ColdFusion - Google Duplex A.I. - How Does it Work?

Starting point is 00:00:00 You are watching Cold Fusion TV. Hi, welcome to another Cold Fusion video. So I'm sure by this stage, most of you guys have heard of Google Duplex. I was actually in Bulgaria when the news broke about this, so I quickly had to make my way back to Australia to make this video. So if I'm out of breath, that's why. Okay, so in this video, we'll take a deeper look at Google Duplex. So what is Duplex?

Starting point is 00:00:30 It's basically an extension of Google Assistant, of Google Assistant that can make phone calls to real humans just by you asking it to do so. It's a deep neural network that builds off WaveNet technology. WaveNet is a speech synthesis program that worked by joining very short units of sound together to create speech. It was a breakthrough in natural speech synthesis when it came out. We've already taken a look at it in a previous video, but here's a quick recap of what it can do. Aspects of the sublime in English poetry and painting 1770 to 1850 Aspects of the sublime in English poetry and painting 1770 to 1850

Starting point is 00:01:07 This is what happened when you don't type in anything for Wavenet to say It's still generating raw audio to randomly imitate human sounds and it still sounds like words Beel-a-la-n-boo He adds the hatching with going to the several of pain So it just tie pictures and tell. So it's exciting. Do you have a very shared about it to be given as well. Duplex is another neural network built on top of WaveNet.

Starting point is 00:01:37 The final result is an AI that can have realistic conversation, but with WaveNet vocal precision. And here's the grand unveiling of Duplex at Google's IO event in May of 2018. But even in the US, 60% of small businesses don't have an online booking system set up. So what you're going to hear is the Google Assistant actually calling a real salon to schedule the appointment for you. Let's listen. A woman's haircut for a client. I'm looking for something on May 3rd. Sure.

Starting point is 00:02:16 Give me one second. Mm-hmm. Sure. What time are you looking for a while? At 12 p.m. Okay. We have at 10 o'clock. 10 a.m. is fine.

Starting point is 00:02:31 Perfect. So I will see Lisa at 10 o'clock. May 3rd. Okay, great. Thanks. Great. Have a great day. Bye. Hi. I'd like to reserve a table for Wednesday the 7. For seven people? Um, it's for four people. Four people when... Next Wednesday at 6 p.m.

Starting point is 00:03:05 Oh, actually we live here for like upro like a five people. For people you can come. How long is the wait usually to uh... usually to be seated. When tomorrow or week A or? Four next Wednesday. The seven. No, it's not too busy.

Starting point is 00:03:26 You can count for people, okay? Oh, I got you. Thanks. Bye bye. But the assistant understands the context, the nuance. It knew to ask for wait times in this case and handle the interaction gracefully. According to Google's blog, the ums and art that you hear are put in sometimes synthetically but actually it sometimes is there to signal that the system is still processing just like a human would the public reaction to this was on the

Starting point is 00:03:57 side of shock and horror and even anger by some some people thought that it was very deceitful to have an AI talk to someone over the phone without them knowing but Google has made it very clear that they're going to be transparent they'll be letting the people on the other end of the phone know that they're talking to duplex duplex has been trained in the narrow field or scheduling appointments or bookings and inquiring about a business as opening hours on holidays. To be clear, Duplex cannot have general conversations, but I have little doubt that the scope will widen in the coming years.

Starting point is 00:04:28 Google stated that there were unique challenges when it came to training such a neural net. How do you get an AI to robustly understand natural language and reply in a realistic manner? This would be pretty hard to do. For example, people tend to talk differently to one another than with computers. We talk faster, correct ourselves with conversation,

Starting point is 00:04:46 and even emit parts of conversation and rely on context instead. Throw in the poor quality and noisiness of a phone line and you have a pretty hard challenge on your hands. To solidify this point a little bit, let's think about the phrase, okay for four. It's such a simple sentence, but it relies on many previous sentences for context. This phrase could refer to a time or an amount of people. We as humans take such things for granted, but it's interesting to take a brand new look at this through the eyes of an AI or a team of researchers trying to solve this problem. Okay, so how does Duplex work?

Starting point is 00:05:18 Duplex uses something called a recurrent neural network. If you don't know what a neural network is, it's basically a massive matrix multiplication function where each part of the matrix is built up of artificial neurons called nodes. The nodes contain a mathematical formula and are arranged in layers, and each node has an input and an output. After receiving the inputs, whatever they may be, the end goal of the whole entire matrix or neural network

Starting point is 00:05:45 is basically to find out how to reduce how wrong it is, or in other words, reduce the amount of error. Perhaps the strangest thing about neural networks is that no one actually knows how they come to their conclusion. You just give them the inputs and they somehow get an answer. Neural networks have been around for a while. In fact, Ted Hoff, the guy who helped create the very first CPU at Intel back in 1971, actually worked on neural networks in his early career.

