Memory-Centric AI, Part I: How Kioxia’s Top Engineers Are Developing an AI That Relies on Memory - “In the Pipeline” Combining Memory and Technology in Pursuit of Future Horizons -

07.05.2022 Memory-Centric AI, Part II: An Internet of Memories: Brainstorming Uses for Memory-Centric AI - “In the Pipeline” Combining Memory and Technology in Pursuit of Future Horizons –

If you had access to artificial intelligence (AI) that knew you intimately, what roles would that AI play in your life, and in what areas? In this interview, Sputniko! examines the possibilities of Memory-Centric AI through the lens of speculative design, a field in which she developed her expertise as a student at the Royal College of Art in London and later as an assistant professor at the MIT Media Lab.

The Future of NFT (Non-fungible Token) Art

Memory-Centric AI is capable of learning new things, gaining experience, and maturing, just as we humans do. In other words, it can be ultra-personal artificial intelligence tailored to an individual—a significant departure from the one-size-fits-all solutions offered by conventional AI. In Memory-Centric AI, memory plays an important role in helping the model to get to know its user. What possibilities will such technology bring about, and how will our lives change as a result? In tackling this question, the three participants in the conversation begin by focusing on the field of art.

Sputniko!: I understand that Memory-Centric AI “matures” by learning all about its user through data, which allows it to make interpretations and decisions that make sense to the user. But another thing that artificial intelligence is good at is making suggestions that we as humans would never have come up with. I know that there are a lot of people, especially in creative fields, who use AI in this way.

Deguchi: By reconsidering the “black box” problem [discussed in the first part of this interview] from a creative point of view, we might conclude that the black box nature of AI is actually a good thing because it gives models the freedom to produce content that we would never come up with ourselves. If you were to look at the black box problem in a positive light but utilize Memory-Centric AI for creative purposes, how would you propose using the AI?

Sputniko!: Being able to see how data is used for decision-making will give the artist more control when fine-tuning an AI-generated work, which I think is a good thing.

NFT art is an emerging trend at the moment, especially in the United States. I myself have been creating NFT art, and I get the impression that AI-generated NFT artwork is becoming increasingly common. I felt a little threatened at first seeing so much NFT art created by artificial intelligence. I was incredulous, thinking that AI couldn’t possibly be an artist.

For example, I saw an article about artificial intelligence that can produce artwork like Rembrandt or Van Gogh. I told myself that it would be no more than an imitation, that AI wouldn’t be able to truly compete with an artist because it doesn’t possess an artist’s creativity. But as I look at all of the AI art appearing in the NFT community these days, I realize that there are many artists who use an artificial intelligence as a collaborator.

Traditionally, artists are taught in art school how to visually express themselves and then go on to create their own art. But these newcomers are creating art with AI programs, without any artistic training. What’s more, if you collaborate with AI, you’ll start to see some pretty interesting works after just ten tries or so. I really felt threatened by that. And some people are selling art that they’ve created using artificial intelligence for as much as 10 ether [about 4 million yen] apiece. It makes me realize what an incredible time we’re living in.

Deguchi: That’s really interesting.

Sputniko!: Isn’t it?

Deguchi: Conventional AI generates multiple different solutions before choosing one to provide to the user. With Memory-Centric AI, I imagine those solutions would be a lot more tailored to your preferences.

Sputniko!: Right. Even in AI collaboration, human judgment ultimately comes into play in terms of sensibility and style. I think being able to customize their AI will give artists a more direct sense of ownership. There’s an idea—why don’t you create a Kioxia Art Department to make NFTs?

Yoshimizu: With today’s artificial intelligence, it has become possible to enter a question in text and get back an answer as an image or as a sound. Or you can input a picture and get a sound as an answer. I think that the ability to freely translate between the five senses—for example, detecting a smell and describing it in words—is an interesting, creative benefit that AI could offer artists.

Sputniko!: Yes. For example, you can input keywords like “natto” (fermented soybeans) and “giraffe,” and the AI would generate a strange natto-covered giraffe or something. [Laughs.]

Yoshimizu: It’s like forced matching.

Sputniko!: In the NFT era, people are turning words into pictures and pictures into sounds, which I feel has redefined each of the five senses. Algorithms have been creating generative art since the ’70s or so, but in the age of NFT art, for an artist like me artificial intelligence and algorithms are an almost indispensable means of creating more and more works, as well as collectibles.

At first, AI-human collaboration was interesting simply because of the novelty factor, but now it is performing an increasingly important role in the art market. I, too, would like to use Kioxia’s tools to create NFT art.

