Design for our times: workshop 4: speculative design, AI, machine learning and design

Par Conseil national de recherches du Canada

Design for our times: workshop 4: speculative design, AI, machine learning and design - Transcription

Dr. Sara Diamond: Although we meet today on a virtual platform. I'd like to take a moment to acknowledge the importance of the lands, where each of us live and work. From coast to coast in Canada, we acknowledged the ancestral and and unceded territory of the Inuit, Métis, and First Nations people that call this land home. For example, in Toronto, where I am today, we are in the traditional territories of the Mississaugas of the Credit, the Haudenosaunee, the Anishinaabe and the Huron-Wendat - but also in a city that is a major gathering place for Indigenous people from all over Canada, including Metis and Inuit. The discovery of 751 unmarked graves at Cowessess First Nation in Saskatchewan last week is yet another reminder that we must acknowledge the wrongs of the past and take action on the findings of the Truth and Reconciliation Commission. Canada will benefit if we engage the values, perspectives, languages, and cultures, which can guide and inspire us, acknowledging the role of Indigenous people as guardians and stewards of the lens and consider how we can each in our own way, move forward in a spirit of reconciliation and collaboration. Thank you.

So today we will briefly review strategic foresight, which is a methodology that analyzes systems to create plausible possible futures. We provided an in-depth overview of these tools and a case study. And these are available in the repository for this site, both in slide form, and as a video from workshop three and we'll tell you at the end of today's workshop, how to, again, how to access those. Then we're going to expand the discussion of speculative design that we began in workshop three, discuss industrial design in the ways it bears relevance to the NRC and as being substantially transformed by artificial intelligence and machine learning, including by generative design, which is a specific subcategory of AI and ML. We'll touch on a few examples of what we think are exciting AI applications in the product and industrial design context. And move on to discuss generative design and architecture and urban planning, which has relevance to new building materials research, amongst many other things. The last years have seen new tools for data collection and a growing awareness of algorithmic bias as well, distrust by some in adopting AI solutions. So we're going to look at the ways that designers can bring value to explainable AI, and then consider the balance between a data collection, AI solutions and user understanding and control. I think that will be a thing throughout the next 90 or so minutes. And we'll end with some thoughts on the future of AI and the role of design in imaging and imagining that future.

Just a bit of a warning today has some fairly technical sections. And you can ask for clarity on what the definitions are. The idea is to wet your appetite. If you're immersed in working in this space, you may find this well, I already know this, but we're hoping that will be just in one section of our workshop as we move on to some of the other AI and machine learning applications. And every week, by the way, I wear a Canadian designer and today it's Joseph Ribkoff from Montreal, just promoting that great Canadian fashion design industry. Also every workshop we've featured a Canadian design company, touching on fashion, services and consulting. And today we are briefly going to explore Myant, a scale-up company that's based in Etobicoke, Ontario, that brings together many features of industrial design in the context of health and wellness and also seniors care and the Internet of Things. Myant has patented smart, sometimes known as intelligent textiles, also known as textile computing, and these are able to collect highly accurate data and are connected by the Internet of Things for some two different kinds of outputs and applications. For some applications Myant has engineered a pod that captures and transmits data to their bespoke AI driven analytics engine and a visualization interface. So its products are relevant for many industries – health and wellness, fitness, aerospace, and automotive workplace safety and home textiles are, but a few. Myant was founded in 2010 by Tony Chahine and Ilaria Varoli and they lead Myant. And Myant currently employs a hundred plus engineers, researchers, and designers. It has over 70 patents and an 80,000 square foot manufacturing facility in Canada where they manufacture some more textiles. So they've really tried to keep their production in Canada. And after many years of experimentation, Myant is about to release its first product. So the Myant Skiin Connected Life App, which is called Skiin is a family informatics application, which will connect family members, friends, and caregivers by engaging them together and enabling health and wellness related data sharing and support. And Myant's garments collect the activity and the related biometric data such as heart rate activity and body temperature data. And you can also input your mood. So there's the opportunity to do analog augmentation of the data through knitted sensors on the garment. And it also tracks location. So data are transferred from the garment pod to the mobile Skiin app via Bluetooth and users create a health profile and identity and pair their Skiin Pod with the application and then connect to their loved ones, friends and caregivers, with whom they can share data. And then also see the data that their connection chooses to share with them. Data are analyzed and flagged on the Myant Platform by machine learning, trained to spot deviations from a specific user's baseline and are able to aggregate longitudinal user data. So the user can link to outside sites to access contextual data and social networks to converse, and they can plan activities with their support network, as well as monitor safety and location. So for seniors care, this is quite important for families and caregivers, for example. Skiin is also a solution for telehealth intervention and delivery and it's just starting a trial in Northern Ontario in partnership with Algoma University and the Sault Area Hospital with Skiin kits to be distributed to about 2,500 trial participants. Myant they're small, but they're very energetic. They're research and design centric. The company collaborates with the universities. So the University of Toronto, OCAD university, Ryerson University, Algoma U and also with Toronto Rehab and UHN’s KITE, which is a world-class rehabilitation center. Southlake Hospital, Myant has an agreement with them and especially with their elder care there, and they have international partnerships with the Mayo Clinic one with Canadian Tire, with SMX, which is looking at creating a textile computing supply chain including with a strong Canadian component, with Helly Hansen for sports wear and Myant depends and here's, I think a very important point for us as designers, on excellent UX and UI. So the interface design the experience design as well as the data visualization and of course, industrial design underlies this, and it has a strong personalization strategy. And, we'll keep reminding ourselves that's one of the lessons of 21st century design is even with mass products to find means of personalizing them for the user. So that's our Canadian company for today.