Starting point is 00:06:10 But practical neural networks have only been possible in the past five years or so. Since about 2012, the general complexity of neural networks has advanced 500 times and I think it's one of the most fascinating fields of computer science. If you want a more detailed explanation of how neural networks work, there's an absolutely brilliant video by the YouTube channel 3 Blue 1 Brown. It's a remarkable explanation. I'll leave a link in the description below. But anyway, I'm getting sidetracked. So the specific type of neural network that duplex is using, as I mentioned, is called a recurrent neural network. These type of networks have a small internal memory that allows them to remember specific

Starting point is 00:06:46 inputs to help understand context. For this reason, it's pretty much the perfect neural network for speech recognition and is at the heart of most speech recognition algorithms. So as I mentioned earlier, Duplex was trained on a whole bunch of different phone conversations, but how did it learn to understand what was going on? Well, the conversations, of course, start an analog speech, and then this speech in audio form is fed into Google's automatic speech recognition system. From this point on, this audio is now understood as text. In this text, one of the text, one converted into a format that the neural network can understand is then fed into the system. Other metadata and other wider context information from the calls, such as the correct

Starting point is 00:07:24 time for the appointment schedule and question, all the time of day, is also fed into the system. So tying it all together, when the neural network listens to a whole bunch of phone recordings as inputs, it eventually learns how to reduce the amount of error it has, meaning that it has better responses when spoken to. The final trained neural network that comes out of all of this is Google Duplex. How can I help you? Hello? Hello, what's up, man?

Starting point is 00:07:51 Hey, I wanted to know what are your hours for today? 10 a.m. to 6.30 p.m. Okay, got it. Thank you for your time. No problem, sir. What? So what can Duplex do? In addition to what was shown in the Google demo, Duplex can also do some pretty interesting things, such as handle interruptions. Okay, what's your film number?

Starting point is 00:08:11 2, 2, 3. 2, 2, what? 2, 2, 3. Okay, 2, 23. Elaborate? Hi, I would like to reserve a table for May 25th. Sorry, what day? For Friday, May 25th.

Starting point is 00:08:32 And respond to audio issues over the phone. Are you here? Yeah, I'm here. So there was a lot of talk about this AI passing the Turing test. So for those of you that don't know, the Turing test was a thought experiment thought up by Alan Turing, one of the fathers of computer science back in 1950. He proposed a test that goes as follows.

Starting point is 00:08:54 A person would interact with a machine that they couldn't see over text. This individual doesn't know if they're talking to another person or a machine. If the machine can interact with this individual without them suspecting that they're talking to a computer, then that machine passes the Turing test. Alan Turing made the prediction that by the year 2000, we would have machines that would be able to pass the test. and he wasn't that far off. The first machine to pass the test

Starting point is 00:09:19 was a text-based chat bot in 2014. The touring test was originally just for text and Google Duplex seems to have passed the touring test in the very narrow field of conversation when it comes to making appointments, but through voice, not just text. I bet that's definitely something that a lot of leading computer scientists

Starting point is 00:09:37 didn't see happening just yet, so I found that very interesting on that point. So some final thoughts. Some people may think that this may take jobs away from telemarketers and other phone-based work in the future. This may very well be possible, but I'd just be being reactionary if I was to say that this was a certainty, or if I was to say that this will be a net negative on society in any way, because at this point, it's far too early to tell conclusively. A reminder, Duplex isn't for general

Starting point is 00:10:03 conversation, it's only specifically for booking reservations and inquiring about open hours during holidays, but I'm sure its scope is going to grow. As far as the assistant goes, Due to its limited scope, it doesn't really change how we live our lives, unless you have a disability, I'd say. If a store has an online page, it's far quicker just to book that way. But then again, Duplex is using a neural network, and time and time again on this channel, we've seen how those things can surprise us. The scope may just grow quicker than we think. But on a wider note, this technology is pretty cool. It's becoming clear that we're at an inflection point when it comes to AI.

Starting point is 00:10:39 In the past couple of years, we've been marveling at AI breakthroughs like AlphaGo and others. But now, it seems like we're starting to see real-world applications of AI coming into view, each time making what was once thought impossible, possible. I think this will start to get more and more commonplace, as breakthroughs keep occurring in parallel. So, am I worried about duplex? Not at all, really. I don't see anything that should make me.

Starting point is 00:11:02 Google has stated that they're going to be transparent when you're talking to a duplex AI. But if this technology becomes commonplace, ask me that question again. But I think all in all, society has to accept that we're at the doorstep, of a brand new error, a time with unimaginable possibilities. And I think that's a bit of a privilege as it stands right now. Anyway, that just about wraps up this video. Thanks for watching. This has been Degogo, you've been watching Coldfusion.

Starting point is 00:11:26 Feel free to subscribe if you just stumbled across this channel. And I want to thank all of you guys that came to the Dubai Blockchain Summit. It was really cool meeting some of you guys. Anyway, that's it. I'll catch you again soon for the next video. Cheers guys. Have a good one.

ColdFusion - Google Duplex A.I. - How Does it Work?

There aren't comments yet for this episode. Click on any sentence in the transcript to leave a comment.