:18:39.034 Deep learning is one subset. It's one set of algorithms. 291 00:18:39.769 --> 00:18:43.529 There are many more that are not deep learning, but deep learning 292 00:18:43.529 --> 00:18:47.370 is the most popular one. It everyone 293 00:18:47.370 --> 00:18:51.155 talks about deep learning. And the reason is is 294 00:18:51.375 --> 00:18:55.075 that deep learning is exactly open to learn anything. 295 00:18:55.295 --> 00:18:59.120 And you can expand the number of parameters. You can expand the amount of 296 00:18:59.120 --> 00:19:02.580 knowledge practically infinitely, 297 00:19:05.015 --> 00:19:08.855 and it will still keep learning. It can't just take anything in 298 00:19:08.855 --> 00:19:12.620 in itself. The price is that it requires lots computation, 299 00:19:12.620 --> 00:19:16.380 lots of data, but it can learn anything. That's exactly the problem we needed 300 00:19:16.380 --> 00:19:19.895 to solve. Right? Now AI is 301 00:19:19.895 --> 00:19:23.735 an term where you that you refer 302 00:19:23.735 --> 00:19:27.095 to in a popular way in the 303 00:19:27.095 --> 00:19:30.780 resulting machine that you have built. Joe you you put 304 00:19:30.780 --> 00:19:34.380 some deep learning, some other machine learning algorithms. You put some 305 00:19:34.380 --> 00:19:38.015 other algorithms that are not machine learning and yet are AI 306 00:19:38.015 --> 00:19:41.395 algorithms. You put everything together. You pack it. You make interface. 307 00:19:42.175 --> 00:19:45.155 Now you have an AI. So 308 00:19:47.260 --> 00:19:50.720 that's how I I, see it, how I understand it, 309 00:19:51.419 --> 00:19:55.154 how I use it. Okay. Can you give us 310 00:19:55.154 --> 00:19:58.134 some hints? Because you've you've been talking about, 311 00:19:59.875 --> 00:20:03.409 you've been talking about let let me take one 312 00:20:03.409 --> 00:20:07.090 step back. Let us because a lot of people who are out 313 00:20:07.090 --> 00:20:10.549 there own companies, work for companies, 314 00:20:11.554 --> 00:20:14.774 that would profit from a new AI 315 00:20:15.955 --> 00:20:19.740 teaching methodology algorithm, like you said, when 316 00:20:19.740 --> 00:20:23.580 when they can still learn, when they're in production. The thing is, let 317 00:20:23.580 --> 00:20:26.800 us get a little bit through that by first telling 318 00:20:27.340 --> 00:20:31.165 what you are allowed to learn, what not, and maybe 319 00:20:31.165 --> 00:20:34.785 we drew your knowledge from to do that, 320 00:20:34.925 --> 00:20:38.539 and then we could expand into where you think, 321 00:20:38.539 --> 00:20:41.900 theoretically, this could be applied. Because my understanding 322 00:20:41.900 --> 00:20:45.340 is you develop the algorithm, but it's 323 00:20:45.340 --> 00:20:48.195 not, broadly applied yet. 324 00:20:49.135 --> 00:20:52.975 Yeah. That's correct. Actually, you know, the 325 00:20:52.975 --> 00:20:56.730 challenge we have right now is to find the right use 326 00:20:56.730 --> 00:21:00.250 cases. So it's it's not the 327 00:21:00.250 --> 00:21:03.870 same you know, one step is to create great technology, 328 00:21:04.485 --> 00:21:08.164 and the other step is to find a great use case cases 329 00:21:08.164 --> 00:21:11.880 for the great technology. And you can know really 330 00:21:11.880 --> 00:21:15.560 be sure that you have a great technology, but still it takes time, and 331 00:21:15.560 --> 00:21:18.920 it's still an effort to find the right use 332 00:21:18.920 --> 00:21:22.695 cases, to find the right market fit. And that's our challenge 333 00:21:22.695 --> 