WEBVTT - Making Computers Think Like Brains 0:00:00.160 --> 0:00:07.240 Brought to you by Toyota. Let's go places. Welcome to 0:00:07.400 --> 0:00:15.280 Forward Thinking Heater, and welcome to Forward Thinking, the podcast 0:00:15.320 --> 0:00:17.520 that looks at the future and says you better think. 0:00:18.120 --> 0:00:21.599 I'm job and Strickland, I'm Lauren Vocaban, and I'm Joe McCormick. 0:00:21.880 --> 0:00:25.360 So hey everybody. Hey, last time we were in the 0:00:25.400 --> 0:00:28.800 podcast studio, which in reality, was a few minutes ago, right, 0:00:28.840 --> 0:00:32.720 we never left Peek behind the curtain, we talked about 0:00:32.760 --> 0:00:35.920 the differences between brains and computers a lot of times, 0:00:35.920 --> 0:00:37.560 a lot of them. Yeah, a lot of times people 0:00:37.600 --> 0:00:40.960 talk about one sort of using the other as a 0:00:41.000 --> 0:00:43.720 metaphor for it, like the brain is like a big computer, 0:00:43.960 --> 0:00:47.400 or computers are thinking. You might say that, and it 0:00:47.440 --> 0:00:49.320 seems like a really good metaphor on the surface, but 0:00:49.360 --> 0:00:50.960 when you get down into the nitty gritty of how 0:00:51.000 --> 0:00:53.960 each one works, they are incredibly different. Yeah. So they 0:00:54.040 --> 0:00:58.520 have some similarities. They both process information, they both store information, 0:00:58.640 --> 0:01:03.240 have memory, they both can perform logic, they both uh, 0:01:03.400 --> 0:01:07.200 they both use electricity. Let's say that there are a 0:01:07.240 --> 0:01:12.840 lot of they have input and output, but they're fundamentally 0:01:12.959 --> 0:01:16.080 different in a lot of other ways. Fundamentally different, and 0:01:16.680 --> 0:01:20.559 especially in the fact that we really understand how computers 0:01:20.600 --> 0:01:23.880 and code information, and we don't fully understand all the 0:01:23.920 --> 0:01:27.120 ins and outs of how information functions in the brain. 0:01:27.480 --> 0:01:31.080 But we do know a lot about things brains are 0:01:31.120 --> 0:01:34.679 good at that computers aren't good at, and we talked 0:01:34.680 --> 0:01:37.240 about some of those in the last podcast. For example, 0:01:37.560 --> 0:01:40.680 brains are more energy efficient than computers for the amount 0:01:40.720 --> 0:01:44.960 of processing they do. Brains are much more versatile than computers. 0:01:44.959 --> 0:01:48.560 So computer might very be able to very quickly execute 0:01:48.880 --> 0:01:54.560 a very specific, pre programmed set of instructions. But brains 0:01:54.600 --> 0:01:56.960 can learn how to do new things, they can adapt 0:01:57.040 --> 0:02:01.360 to new scenarios, they can reprogram them themselves to do 0:02:01.400 --> 0:02:04.880 a task better. Computers can't do any of this without specific, 0:02:05.040 --> 0:02:08.359 you know, augmentation to allow them to do it. So 0:02:09.000 --> 0:02:10.960 we want to talk today about how we can make 0:02:11.000 --> 0:02:16.120 computers more like brains. Yeah, this is not a new idea. 0:02:16.560 --> 0:02:19.919 Um people have been I mean we've we've had discussions 0:02:19.919 --> 0:02:24.119 about this. Even even Touring had some thoughts about how 0:02:24.280 --> 0:02:27.520 a computer could at least simulate intelligence enough for it 0:02:27.560 --> 0:02:30.480 to seem like it was thinking like a person. Now 0:02:30.520 --> 0:02:33.799 in that case, he was really talking about, uh kind 0:02:33.840 --> 0:02:37.040 of a simulation the idea that if it's simulated so 0:02:37.120 --> 0:02:39.920 close to human behavior that we might as well assume 0:02:39.960 --> 0:02:43.560 it has intelligence. We're not necessarily talking intelligence here in 0:02:43.600 --> 0:02:47.160 the sense of a a self aware of thinking machine. 0:02:47.480 --> 0:02:51.880 We're talking about giving a computer some capabilities that are uh, 0:02:52.000 --> 0:02:54.160 that come very easily to brains, but not so easily 0:02:54.200 --> 0:02:57.280 to computers. Yeah. I think the difference between this and 0:02:57.639 --> 0:03:00.000 the kind of artificial intelligence we talked about in the 0:03:00.040 --> 0:03:03.760 Turing test is that's more like the appearance of intelligence. 