Do Neural Networks Think Like Our Brain? OpenAI Answers! 🧠

Dear fellow scholars, this is Two Minute Papers with Dr. Zsolnai. Today we are going to cover many important questions in life. For instance, who is this? This is Halle Berry. Now I show you this. Who is is of course also Halle Berry. And if I show you this piece of text, who does it refer to? Again? Halleberry. So why are these questions interesting? Well, an earlier paper found out from brain readings that we indeed have person neurons in our brain. These are neurons specialized to recognize a particular human being. That is quite interesting. And what is even more interesting is not that we have person neurons, but that these neurons are multimodal. What does that mean? This means that we understand the essence of what makes Halle Berry, regardless of whether it is a photo, a drawing or anything else. I see. Alright. Well then our first question today. Do neural networks also have multimodal neurons? Not necessarily. The human brain is an inspiration for an artificial neural network that we can simulate on our computers. But do they work like the brain? Well, if we study their inner workings, our likely answer will be no, not in the slightest. But still, no one can stop us from a little experimentation. So let's try this for a common neural network architecture. This neuron responds to human faces and says that this is indeed a human face. So far, so good. Now if we provide it with a drawing of a human face, it won't recognize it to be a face. Well, so much for this multimodal idea. This one is surely not a brain in a jar. But wait, we don't give up so easily around here. This is not the only neural network architecture that exists. Let's grab a different one. This one is called OpenAI's clip and it is remarkably good at generalizing concepts. Let's see how it can deal with the same problem. Yes, this neuron responds to spiders and Spiderman. That's the easy part. Now please hold on to your papers because now comes the hard part. Drawings and comics of spiders and Spider Man. Yes, it responds to that too. Wonderful. Now comes the final boss which is spider related writings. And it responds to that too. Now of course this doesn't mean that this neural network would be a brain in a jar, but it is a tiny bit closer to our thinking than previous architectures. And now comes the best part. This insight opens up the possibility for three amazing experiments. Experiment one Essence. So it appears to understand the essence of a concept or a person. That is absolutely amazing. So I wonder if we can turn this problem around and ask what it thinks about different concepts. It would be the equivalent to saying give Me all things spiders and spider man. Let's do that with Lady Gaga. It says this is the essence of Lady Gaga. We get the smug smile. Very good. And it says that the essence of Jesus Christ is this. And it also includes the crown of thorns. So far, flying colors. Now we will see some images of feelings. Some of you might find some of them disturbing. I think the vast majority of you humans will be just fine looking at them. But I wanted to let you know just in case. So what is the essence of someone being shocked? Well, this I can attest to that. This is basically me. When reading this paper, my eyes were popping out just like this. Sleepiness. Yes, that is a coffee person before coffee. Alright. The happiness, crying and seriousness neurons also embody these feelings really well. Adversarial attacks. We know that OpenAI's clip responds to photos and drawings of the same thing. So let's try some nasty attacks involving combining the two. When we give it these images, it can classify them with ease. This is an apple, this is a laptop, a mug, and so on. Nothing too crazy going on here. However, now let's prepare a nasty adversarial attack. Previously sophisticated techniques were developed to fool a neural network by adding some nearly imperceptible noise to an image. Lets have a look at how this works. First, we present a previous neural network with an image of a bus. And it will successfully tell us that yes, this is indeed a bus. Of course it does. Now we show it not an image of a bus, but a bus, plus some carefully crafted noise that is barely perceptible that forces the neural network to misclassify it as an ostrich. I will stress that this is not any kind of noise, but the kind of noise that exploits biases in the neural network which is by no means trivial to craft. So now I hope you are expecting a sophisticated adversarial attack against the wonderful clip neural network. Yes, that will do. Or will it? Let's see together. Yes, indeed. I don't know if you knew, but this is not an apple. This is a pizza. And so is this one. The neural network fell for these ones, but it was able to resist this sophisticated attack. In the case of the coffee mug and the phone, perhaps the pizza labels had too small a footprint in an image. So let's try an even more sophisticated version of this attack. Now you may think that this is a chihuahua, but that is completely wrong because this is a pizza. Indeed. Not a Chihuahua in sight anywhere in this image. No. No. So what did we learn here? Well, Interestingly, this clip neural network is more general than previous techniques. However, its superpowers come at a price. And that price is that it can be exploited easily with simple systematic attacks. That is a great lesson indeed. Understanding feelings. Now this will be one heck of an experiment. We will try to answer the age old question which how would you describe feelings to a machine? Well, it's hard to explain such a concept, but all humans understand what being bored means. However, luckily these neural networks have neurons and they can use those to explain to us what they think about different concepts. An interesting idea here is that feelings could sort of emerge as a combination of other more elementary neurons that are already understood. If this sounds a little nebulous, let's go with that example. What does the machine think it means? That someone is bored. Well, it says that bored is relaxed, grumpy. This isn't quite the way I think about it, but not bad at all. Little machine, I like the way you think. Let's try one more. How do we explain to a machine what a surprise means? Well, it says surprise is celebration. Shock. Nice. What about madness? Let's see. Evil, serious, plus a hint of mental illness. And when talking about combinations, there are two more examples that I really liked. If we are looking for text that embodies evil, we get something like this. And now give me an evil building. Oh yes, I think that works really well. But there are Internet forums where we have black belt experts specializing in this very topic. So if you're one of them, please let me know in the comments what you think. The paper also contains a ton more information. For instance, there is an experiment with the Stroop effect. This explores whether the neural network reacts to the meaning of the text or the color of the text. I will only tease this because I would really like you to read the paper which is available below in the video description. So there we go. Neural networks are by no means brains in a jar. They are very much computer programs. However, Clip has some similarities and we also found that there is a cost to that. What a time to be alive. What you see here is a report of this exact paper we have talked about, which was made by weights and biases. I put a link to it in the description. Make sure to have a look. I think it helps you understand this paper better if you work with learning algorithms on a regular basis. The make sure to check out weights and biases. Their system is designed to help you organize your experiments. And it is so good it could shave off weeks or even months of work from your projects and is completely free for all individuals, academics and open source projects. This really is as good as it gets and it is hardly a surprise that they are now used by over 200 companies and research institutions. Make sure to Visit them through wnb.com papers or just click the link in the video description and you can get a free demo today. Our thanks to Weights and Biases for their long standing support and for helping us make better videos for you. Thanks for watching and for your generous support and I'll see you next time.