We're now going to talk a little bit about strategic foresight and there really is a review, and we know some of you were with us last time. So we'll be brief. So the Myant saga though really speaks to effective strategic foresight. Its leadership has really looked at changes and needs and in demand and has pivoted the company to meet those needs. And just as a reminder, the process of strategic foresight begins with a horizon scan. So in that stage, you identify trends and drivers of change, and these are collected, clustered and analyzed. And that process happens both through secondary research and through data analysis. These are then used to create multiple and diverse scenarios of possible futures. And I should go back and say that foresight strategists/designers often will use survey instruments to collect data that's fresh data. And we'll talk towards the end of this on surveys and their strengths and challenges. In any case, the idea is to create diverse scenarios of possible futures. And these different futures are enriched with visuals and through storytelling to make them as lively as possible. Finally, the scenarios are used in a process called wind tunneling. And you get this visual of, testing. Where strategies for the future are tested against the scenarios, changes and adjustments can be made so that strategies can be more robust. And you always look also for this sort of intersection place where you see trends that are evident before you go out to the more different possibilities in the quandrant. And so in our last workshop, we explored the question: What is the future of sustainable food adoption in Canada and the market for Canadian sustainable food products. And we built four intersecting worlds: localized monoculture of garden city. So you can see this on the left. The localized biodiversity of slow growth. And you can see that our axes are biodiversity and monoculture and localized and global. And then the globalized monoculture of lab-meals. And the globalized bio-diversity of magic mushrooms. If you were not here and you're now intrigued, please access the slides and video of workshop three, and you'll learn more about it. And we had great participation from NRC colleagues in helping us to shape up the process as we went through it. So the practice of foresight and of applying a future oriented mindset when creating and strategizing evolved over many years and branched into other ways of designing with the future in mind, this is constantly evolving. So new design practices such as design fiction designed futures or speculative design have emerged. Speculative design, which we touched on very briefly last workshop is an extension of strategic foresight that imagines future worlds that are impossible, ridiculous, or on the edge of reality where systems or technology do not yet exist. In order to understand how pressing issues of the future can be accounted for today. Speculative design was invented by Anthony Dunne and Fiona Raby and it hopes to spur a public debate on what is a predictable societal development. So this is really a method that is focused on policy development, but also on decisions around emerging technologies and Dunne and Raby argue that design should be used for problem finding not only to solve problems. It can be used to imagine outcomes of research very early in the planning cycle and speculative strategies can help us to imagine change in how we live, work, and play. It provides researchers at the NRC with a means to rigorously examine the future. And asks big questions, like what is the NRC's response to the possibility that there may be an exponential increase in Arctic melt, or what if General Artificial Intelligence can self-replicate? So it allows us to think through these large possibilities and then work back to thinking about design strategies and technology strategies that would respond. And with speculation being both a means to outline possibilities and to resolve them. There's essentially five distinctive steps. So the first is to define, so to speculate upon the future requires specificity. So it can be an identification of a fringe technology or a cultural trend. And these are put into focus, which allows for future scenarios to unfold. And you're asking the question of how can emergent technology and subsequent experimental uses present signals that suggest alternate yet plausible outcomes. The second is to ideate so really to choose a timeframe, and usually for speculative design, it's five to 15 years ahead and it depends on the industr, to allow for a substantial projection. Then there's narration, which is essentially storytelling. So here you design a solution or an artifact or a system that might solve an issue in that future world and have that choice, tell a story about its future world. So this allows for the participants to begin imagining alternate futures to contextualize the design. Then you generate which is the final step. So to engage in discussion surrounding the fictional story to project, what issues may arise or the fictional technology or the fictional objects. How would that solution be implemented and what are the cultural, technical or economic restrictions that could arise in that scenario? And what would that solution be? Would it be integral to this future? And then you respond. So the questions are asked in the generate phase are then setting the stage for a research agenda that becomes sensitive to these issues of what the future might become. So those are the stages. And we're going to give a case study right now, which is the Strange Telemetry Case Study In 2015, the United Kingdom hired the consultancy Strange Telemetry to speculate on a set of challenges and opportunities that aging in place could present. So speculative design was used as a kind of cautionary exercise to get ahead of the rise in aging populations. And it marked the first act of use of speculative design in the UK government policy process. The designers created what they would describe, or we would describe as visual artifacts that were customized for this project. So these were a deck of cards that allowed the residents of the area to discuss and debate possible developments in employment, in services, and in transport. The participants were almost all over 50 years of age and they looked over the projected 25 year horizon with some of them noting they may not live through that future, but were really excited about informing it. And they used the card deck to respond, to propose changes in access to urban transit and the transition to digital as current trends, two of which could deeply affect the UK's aging population. The cards were color coded and they were prompts. So residents would pose alternate futures by responding to questions. Purple cards asked, and this is just an example, what changes would you want others? So policy makers, local government and companies to make, if this scenario might be in your future or part of your future. And they held a number of workshops. I'm only going to describe two of them. So the first looked at the future of work for seniors. And they were prompted to respond to issues of digital adoption and automation. So this is an example, 25 years ahead of a family owned robot repair store. Another scenario was an e-lancer’s living room. So an e-lancer is an electronic freelancer in this scenario, a very pervasive gig economy. And then a third was a piece work call centers. So that exists now, but protected, into the future. Another workshop explored the future of urban transit for aging citizens. And then within this, there were two quite contrasting scenarios. One was a state-backed, localized, social transit infrastructure, and the third was more of a privatized kind of multi solution ride sharing, automated vehicles et cetera, et cetera, solution to transit. Cars being retained within scenario two. And the design researchers created reports for the UK government on the findings. So the participants in general expressed greater trust in state power in supporting transit than corporate control, but they disagreed around what forms of work might be desirable to an older generation. Some saw the primary need as being for income. Others thought it was important to do socially meaningful work that was recognized by society, especially as an older person and also others saw work, which would allow them to build on their existing expertise. So participants also questioned what was a real interpersonal interaction, whether online or in person, in terms of types of work and service, provision and mobility. So how much would they be interacting with thoughts AI agents, at every stage of their needs versus with actual people or they themselves performing labor, like at a call center as opposed to a bot like an automatic chat agent, for example. And there was another report that was created on the methodology of speculative design. So designers use design fictions to engage seniors in imagining the design and use of technologies that would provide home support and monitoring have also been another direction. This builds a little bit on our conversation around Myant and the previous story. In Aseen Ahmadpour and their colleague’s experiment , older adults developed concepts of wearable monitoring devices through imaginative dialogue. It turned out that they were particularly interested in anthropomorphic or animistic toys, not bracelets, not watches or wearables. And they imagined a version of Gorby. If you have kids or grandkids, you may know, or if you yourself like robotic toys. Gorby is a popular, stuffed animal, and they imagined a version of Gorby as a pet robot, expanding on the capacities of the commercial toy who would registered heartbeat and body temperature. So what the seniors imagined as they would hug the stuffed toy and the sensations could be felt by the user through vibration. So they were hugged back and then there'd be temperature changes of the toy as well when it was hugged and so there was a virtual hug through this virtual channel as the toy at the same time collected the biometric data from the senior. Gorby could also send hugs through the Internet of Things to the senior’s network, facilitating relatedness with their extended network or family. So this is a use of design fiction and participants offered these insights into the future of wellbeing monitoring technologies. And that really surprised the researchers that this was the form that it took. We chose Myant, Gorby Design Fiction research and Strange Telemetry as these may be relevant methods to the Aging in Place Challenge program at the NRC , which is developing sustainable models for longterm care, by shifting towards home-based and community based care. I'm quite familiar with this research area, I chair the research advisory committee at Baycrest and you are a very important partner to Baycrest. We're just going to pause here for a minute and see if you have any comments or questions, please put them in the chat about ways that you think these tools could be useful for your research or any questions about the methodology or the actual projects. Not seeing any quite yet. Okay. We have lots of opportunities for you to converse with us as we move through the rest of our presentation today.