0:03:04.320 --> 0:03:10.120 This is working like a brain works. So if you were, example, uh, 0:03:10.320 --> 0:03:15.320 teaching a computer how to recognize certain phrases spoken out loud, 0:03:15.760 --> 0:03:18.000 and then you had someone else come in who was 0:03:18.160 --> 0:03:22.600 a different age, gender, and perhaps even a different dialect 0:03:22.760 --> 0:03:25.280 say that same phrase, the computer would still be able 0:03:25.320 --> 0:03:27.960 to pick up on it, even though the phonetics might 0:03:28.000 --> 0:03:31.840 be remarkably different. Yeah. So people were pretty good at this. 0:03:31.960 --> 0:03:34.560 Not depending on where you are, you might if you're 0:03:34.560 --> 0:03:37.440 in Yorkshire and you happen to be from the South, 0:03:37.640 --> 0:03:40.560 you might not understand a word anyone is saying. I 0:03:41.320 --> 0:03:43.920 maintained that no one in Yorkshire understands what anyone else 0:03:43.960 --> 0:03:47.200 is saying. But but if you're a computer, I mean, 0:03:47.280 --> 0:03:51.360 these issues can become difficult even with subtle changes. So 0:03:51.600 --> 0:03:53.880 it's well, it's not just sounds like that. One of 0:03:53.920 --> 0:03:55.680 the other things I should have mentioned at the beginning 0:03:55.720 --> 0:03:58.680 that human brains are very good at that computers are 0:03:58.720 --> 0:04:01.880 not very good at is dealing with real world data. 0:04:02.080 --> 0:04:06.640 So computers do great jobs with data that's like a formatted, 0:04:06.680 --> 0:04:10.080 structured set of number values. They do not do a 0:04:10.120 --> 0:04:13.960 great job with say a visual image or a sound 0:04:14.280 --> 0:04:16.919 or you know, the feeling of a of a touch. 0:04:17.160 --> 0:04:19.960 All that has to be translated into digital language for 0:04:19.960 --> 0:04:23.520 the computer. It just doesn't work intuitively like it does 0:04:23.600 --> 0:04:26.520 for us. It's also really bad at making associations between 0:04:26.600 --> 0:04:30.960 things without you having built in entire ontology to teach 0:04:31.040 --> 0:04:33.320 it what things are in their relationships to each other. 0:04:33.640 --> 0:04:37.960 So making a computer think more or perhaps work more 0:04:38.320 --> 0:04:41.200 like a brain has a lot of a lot of 0:04:41.200 --> 0:04:44.000 of appeal to it. And like I said this, this 0:04:44.040 --> 0:04:45.560 appeal goes back quite a way. It is in fact 0:04:45.640 --> 0:04:47.880 John von Neumann. In our last podcast, we talked about 0:04:47.920 --> 0:04:51.400 von Neuman architecture and the fact that the their your 0:04:51.440 --> 0:04:55.039 basic computer today is based on von Neumann architecture. This 0:04:55.120 --> 0:04:58.520 idea of essentially a CPU and RAM. That's that's your 0:04:58.720 --> 0:05:00.880 basic when you ole it all down, that's what it 0:05:00.880 --> 0:05:04.240 comes down to. Um. He actually was working on a 0:05:04.279 --> 0:05:07.919 book and the book was titled The Computer and the Brain. 0:05:08.120 --> 0:05:11.880 He passed away before finishing it, but even then he 0:05:11.960 --> 0:05:16.040 was sort of looking at the differences between computers and 0:05:16.080 --> 0:05:19.200 brains and how could you perhaps start to close the 0:05:19.240 --> 0:05:24.200 gap between the two, Uh in which way? Well, and 0:05:24.720 --> 0:05:27.599 in the way of making computers more like brains, not brains, 0:05:27.640 --> 0:05:31.440 more like computers. He wasn't suggested, He wasn't suggesting that 0:05:31.480 --> 0:05:34.600 we all have positronic brains installed in our heads. Um, 0:05:34.680 --> 0:05:36.400 some people would like that world. I just thought we 0:05:36.440 --> 0:05:39.640 should be clear. Yeah, that's fair, it's fair. The clarification 0:05:39.680 --> 0:05:43.239 has been made. Then. We also have a professor Carver 0:05:43.400 --> 0:05:46.200 Mead of the California Institute of Technology who began to 0:05:46.240 --> 0:05:49.960 experiment with designing computer chips that were inspired by brains, 0:05:50.000 --> 0:05:52.960 