00:21:26.455 right now. And and, I think we 334 00:21:26.455 --> 00:21:30.060 have, like, lots of examples in other technologies as well. Like, the first 335 00:21:30.060 --> 00:21:33.900 computers people did not know how they will use it. They knew 336 00:21:33.900 --> 00:21:37.695 they have some amazing machine that can do so many different things. But how 337 00:21:37.695 --> 00:21:41.294 exactly it's gonna be used, they didn't know back then. Right? They didn't know that 338 00:21:41.294 --> 00:21:44.815 that we are gonna use it for to play to to 339 00:21:44.975 --> 00:21:48.799 for word processing or Excel sheets or or it took time 340 00:21:48.799 --> 00:21:52.399 to invest the invent those. I know computer games. I don't know. 341 00:21:52.399 --> 00:21:55.539 But similar happens here with large language models. 342 00:21:56.975 --> 00:22:00.815 We are still searching the right use cases. Although we we have some 343 00:22:01.215 --> 00:22:04.655 we know we have some amazingly powerful machine 344 00:22:04.655 --> 00:22:07.919 there. So the same challenge 345 00:22:09.500 --> 00:22:12.865 robots Go Menninger is facing right now. Find the right use case. 346 00:22:13.425 --> 00:22:17.025 We have identified several potential ones that we are working on. 347 00:22:17.025 --> 00:22:20.785 Right? For example, a car that is adjusting 348 00:22:20.785 --> 00:22:24.480 itself, its visual system constantly while 349 00:22:24.480 --> 00:22:28.320 it's driving. Like, for example, when a when a when you are 350 00:22:28.320 --> 00:22:31.985 driving on a sunny day and suddenly you enter a tunnel. The 351 00:22:31.985 --> 00:22:34.725 the the laws of visual, 352 00:22:36.545 --> 00:22:40.390 shades, the lightning, and so on, it completely changes suddenly. Right? 353 00:22:40.630 --> 00:22:44.150 And what our human brain does, it adjusts. It takes a few 354 00:22:44.150 --> 00:22:47.830 seconds until you can see clearly when you enter it to 355 00:22:47.830 --> 00:22:51.415 know. Right? The the 356 00:22:51.415 --> 00:22:54.555 AI solutions that we have today, they don't adjust. They have no 357 00:22:54.855 --> 00:22:58.615 capability of adjusting. Right? With our solution, it should 358 00:22:58.615 --> 00:23:02.059 be possible. Right? Similar thing with when you when 359 00:23:02.059 --> 00:23:05.899 you have, say, conversation. When machine has a conversation 360 00:23:05.899 --> 00:23:09.065 with humans and has to listen to to to, 361 00:23:09.465 --> 00:23:13.305 people's speech, different people have 362 00:23:13.305 --> 00:23:16.905 different accents. Right? And we adjust to accents. Right? So when I 363 00:23:16.905 --> 00:23:20.480 start talking to you in the morning, like, the first few sentences, you may 364 00:23:20.480 --> 00:23:23.840 not be really it would be hard for you maybe to understand because I have 365 00:23:23.840 --> 00:23:27.545 a unique accent, which comes from my Croatian 366 00:23:27.765 --> 00:23:31.605 language. But after a few minutes, you get adjusted. Oh, doctor, don't 367 00:23:31.605 --> 00:23:35.070 worry. Be before 10 AM, I barely under understand 368 00:23:35.070 --> 00:23:36.850 anybody. I'm not a boring person. 369 00:23:38.750 --> 00:23:42.455 Okay. So, you know, that's another use 370 00:23:42.455 --> 00:23:46.294 case. Right? So we are we are we are 371 00:23:46.294 --> 00:23:49.760 looking for those, and and the challenge is to find the right 372 00:23:50.059 --> 00:23:52.000 market fit. Mhmm. 373 00:23:54.059 --> 00:23:57.360 Can you talk about industries? We just 374 00:23:57.725 --> 00:24:01.404 talked about self driving cars. I do 375 00:24:01.404 --> 00:24:04.945 know there was a lot of AI involved, for example, 376 00:24:05.590 --> 00:24:09.270 during the development of the COVID vaccines to end the 377 00:24:09.270 --> 00:24:12.570 pandemic. Talk about health care, biotech. 