Deguchi: At the moment, it is the artwork generated by AI models that have value, but it may be possible to instead provide the AI itself as a product once it has the know-how to create art. That’s because, with Memory-Centric AI, it should become possible to sell a portion of your personal knowledge data.

Sputniko!: Do you mean that by gathering knowledge, you could nurture an artificial intelligence capability that acquires its own unique style as a creator?

Deguchi: Yes, I think so. I think it may be that the knowledge itself will have value.

Sputniko!: That’s amazing! To be able to say, “I’ve been collaborating with this absolutely brilliant AI!” [Laughs.]

Deguchi: Artists might be able to share algorithms like you might share an art tool.

Yoshimizu: You can look at this kind of artistic AI as a rival, but you can also nurture it as a way to nurture yourself. Even if you were to lose the ability to control your body, your knowledge and way of thinking would still remain a valuable resource.

Sputniko!: Creative tools based on conventional AI are interesting, but there are only about twenty different types available, and so everything ends up being very similar. But with Memory-Centric AI, I could develop a creative collaborator that’s perfect for me and my style. This is so exciting—I want to start collaborating with your AI immediately!

Deguchi & Yoshimizu: We’ll work on it!

Warm and Welcoming AI

Deguchi: With our Memory-Centric AI, we want to present a model that relies on memory much more than conventional AI does. In doing so, I hope we will make the world a better place by helping people to see artificial intelligence in a more approachable light.

Yoshimizu: The availability of personal AI that matures in tandem with its users will, I think, help people begin to see their AI as a trustworthy partner. When you call your AI, it won’t respond with a generic reply determined by studying data in the cloud; it’ll give you an answer that only someone who knows you well could give. I would like to see a world of warm, welcoming AIs.

Sputniko!: The parents of people in my generation are aging. I hear that many elderly people are unable to go out much due to Covid and feel lonely because their family members live far away. I try to call my parents regularly for simple conversations, like asking them how they are or what they did today or just chatting about the weather. But people need to talk and interact with others for longer than just ten minutes a day, so it would be great to have artificial intelligence that can converse with my parents the same way I would.

It’s of course important to call them myself, too, but I want to make time for my parents during the many hours that I’m not on the phone. I wouldn’t want to leave it all to AI, but I would like to see AI become a part of their lives and offer them a little more interaction. If I could personalize an AI capability based on my speech mannerisms, my memories, and my parents’ memories, it could interact with my parents in a way that’s completely different than conventional AI. It would be easy to talk to.

Deguchi: That fits perfectly with our ideal of artificial intelligence that relies on an individual’s memories.

Sputniko!: Also, I recently gave birth to my first child, who is now six months old. Whenever I look at her, I think about how much she’s changing week by week. Last week, she would calm down when I gave her a pacifier, but this week, she’s been spitting it out. I am just like an AI module, learning about the best ways to soothe her as she changes. I think it would be wonderful to have an AI partner that could watch her grow up with me.

Maybe it would be installed in the cradle where she sleeps, so that it learns how to soothe her when she cries at night. I would think it would be possible to develop Memory-Centric AI based on a baby’s growth and memories.

Deguchi: A personalized AI-driven cradle that knows how to rock a baby and calm it down? It would be learning constantly.

Sputniko!: It would have to, since each baby is unique.

Deguchi: It’s a great idea for using Memory-Centric AI.

Yoshimizu: I think it comes down to memory. When something unexpected happens, many people’s first instinct is to search for advice online. But if you had a dictionary-like AI that continually updated information related to your child’s idiosyncrasies, you could get a personalized answer telling you what to do.

Sputniko!: Right. It would combine general child-rearing knowledge from the Internet with its specific data on your child to help you make the best decision.

Yoshimizu: There are people who are conducting trials using robotics in nursing care. That might be another field in which Memory-Centric AI would be helpful.

Cartoon Characters and AI

Sputniko!: As an artist, I sometimes conduct strange experiments with AI. For example, I recently had an AI learn all of Yukio Mishima’s novels. After creating an AI that could think like Mishima, I experimented with it, asking, “What would Yukio Mishima say?” when I was in doubt about something. And this is not limited to novelists like Mishima—you could also teach an AI about Mickey Mouse, for example, to create an AI module that responds like Mickey. *To the extent that it does not infringe on the copyrights or intellectual property rights of others. The same applies hereinafter.

I was actually bullied in preschool, and I think about how comforting and encouraging it would have been if I’d had an AI Mickey Mouse to talk to. To be able to talk with your personal hero or heroine in a way that helps you—that’s what I dream of.

Deguchi: That’s an interesting idea.