and he called them neuromorphic chips. Will be talking a 0:05:53.000 --> 0:05:55.960 lot about those today. And one of his collection of 0:05:55.960 --> 0:05:59.479 transistors mimicked how our brains processed visual data, and it 0:05:59.480 --> 0:06:01.760 actually allowed them to create a computer that could recognize 0:06:01.800 --> 0:06:04.720 the edges of objects in an image. So being able 0:06:04.760 --> 0:06:09.360 to tell where one object ends and something else begins, 0:06:09.360 --> 0:06:12.479 whether that's just space or another object or whatever, which 0:06:12.520 --> 0:06:15.799 again very easy for humans to do, not necessarily easy 0:06:15.839 --> 0:06:17.400 for a machine to do. You know, you have to 0:06:17.400 --> 0:06:19.480 figure out how to how how do you have the 0:06:19.520 --> 0:06:22.960 machine identify that actually very difficult for machines to do. 0:06:23.600 --> 0:06:26.960 Uh So, his work emulated brains but didn't simulate them, 0:06:27.120 --> 0:06:30.000 and he also admitted his work was much harder than 0:06:30.040 --> 0:06:34.200 he anticipated and that achieving a brain like computer system 0:06:34.240 --> 0:06:36.640 was a non trivial challenge that was going to require 0:06:37.520 --> 0:06:41.720 a lot of innovation because you could brute force it 0:06:42.120 --> 0:06:43.960 and as in other words, you could you could use 0:06:44.000 --> 0:06:50.200 a lot of classic processors to carry that heavy load, 0:06:50.560 --> 0:06:53.200 but that defeats the purpose. The whole purpose of making 0:06:53.560 --> 0:06:56.920 a brain like computer is so that you can take 0:06:56.960 --> 0:07:01.200 advantage of the brain's properties, including that I'm amazing. Energy 0:07:01.200 --> 0:07:03.800 efficiency doesn't be any good if you are able to 0:07:03.839 --> 0:07:06.840 brute force your way into behaving kind of like a brain. 0:07:06.920 --> 0:07:11.400 But you're, yeah, yeah, you're you're. Every time you turn 0:07:11.440 --> 0:07:13.480 it on, all the lights in the neighboring houses go 0:07:13.680 --> 0:07:16.360 womb where they fade out for a moment in the cup. 0:07:16.800 --> 0:07:20.960 Some crunching numbers. Again, that's not useful, So that would 0:07:21.000 --> 0:07:23.440 be uh, you know, that was in the nineteen eighties 0:07:23.600 --> 0:07:27.000 and there were some like I said, some success skipping 0:07:27.000 --> 0:07:29.640 way ahead. Uh, there are a lot a lot more 0:07:29.640 --> 0:07:32.520 work has been done in the fairly recent past. Back 0:07:32.560 --> 0:07:35.520 in two thousand eight, DARPA announced the Synapse program with 0:07:35.560 --> 0:07:37.880 the goal to break free from von Neuman architecture and 0:07:37.920 --> 0:07:41.200 create a new model of computer designs. So this is 0:07:41.880 --> 0:07:45.800 about as dramatic a departure as you can imagine for 0:07:45.840 --> 0:07:49.200 a computer science You're talking about not just getting away 0:07:49.200 --> 0:07:52.640 from that CPU and RAM model, but possibly getting away 0:07:52.640 --> 0:07:57.840 from binary data entirely. So that's I mean, that's that's 0:07:58.000 --> 0:08:02.200 so different from what we've been doinging that it's uh, 0:08:02.320 --> 0:08:06.200 it's it's completely revolutionary. I mean there's really hard to 0:08:06.360 --> 0:08:08.800 it's hard to imagine it just because we've been relying 0:08:08.800 --> 0:08:11.480 on computer so long that are stuck on this binary 0:08:11.520 --> 0:08:16.160 system binary just zeros and ones or off and on switches, 0:08:16.200 --> 0:08:18.440 if you want to think of it that way. Um, 0:08:18.480 --> 0:08:20.680 So in a broader category, you could just think about 0:08:20.720 --> 0:08:23.600 it not in terms of binary but just digital data 0:08:23.640 --> 0:08:27.040 of any kind, right, Well, I mean with binary specifically 0:08:27.080 --> 0:08:32.920 you're talking about your basic unit has two states, right, 0:08:32.960 --> 0:08:35.040 That's that's what you're limited to. You can't have any 0:08:35.040 --> 0:08:38.120 more than two states. And that's one thing that Darka 0:08:38.200 --> 0:08:39.720 was saying. Let's try and get away from