378 00:24:13.065 --> 00:24:16.585 Would you think your algorithm could also be 379 00:24:16.585 --> 00:24:20.365 applied there, financial services and so on and so forth? 380 00:24:21.385 --> 00:24:24.549 Yeah. I I I'm sure it can be. 381 00:24:25.570 --> 00:24:29.169 But how exactly? I don't know yet. Right? But what what we 382 00:24:29.895 --> 00:24:33.674 we have a very nice piece of evidence that it the algorithm 383 00:24:33.735 --> 00:24:37.115 can reply there because we have, performed 384 00:24:37.255 --> 00:24:39.195 one project together with NASA, 385 00:24:40.860 --> 00:24:44.380 exactly in this field. And, what we 386 00:24:44.380 --> 00:24:47.680 did, we trained the neural network to learn, 387 00:24:48.965 --> 00:24:51.865 how to deal with RNA sequences. 388 00:24:52.865 --> 00:24:56.105 Joe you basically one one general problem 389 00:24:56.405 --> 00:25:00.200 that is relevant for for biotech industry 390 00:25:00.340 --> 00:25:04.180 is to process RNA information. You have a sequence of 391 00:25:04.180 --> 00:25:07.845 of nucleotides, and you need to work with 392 00:25:07.845 --> 00:25:11.445 that as an AI solution. Now like 393 00:25:11.445 --> 00:25:15.205 every other AI, you need a lot of 394 00:25:15.205 --> 00:25:18.910 data to train anything. But NASA 395 00:25:18.910 --> 00:25:22.450 had 6 mice only in space, 396 00:25:22.590 --> 00:25:25.985 and they couldn't increase the data amount of data because it's just 397 00:25:26.385 --> 00:25:30.145 prohibitively expensive. And they they just can't pay, I don't 398 00:25:30.145 --> 00:25:33.125 know, 100,000,000 to send another 399 00:25:33.620 --> 00:25:37.460 600 mice in space. So they had to find a solution how to train a 400 00:25:37.460 --> 00:25:40.915 model with RNA coming from 6 month 401 00:25:41.315 --> 00:25:44.775 month mice in space, and they had it at the same time as a control 402 00:25:44.915 --> 00:25:48.675 6 mice on the ground. And they 403 00:25:48.675 --> 00:25:52.009 had to train the model with it. So they discovered us. 404 00:25:52.309 --> 00:25:55.909 We worked together, and and it worked. It 405 00:25:55.909 --> 00:25:59.345 worked. So we could use the Earth, a lot 406 00:25:59.345 --> 00:26:02.725 of massive Earth data to 407 00:26:03.105 --> 00:26:06.610 train the model, to learn how to learn. And then after 408 00:26:06.610 --> 00:26:09.650 that, it could deal effectively with just, 409 00:26:10.290 --> 00:26:13.350 6 data points from space. And 410 00:26:14.184 --> 00:26:17.785 and, we wrote this paper recently Joe one can find 411 00:26:17.785 --> 00:26:21.460 it. Of of course. We link it down here in Joe 412 00:26:21.560 --> 00:26:25.086 show notes on our blog post. You'll find 413 00:26:25.086 --> 00:26:28.612 it at stethubreight.oforward/block. And and then I 414 00:26:28.612 --> 00:26:32.275 I was curious. So your model was basically 415 00:26:32.895 --> 00:26:36.335 your algorithm was first applied to teach the other 416 00:26:36.335 --> 00:26:39.650 algorithm how to learn? Well, there's a 417 00:26:40.290 --> 00:26:43.430 yeah. There are 2 algorithms. One algorithm 418 00:26:43.650 --> 00:26:47.445 that, knows how to learn, but this 419 00:26:47.445 --> 00:26:50.885 one has to learn first. So there's the first an algorithm that teaches the second 420 00:26:50.885 --> 00:26:54.645 algorithm how to learn. Right? So there 421 00:26:54.645 --> 00:26:57.840 there is a hierarchy of algorithms. 