Yoshimizu: Everyone has their own reasons for liking Mickey Mouse, so each person could shape Mickey based on what they like about him. There would be a universal Mickey dataset, but each person’s version would evolve based on how they answer Mickey when he asks, “What do you like about me?”

Sputniko!: Your Mickey would begin to show more and more of the traits that you like best. Many people have characters that they love, so I like the idea of them being able to shape a character into something that’s just right for them.

Yoshimizu: That would be a great product. And it would be easily accessible, without having to be connected to Wi-Fi all the time.

Sputniko!: Some people in Japan even fall in love with fictional characters. If you were to market this kind of AI to these fans, I think there would be tremendous demand.

Deguchi: Fans like new things, don’t they? It would be an interesting collaboration.

Sputniko!: Assistants like Siri, Alexa and Google Home™ are designed to have as little personality as possible. But what we’re talking about would be quirky and a lot more fun.

Deguchi: That kind of all-purpose virtual assistant requires a lot of data and a lot of power, since they need to make generic decisions determined by studying a huge amount of data. Personalized AI may not require as much power and may be more efficient to build.

Knowing a Person Through Their Life Log

Sputniko!: So far, we’ve been talking about artificial intelligence that would learn our actions and speech mannerisms. But since we are human beings, we also have a lot of biodata they could pull from. I think AI that learns from our biodata would be interesting.

For example, every month, women go through hormonal changes associated with their menstrual cycles. To some extent, women learn how their bodies change and plan their daily lives around that, but I would be grateful to have an AI facility that could track my body for me and work with me to plan my work and private life.

Yoshimizu: AI can handle anything that can be converted into data, so for starters we could partner it with a biometric sensor manufacturer to track your physical condition. Also, if AI tracks your state of health intelligently on the basis of the sorts of data that can be obtained now, it should be possible to tailor its recommendations to your actual physical condition.

Deguchi: For example, let’s say you have an AI program that suggests meals to prepare. If you input what you have in the fridge and what you want to eat, it suggests a recipe. That can already be done with existing technology. But our AI could also sense when you’re hungover, for example, and don’t have the brainpower to make any decisions. It knows the contents of your refrigerator, so it could tell you to try, say, vegetable soup. You would only want to customize an AI program with data from your day-to-day life if it’s trustworthy and has gotten to know you, so I think this might be a good application of the kind of artificial intelligence we’re working on, which relies on memory.

Sputniko!: I see potential in AI that can get to know me based on my biodata. It can’t truly know me without knowing about my body.

The Internet-of-Memories Era

Yoshimizu: One thing we will need to explore in the coming years is how our AI is going to record and analyze our life logs. To me, it seems intuitive to capture all this data through a virtual world such as a metaverse. Besides, presenting artificial intelligence as a two-dimensional partner in the metaverse would be a good fit from a sensory perspective.

Deguchi: In the metaverse, all data is digitized. Perhaps we will be able to create AI that helps you maintain your identity in the virtual world where all your personalized data is stored. Or perhaps an alternate personality that uses your data to act on your behalf within the metaverse when you’re not around.

Sputniko!: Since we’re spending more time in the digital space, it would be easier to compile a life log in the metaverse and use that data to create a personalized sidekick.

Deguchi: If we want to gather data quickly, that might be the best place to start.

Sputniko!: You might be right.

Yoshimizu: Personally, I’m more interested in real-world spaces. Since our product is physically tiny, we haven’t yet figured out the size of the computer that would be used to control it. However, if storage capacity is the limiting factor for raising Memory-Centric AI, we have an advantage in that we produce technology that allows a lot of information to be stored in a small space. In the metaverse, there’s a lot you can do simply by using cloud data, but what we really want is for every device in the real world to serve as an intelligent assistant.

Sputniko!: You want the intelligence to be stored in the device itself.

Yoshimizu: Exactly. We want to be able to put our very thoughts into devices. I also want to ensure that our product is compact. That is the thing about Kioxia: we want to create a world where each and every object has memory.

Sputniko!: It’s true that there is memory everywhere. You know, the term Internet of Things, or IoT, was coined a while ago. I think Kioxia could make an Internet of Memories, or IoM. Something that connects memories, not just things.

Deguchi: IoM—I love that.

Sputniko!: I hope you’ll start spreading the word on IoM.

Yoshimizu: Absolutely. We’ll make it happen!

Siri is a trademark of Apple Inc.
Google Home is a trademark of Google LLC.
Any other company names, product names, and service names may be trademarks of their respective companies.

The content and profile are current as of the time of the interview (March 2022).