that. Let's 0:08:39.720 --> 0:08:42.360 look at a way of doing computation where we're not 0:08:42.440 --> 0:08:45.600 limited by this too state of your basic unit, which 0:08:45.640 --> 0:08:51.000 is incredible. I mean, it requires a total different approach. Now, 0:08:51.040 --> 0:08:54.600 two thousand eleven, IBM unveiled at the first synapse chips 0:08:54.640 --> 0:08:58.560 that could play Pong. Well, come on, now, they couldn't 0:08:58.600 --> 0:09:02.760 just play, actually learn how to play. Right, So you 0:09:02.840 --> 0:09:06.400 sit your little brother a little little sister down at 0:09:06.400 --> 0:09:09.280 a video game, and because you're mean, you don't tell 0:09:09.320 --> 0:09:11.600 them how to play, and you just keep beating them 0:09:11.600 --> 0:09:15.560 over and over at Mortal Kombat or whatever. Eventually they 0:09:15.600 --> 0:09:17.760 figure out how to play. They might not be real good, 0:09:17.800 --> 0:09:20.320 but they figure out what the buttons do and how 0:09:20.360 --> 0:09:22.600 to win, and because they're younger than you are, they've 0:09:22.640 --> 0:09:25.440 got amazing twitch skills and eventually they dominate you in 0:09:25.440 --> 0:09:28.440 every game. And then you just don't want to play. 0:09:29.160 --> 0:09:32.840 That's the killer instinct, right, Yeah, that's what this computer 0:09:32.920 --> 0:09:35.199 did at Pong. They sat it down they were mean 0:09:35.240 --> 0:09:37.360 to it. They didn't tell it how to play. They 0:09:37.440 --> 0:09:39.520 just let it know when it was doing good or 0:09:39.520 --> 0:09:42.040 doing bad. Right if it if it blocked the ball, 0:09:42.080 --> 0:09:43.960 then it got essentially a little bit of a reward 0:09:44.080 --> 0:09:46.720 like good job. And if it let the ball go 0:09:46.920 --> 0:09:50.440 by and uh and fly past and the opponent got 0:09:50.559 --> 0:09:53.360 a point, essentially got a message saying, you're not doing 0:09:53.400 --> 0:09:56.640 so well. And so it began to learn that what 0:09:56.640 --> 0:09:58.640 it was supposed to do was block the ball from 0:09:58.640 --> 0:10:00.800 going in, and so it got better and better and 0:10:00.840 --> 0:10:03.440 better at doing that. So in in human terms that 0:10:03.520 --> 0:10:05.719 sounds kind of trivial because we can all do that, 0:10:05.800 --> 0:10:08.839 but in a computer that's a big deal. The idea 0:10:08.880 --> 0:10:12.840 of learning computers could very well revolutionize the kinds of 0:10:12.920 --> 0:10:16.720 jobs computers can do. And another example they had, which 0:10:16.840 --> 0:10:19.760 you know is less trivial than learning to play pong, 0:10:20.360 --> 0:10:23.600 is they had this these kind of chips as a 0:10:23.640 --> 0:10:26.840 guidance system, part of a guidance system for a flying 0:10:26.920 --> 0:10:30.920 drone that could just follow a yellow line, specifically yellow 0:10:30.920 --> 0:10:33.319 lines on a road, and so it could actually follow 0:10:33.360 --> 0:10:36.960 along a road and stay on track because it new 0:10:37.000 --> 0:10:39.520 to recognize this particular yellow line and that that was 0:10:39.600 --> 0:10:42.440 the course it was supposed to follow. So it was 0:10:42.480 --> 0:10:47.640 able to, you know, observe an environment and identify the 0:10:47.679 --> 0:10:50.680 important feature of that environment and follow that feature, which 0:10:50.720 --> 0:10:53.240 again something very easy for humans to do. You tell 0:10:53.240 --> 0:10:55.920 a human as long as they have their senses, and 0:10:55.960 --> 0:10:58.440 you tell human, hey, you need to do when when 0:10:58.520 --> 0:11:01.600 you see or experience this, you need to do this. 0:11:02.320 --> 0:11:06.160 It's very easy for a machine, not necessarily that easy. Well, 0:11:06.559 --> 0:11:10.160 like us in navigating the real world. Computers like these 0:11:10.160 --> 0:11:12.520 are learning, not just being told what to do, but 0:11:12.679 --> 0:11:15.280 figuring out what it is they need to do. Yeah. 0:11:15.840 --> 0:11:19.560 So two thousand thirteen, so very recent past, HRL built 0:11:19.600 --> 0:11:23.319 a chip that can alter the connections between neurons synthetic 0:11:23.360 --> 0:11:26.480 neurons in this uh in this chip, and it's similar 0:11:26.520 --> 0:11:28.679 to the way the brain can change synapses. So it 0:11:28.800 --> 0:11:32.000 learns quote unquote like a brain. It's able to make 0:11:32.080 --> 0:11:37.120 these connections that are emulating what happens within a human brain. Now, 0:11:37.200 --> 0:11:40.720 keep in mind, as we said before, we still don't 0:11:40.800 --> 0:11:44.679 fully understand all the complexity that is the human brain 0:11:44.720 --> 0:11:47.520 and the human nervous system and and how we think. 