One of the related. And growing branch of speculative design is Indigenous futurism and Dr. Grace Dillon is the person who developed this term. She articulated the idea and then the philosophy. And so she draws on the Indigenous view, one that is held by many different Indigenous people, different communities, that it's important for humans to look ahead, seven generations to imagine and take responsibility for the consequences of contemporary decisions. There's increasing creative works of indigenous futurism and also the methodology being used in sustainable design planning in circular economy planning, which we talked about in workshop one which is looking at every point of a product, looking at how it can be reworked, reused, recycled and in actual technology invention. In this work inside communities, there is a focus on developing images, fiction, fantasy, and reinterpretations of histories. As I said, often with a focus on technological innovation and indigenous technologists and philosophers will tell you that indigenous people have always been tool builders. They were in some ways, some of the first tool builders. So we want to just give you an example of an artist, Skawennati who is based in both Mohawk Territory here in Canada and in Montreal. And she's a digital media artist. She was born in Kanien’kehá:ka Mohawk territory. She produced the series "She falls for ages", which is a science fiction, retelling of the Haudenosaunee (Iroquois) creation story that re-imagines Sky World as a kind of futuristic, utopian space and Sky Woman as a brave astronaut and world builder . Sky woman returns to a planet that she created when its tree of life is threatened. So obviously a metaphor for earth. Also in her other work, she crafted a series called the "Time Traveler Series" it's built in Machinima, which some of you may be familiar with. It's actually a gaming engine, animation gaming engine, that a lot of digital media artists have kind of converted into something that can be used for both linear or interactive storytelling. So the series explored indigenous possible futures throught its characters, and she created this with her partner, Jason Lewis, who's of Hawaiian and Samoan descent. And he's the Concordia University Research Chair in Computational Media and the Indigenous Future Imaginary, and Lewis founded AbTeC , which is Aboriginal Territories in cyber space, and it teaches indigenous youth games and animation design and basic programming too. And they create interactive experiences to look at the future of their communities and the planet. Lewis's AbTeC is grounding students in well-being by allowing an imagined and tangible future. So we thought a few examples here of student work it's quite beautiful. And, some are about looking at how to bring back a world that's been devastated and looking at different kinds of communities and clans. So those are the images that you're seeing there and we'll pick up on Jason's work as a designer on the future of deep learning and ethics and AGI later on. But he and AbTeC are an interesting potential partner for for the NRC in terms of a really great group of talented youth. So I just want to jump back to the present now, keeping in mind that foresight work and speculative design are tools that can help to understand potential devices, systems, and services.