422 00:26:59.340 --> 00:27:03.100 Like you said, the the parents and the there's a 423 00:27:03.100 --> 00:27:06.934 parent and the The parent algorithm 424 00:27:06.934 --> 00:27:07.434 algorithm. 425 00:27:11.815 --> 00:27:14.955 I see. I see. I see. Very, very interesting. 426 00:27:15.246 --> 00:27:19.000 Joe, everybody who'd like to 427 00:27:19.000 --> 00:27:22.760 learn more about your algorithm, we'll link down here 428 00:27:22.760 --> 00:27:26.515 in the show notes. First, your company's website. Secondly, your 429 00:27:26.515 --> 00:27:30.275 LinkedIn profile. As always, we will be ending with a 430 00:27:30.275 --> 00:27:33.760 few questions. First of them, 431 00:27:34.700 --> 00:27:38.140 are you looking for talented people, HR, new 432 00:27:38.140 --> 00:27:41.745 hires? I'm I'm 433 00:27:41.965 --> 00:27:45.725 looking for talented people all the time. Right? There 434 00:27:45.725 --> 00:27:48.705 is, it's 435 00:27:49.165 --> 00:27:52.940 great pleasure to find talented machine 436 00:27:52.940 --> 00:27:56.240 learning people. They are not easy to find, I have to say. 437 00:27:57.024 --> 00:28:00.644 And every time I, discover someone, 438 00:28:02.304 --> 00:28:05.850 it's a great pleasure to work with them. Right? And if there 439 00:28:05.850 --> 00:28:09.690 are, any, say, PhD 440 00:28:09.690 --> 00:28:13.130 students interested in a project, people who have 441 00:28:13.130 --> 00:28:15.815 finished PhD interested in a great project, 442 00:28:16.675 --> 00:28:20.275 please contact me. We have we 443 00:28:20.275 --> 00:28:23.910 have really interesting problems to solve. Right? 444 00:28:23.910 --> 00:28:27.130 That that are academically interesting, commercially 445 00:28:27.270 --> 00:28:31.035 interesting. I would be interested. You're 446 00:28:31.355 --> 00:28:34.815 likely not looking for the normal coders here. 447 00:28:35.035 --> 00:28:38.795 Are you looking more for people with a statistics, computer science 448 00:28:38.795 --> 00:28:42.060 background, or do you look more for people with a psychology 449 00:28:42.280 --> 00:28:45.260 background? More statistics 450 00:28:46.200 --> 00:28:48.145 slash machine learning, 451 00:28:49.965 --> 00:28:53.665 which includes, of course, a lot of computer science understanding. Right? 452 00:28:53.805 --> 00:28:57.010 That's the is 453 00:28:57.549 --> 00:28:59.090 the the rarest breed. 454 00:29:02.429 --> 00:29:04.210 These are the hardest to find. 455 00:29:08.475 --> 00:29:11.850 That is already pretty good. Would you 456 00:29:11.850 --> 00:29:14.990 guys be open to talk to new investors? 457 00:29:15.929 --> 00:29:19.605 Yes. Definitely. Definitely. Yes. Yep. Yes. 458 00:29:19.924 --> 00:29:23.684 And the last question, this is, supported by Hessen 459 00:29:23.684 --> 00:29:27.065 Trade and Eves and the European Enterprise Network 460 00:29:27.285 --> 00:29:31.130 Hessen. You would now have the chance to 461 00:29:31.130 --> 00:29:34.750 address a question, a concern, an idea for improvement 462 00:29:35.210 --> 00:29:38.835 for the state's Startupradio scene. If you 463 00:29:38.835 --> 00:29:41.975 could address those decision makers, what would it be? 464 00:29:43.960 --> 00:29:47.180 Well, one experience I had 465 00:29:47.480 --> 00:29:49.740 by, you know, starting a company 466 00:29:51.320 --> 00:29:54.915 was that in general in Germany and in Hessen, 467 00:29:55.775 --> 00:29:59.615 it's very hard to start a deep tech company like we are. And deep 468 