0:11:48.000 --> 0:11:51.240 We're still learning quite a bit about that. But by 0:11:51.320 --> 0:11:54.720 emulating some of the physical processes. We're starting to see 0:11:54.800 --> 0:11:58.800 some improvements and things like energy efficiency for specific types 0:11:58.880 --> 0:12:02.760 of computer tasks. Yeah. Well, there are also very large 0:12:02.880 --> 0:12:06.760 scale adaptations that are being made in computers to make 0:12:06.760 --> 0:12:10.160 them more like brains. For example, you've got neural networks. 0:12:10.280 --> 0:12:13.680 You're familiar with this concept. Yeah, yeah, So in your 0:12:13.679 --> 0:12:16.920 brain you have a neural network. You have a network 0:12:16.960 --> 0:12:20.400 of neurons, and they're these nerve cells that are connected 0:12:20.440 --> 0:12:23.560 in tons of different ways all throughout your brain, sending 0:12:23.640 --> 0:12:27.920 signals back and forth to each other. And somehow this process, uh, 0:12:27.960 --> 0:12:30.679 that's a very simple representation of it. There there are 0:12:30.679 --> 0:12:33.800 other kinds of cells and things going on to somehow 0:12:33.840 --> 0:12:39.160 this whole process creates thinking, It manifests itself and thought right. 0:12:39.880 --> 0:12:43.600 And artificial neural network tries to use some of the 0:12:43.640 --> 0:12:47.880 same basic concepts to make a computer or network of 0:12:47.920 --> 0:12:52.000 computers work more like a brain. So, neural networks are 0:12:52.400 --> 0:12:56.600 systems of interconnected nodes that simulate Each node simulates the 0:12:56.640 --> 0:12:59.960 behavior of a neuron. In basic terms, neurons in the 0:13:00.040 --> 0:13:03.480 brain receive a signal, and if the signal is strong enough, 0:13:03.559 --> 0:13:07.680 they pass that signal along to other neurons. Artificial neuron 0:13:07.760 --> 0:13:12.839 nodes receive quote weighted signals, and they received them from 0:13:12.920 --> 0:13:16.520 lots of other nodes, and then the waiting system determines 0:13:16.720 --> 0:13:19.640 whether or not that signal gets passed along to the 0:13:19.679 --> 0:13:23.079 next layer of nodes or to a final output node. 0:13:24.440 --> 0:13:27.640 And so these networks are very good at things like 0:13:27.880 --> 0:13:31.880 pattern recognition and learning how to deal with new types 0:13:31.920 --> 0:13:35.000 of data sets, and when spread out across lots of 0:13:35.120 --> 0:13:39.160 different processors, artificial neural networks are very good at processing 0:13:39.559 --> 0:13:44.160 huge amounts of data very quickly. Yeah. So an example 0:13:44.200 --> 0:13:46.960 of this is Google Brain. You may have heard of that. 0:13:47.480 --> 0:13:51.640 That was a hush hush Google x project, right, yeah, 0:13:51.720 --> 0:13:54.840 one of many, uh neural network approach just like you 0:13:54.880 --> 0:13:57.880 were saying. So it uses thousands of computers to mimic 0:13:57.880 --> 0:14:00.280 the synaptic connections in our brains and rely is on 0:14:00.320 --> 0:14:04.520 simple sets of rules also called algorithms, right, So it's 0:14:04.559 --> 0:14:07.160 it's relying on that to learn on its own. So 0:14:07.200 --> 0:14:10.680 the example everyone loves to mention, you'll see it all 0:14:10.720 --> 0:14:15.520 over the internet because it's the Internet, is that it 0:14:15.600 --> 0:14:19.560 learned what a cat was by scanning thousands of pictures 0:14:19.600 --> 0:14:22.360 of cats. Now it had been told the term cat, 0:14:22.960 --> 0:14:25.920 it had to be it did not realize it didn't 0:14:26.000 --> 0:14:29.560 see tens of thousands of pictures of cats and is 0:14:29.600 --> 0:14:31.840 a cat like it could have just as easily said 0:14:31.880 --> 0:14:36.440 that is a bucket. That's a different meme entirely. Yeah, no, 0:14:38.400 --> 0:14:41.680 walrus is Yeah. No. The amazing thing about it is 0:14:41.760 --> 0:14:45.680 that that it was able to isolate the visual concept 0:14:45.760 --> 0:14:49.080 of a cat and then be able to identify a cat, 0:14:49.200 --> 0:14:50.920 even if it was a cat that was not among 0:14:50.960 --> 0:14:53.720 the pictures that already scanned, a totally new cat in 0:14:53.760 --> 0:14:56.000 a in a video or a picture exactly so you 0:14:56.000 --> 0:14:57.920 can show it a picture of a cat and had 0:14:58.000 --> 0:15:00.560 never seen before, but because it now knew what the 0:15:00.720 --> 0:15:04.120 concept of a cat was, it could identify that as 0:15:04.160 --> 0:15:07.120 a cat. Yeah. One of the people behind this project 0:15:07.720 --> 0:15:11.280 was gave a quote saying that it basically invented the 0:15:11.320 --> 0:15:14.160 concept of a cat, which is kind of true. Like 0:15:14.240 --> 0:15:17.080 it it had to just I thought of it, and 0:15:17.320 --> 0:15:19.120 I wrote this in the video. Don't know whether you 0:15:19.160 --> 0:15:20.960 said these words or not, but I thought about it 0:15:21.560 --> 0:15:23.840 sort of in the in the terms of like a 0:15:23.920 --> 0:15:28.720 naturalist surveying new animal or plant species, you would you 0:15:28.720 --> 0:15:32.960 would look for patterns, and you're not being told what 0:15:33.080 --> 0:15:36.920 new species to look for. But by observing enough of them, 0:15:37.000 --> 0:15:41.040 you'd say, Okay, I'm seeing all these similarities with slight variations, 0:15:41.040 --> 0:15:45.120 this must be a species. And here's another species that's 0:15:45.200 --> 0:15:48.520 kind of the same but different. And you're creating these 0:15:48.560 --> 0:15:51.960 categories in your mind without direction from the top. You're 0:15:51.960 --> 0:15:54.720 just taking in lots of data and figuring out what 0:15:54.800 --> 0:15:56.840 to do with it. And that's what these neural networks 0:15:56.840 --> 0:16:00.720 are really good at. So now what they're not necessarily 0:16:00.720 --> 0:16:03.880 good at as being super energy efficient, because we're talking 0:16:03.920 --> 0:16:09.000 thousands of computers. Was sixteen thousand computers to figure out 0:16:09.000 --> 0:16:14.680 what a cat was something like it was, So we're 0:16:14.680 --> 0:16:18.400 talking about a very energy hungry approach here. Because each 0:16:18.440 --> 0:16:21.880 computer is behaving as a neuron. It's not like we're 0:16:21.960 --> 0:16:25.520 we've you know, distilled the neuron down to an element 0:16:25.560 --> 0:16:28.480 that's on a transistor. No, we're talking about individual computers 0:16:28.480 --> 0:16:31.560 acting as a neuron. So you can also simulate them 0:16:31.560 --> 0:16:34.600 within a computer, but this is far less powerful. Right, 0:16:34.640 --> 0:16:37.920 if you simulate within a computer, you better have packed 0:16:38.040 --> 0:16:42.560 a lunch and for the next like several months. No, 0:16:42.640 --> 0:16:46.400 it's incredibly slow simulating these things, making virtual brains, which 0:16:46.400 --> 0:16:49.120 people have tried to do even on a small scale 0:16:49.120 --> 0:16:52.960 requires an enormous amount of processing power, because again, that 0:16:53.320 --> 0:16:57.680 simulation is resting upon a classical computer approach. It's not 0:16:58.400 --> 0:17:01.760 something that's broken free of that Von Neumann architecture. It's 0:17:01.800 --> 0:17:04.199 just resting on top of it. Um. One of the 0:17:04.200 --> 0:17:07.000 guys working on, or at least working with some of 0:17:07.000 --> 0:17:09.520 the people who are working on Google Brain is our 0:17:09.560 --> 0:17:15.919 friend R. Kurtzwell Singularity Dude. Yeah, so, yeah, a pioneer 0:17:15.960 --> 0:17:18.879 in the in the field of voice recognition software as 0:17:18.920 --> 0:17:23.760 well as other artificial intelligence developments, and he's working with Google, 0:17:24.520 --> 0:17:28.440 not necessarily on Google Brain. At least the interview I read, 0:17:28.520 --> 0:17:30.720 it seemed to me that he was working, you know, 0:17:30.800 --> 0:17:33.000 some of his colleagues were working on Google Brain, and 0:17:33.000 --> 0:17:35.200 perhaps he had worked on some things that are are 0:17:35.240 --> 0:17:38.320 related to it. His natural language work is something that 0:17:38.359 --> 0:17:41.880 could be very useful in UH and certain approaches with 0:17:42.119 --> 0:17:45.560 a neural network, so you know, it's related to but 0:17:45.640 --> 0:17:49.439 not necessarily directly overlapping the project. Another project that I 0:17:49.480 --> 0:17:52.359 wanted to mention along these lines is an open source 0:17:52.400 --> 0:17:56.359 one called NEST, which stands for Neural Simulation Technology that 0:17:56.400 --> 0:17:58.720 started coming together in the mid nineteen nineties and and 0:17:58.800 --> 0:18:02.159 studies large at works like the brain and and tries 0:18:02.240 --> 0:18:05.800 to write flexible emulators for the brain for for general 0:18:05.800 --> 0:18:08.159 research use. It's something that I ran across while I 0:18:08.160 --> 0:18:09.879 was in the rabbit hole of all of this research, 0:18:09.920 --> 0:18:11.840 and I thought that it was really neat that it's 0:18:11.880 --> 0:18:14.840 you know, we're working towards giving researchers better tools for 0:18:14.920 --> 0:18:18.399 studying this kind of stuff. That's pretty cool, totally. Okay, 0:18:18.480 --> 0:18:22.200 let's let's zoom in a little bit, because we were 0:18:22.240 --> 0:18:26.240 talking about these sort of artificial constructs like neural artificial 0:18:26.280 --> 0:18:28.719 neural networks and you know, take place plus across a 0:18:28.720 --> 0:18:32.320 lot of computers or in some virtual simulation inside a computer. 0:18:33.240 --> 0:18:37.200 Can we make the actual hardware in your device work 0:18:37.400 --> 0:18:41.040 physically more like a brain does? That is the million 0:18:41.040 --> 0:18:43.000 dollar question, and in fact is the thing that I 0:18:43.040 --> 0:18:46.320 think is the most important if you want to capture 0:18:46.359 --> 0:18:49.440 the energy efficiency angle here, right, if it works more 0:18:49.480 --> 0:18:52.200 like a brain, then it's going to require far less 0:18:52.200 --> 0:18:55.920 power than trying to simulate what a brain does using 0:18:55.920 --> 0:18:58.879 this other architecture. Well, based on some really cutting edge research, 0:18:58.920 --> 0:19:01.280 I think the answer is, yeah, overall, there's there's a 0:19:01.320 --> 0:19:04.439 project at IBM right now. Yeah, the neurosynaptic chips. We 0:19:04.560 --> 0:19:07.720 mentioned that just the beginning in the timeline, But we're 0:19:07.760 --> 0:19:11.320 looking at things like collections of six thousand transistors copying 0:19:11.320 --> 0:19:14.400 the electrical spiking behavior of a neuron. So six thousand 0:19:14.400 --> 0:19:17.879 transistors is nothing. I mean, you look at a micro 0:19:18.080 --> 0:19:20.639 processor right now, and there could be more than a 0:19:20.680 --> 0:19:25.840 billion transistors on that single one inch silicon chip. But 0:19:26.000 --> 0:19:30.359 usually transistors are not modeling spiking behavior. No they aren't. 0:19:30.480 --> 0:19:32.439 But what I'm the only point I'm making is that 0:19:32.600 --> 0:19:35.600 we're talking much smaller now. Before before we were talking 0:19:35.680 --> 0:19:39.840 massive networks of individual computers. Now we're talking about discrete 0:19:39.840 --> 0:19:42.600 elements that you would find within a computer that are 0:19:42.680 --> 0:19:45.159 mimicking the behavior of a neuron specifically because of the 0:19:45.200 --> 0:19:47.640 way they've been engineered. It's not like you know, you've 0:19:47.680 --> 0:19:49.960 hit a little switch like, oh, now, the transistor things 0:19:49.960 --> 0:19:52.400 like a brain. No, it's because it's been specifically engineered 0:19:52.440 --> 0:19:54.920 that way. So yeah, it copies this electrical spiking behavior, 0:19:55.359 --> 0:19:58.320 