00:29:59.615 --> 00:30:03.169 tech basically means that you have to 469 00:30:03.169 --> 00:30:06.850 develop a technology before you have a use case, 470 00:30:06.850 --> 00:30:10.610 before you have a market fit, before you have a customer. There's a 471 00:30:10.610 --> 00:30:14.225 lots of work behind that. 472 00:30:15.245 --> 00:30:18.684 And my feeling is that this is this is really hard to 473 00:30:18.684 --> 00:30:22.450 get in Germany. The and that's 474 00:30:22.450 --> 00:30:25.830 probably the reason why we see most of the companies that 475 00:30:26.130 --> 00:30:28.945 come out with such successful 476 00:30:31.404 --> 00:30:35.184 concepts are not coming from Germany, but from United States. 477 00:30:37.450 --> 00:30:39.549 From UK is also much more, 478 00:30:41.290 --> 00:30:45.085 effective in that way. And we somehow, it would 479 00:30:45.085 --> 00:30:48.684 be really great if we found a way that works in 480 00:30:48.684 --> 00:30:52.460 Germany for deep deep tech. Right? It's not just 481 00:30:52.539 --> 00:30:56.299 there is a lots of talk about deep tech, but there's not lots of it's 482 00:30:56.299 --> 00:30:59.440 not coming out really in a great way. We don't 483 00:30:59.659 --> 00:31:03.095 have a list of great 484 00:31:03.155 --> 00:31:06.855 German or Hessian companies 485 00:31:07.155 --> 00:31:10.610 that have actually brought new AI 486 00:31:10.610 --> 00:31:14.210 technologies to the world. Right? Everybody's trying to just pick up 487 00:31:14.210 --> 00:31:18.049 whatever he's using and just go and make a make a business model, which 488 00:31:18.049 --> 00:31:21.675 is great, which is, of course, useful, but it's not 489 00:31:21.675 --> 00:31:25.355 how you become a leader. Right? We need to become a leader. And to become 490 00:31:25.355 --> 00:31:29.030 a leader, you cannot do it with with the system the way how it 491 00:31:29.030 --> 00:31:32.710 works right now. I don't know why what is broken. It's not 492 00:31:32.710 --> 00:31:36.265 only the government. The governments can help a lot. But, 493 00:31:36.265 --> 00:31:39.405 also, I think it's generally mindset. I don't know. 494 00:31:40.025 --> 00:31:43.705 Also, investors are generally not don't have the 495 00:31:43.705 --> 00:31:45.405 stomach for such risks. 496 00:31:47.570 --> 00:31:51.250 They would like to go quickly, you know, have your customer first, and 497 00:31:51.250 --> 00:31:53.514 then we'll talk. And, 498 00:31:55.255 --> 00:31:59.014 fine. Reducing risk is fine. To become 499 00:31:59.014 --> 00:32:02.560 a leader, the world leader, that's not possible that 500 00:32:02.720 --> 00:32:05.760 Right? That's my that's my, 501 00:32:06.560 --> 00:32:10.240 message. Find out a way to 502 00:32:10.240 --> 00:32:14.035 become a leader. World leader help us become a world leader. We wanna 503 00:32:14.035 --> 00:32:17.735 be world leader in the technology, in developing fundamental 504 00:32:17.875 --> 00:32:21.660 technology for AI and help us. 505 00:32:21.660 --> 00:32:25.020 And we don't we did get a small grant from Hassan. Thank you very 506 00:32:25.020 --> 00:32:27.520 much. But we would need 507 00:32:29.154 --> 00:32:32.695 also some some bigger support. 508 00:32:33.315 --> 00:32:36.455 If if not, we'll fight. We'll keep fighting. 509 00:32:38.680 --> 00:32:42.440 Awesome. Danco, it was a pleasure talking to you. Thank 510 00:32:42.440 --> 00:32:46.135 you very much. I really enjoyed the interview. Yeah. 511 00:32:46.135 --> 00:32:49.975 Me too. Thank you. Thank you. Have a good day. 512 00:32:49.975 --> 00:32:51.515 Bye bye. Bye bye.