and the collection of transistors are wired together to make 0:19:58.359 --> 0:20:02.720 a system of synthetic rons. So using these chips was 0:20:02.880 --> 0:20:05.800 a little complicated. When they first showed them off. The 0:20:05.840 --> 0:20:09.360 team had to simulate a physical chip with a virtual 0:20:09.680 --> 0:20:13.240 chip inside a computer program then port that configuration over 0:20:13.359 --> 0:20:16.000 onto the physical hardware that they had built, and that 0:20:16.080 --> 0:20:18.800 chip allowed programmers to create protocols so the computer could 0:20:18.880 --> 0:20:22.639 learn how to recognize handwritten numbers or play pong. This 0:20:22.720 --> 0:20:24.959 is the pong playing chip we talked about at the 0:20:25.000 --> 0:20:27.480 top of the podcast. So, but they could also learn 0:20:27.600 --> 0:20:31.480 like if if you show a computer, like you've got 0:20:31.520 --> 0:20:34.360 some optical recognition software, this is something else Kurts while 0:20:34.440 --> 0:20:38.080 worked on optical recognition software in your computer, and you 0:20:38.119 --> 0:20:41.159 show it a picture of a six, the number six 0:20:41.840 --> 0:20:45.640 that has been printed in a specific font, and then 0:20:45.680 --> 0:20:48.960 show it another number six that either someone has handwritten 0:20:49.440 --> 0:20:51.480 or it has been printed in a different font. The 0:20:51.480 --> 0:20:54.760 computer may have trouble identifying it because it's not exactly 0:20:54.840 --> 0:20:57.040 what you showed it the first time. But let's say 0:20:57.080 --> 0:21:00.359 it really really wants to leave a spam com in 0:21:00.400 --> 0:21:03.080 the comments section on your article. How can it get 0:21:03.119 --> 0:21:07.200 past this problem? Well, using this approach with neurosynaptic chips, 0:21:07.200 --> 0:21:09.680 it could really help because, like I said, they were 0:21:09.680 --> 0:21:13.800 able to use this this network to create a program 0:21:13.800 --> 0:21:17.680 that was able to recognize digits that had been drawn, 0:21:17.840 --> 0:21:21.760 even if they had totally different people drawing them in 0:21:21.800 --> 0:21:25.560 different styles, they were still able to have this machine 0:21:25.640 --> 0:21:28.640 recognize what that number was. So really, when you look 0:21:28.680 --> 0:21:33.199 at it, this particular type of machine was end end 0:21:33.240 --> 0:21:36.360 of which, by the way, the people who created Capture 0:21:36.840 --> 0:21:39.679 are pretty much okay with because they said that the 0:21:40.880 --> 0:21:42.480 yeah and only it was a stop gap. It was 0:21:42.520 --> 0:21:46.320 something that pushed forward the development of artificial intelligence because 0:21:46.359 --> 0:21:48.879 they said, we had to create a thing that was 0:21:49.000 --> 0:21:52.080 difficult for machines to do but easy for humans to do. 0:21:52.720 --> 0:21:55.720 And really, what that does is it points out, hey, 0:21:55.880 --> 0:21:59.160 machines are really bad at this one particular task. How 0:21:59.200 --> 0:22:02.400 can we make sheens better at doing that? And so 0:22:02.560 --> 0:22:05.919 while it's kind of irritating for anyone who wants to 0:22:05.960 --> 0:22:08.040 have controls on their site so they don't get hit 0:22:08.080 --> 0:22:12.119 by spam content, uh, it is very reassuring for anyone 0:22:12.119 --> 0:22:14.840 in the artificial intelligence field because it shows that progress 0:22:14.880 --> 0:22:18.040 is being made. So double edged sword, I guess we 0:22:18.080 --> 0:22:21.679 could say, but at any rate, Uh. The whole idea 0:22:21.720 --> 0:22:23.600 was to emulate the brain's ability to make lots of 0:22:23.640 --> 0:22:26.520 those synaptic connections at once, and the current state of 0:22:26.560 --> 0:22:30.359 the art involves organizing neuro synaptic cores onto a grid 0:22:30.800 --> 0:22:36.000 to create a synthetic cortex. So IBMS neuromorphic programming architecture 0:22:36.040 --> 0:22:40.160 is based off modular blocks of code called core lets, 0:22:40.200 --> 0:22:43.000 and programmers can choose core lets from a predetermined menu 0:22:43.080 --> 0:22:45.320