So we're going to start with industrial design and the impacts in integration of AI and machine learning. So industrial design began as product design. So for those of us who've been with us since workshop one, remember all those chairs in workshop one? And in this century continues to be the design of products, including chairs, for manufacturing and advanced manufacturing and of systems and services. Also of sustainable solutions. Design with, and for emerging technologies, it's expanded to interaction design and experience design, and has moved into areas such as design for health and wellness and a focus on design for the circular economy. And somewhat would argue that design thinking also came out of the industrial design ethos. AI and machine learning, learning cover a very broad spectrum. And Andreas Kaplan and Michael Heinlan classify artificial intelligence into three types of systems. And we'll just use this as a hand wave to the complexity of this space. The first is analytics. So those really are systems that are assist to humans and can manage an analysis of large data sets. Human inspired AI, which includes some effective computing tools. Sentiment analysis, for example, and then humanized artificial intelligence. So we're going to talk about the latter at the end of the workshop, but much applied or good old fashioned AI, which is the first category consists of tools that support the cognitive using a symbolic information processing model, rooted in automation, in which machines assume and assist with human tasks because of their efficiency in processing data. The next phase, which has really humanized AI, it shows characteristics of all kinds of competencies. So cognitive, emotional and social intelligence, and really in the third phase, which is AGI, it's able to be self-conscious and self-aware in interactions with others. And we know this is really the meat of lots of speculative fiction, but also of current research. And as I said, we'll come back to that. And as with many disciplines, AI and ML are disrupting the ways that designers work, the tools that we use and resetting where we fit in the value chain. So it's quite disruptive in terms of creative industries. So example applications of relevance to industrial design. There are many that these are just a few so design for the Internet of Things, robotic human interfaces, such as toys, elder support, integrating robotics into manufacturing, but thinking through both that process in partnership with engineers, because you have to think about the process of making that product. AI and ML embedded systems, so we saw that with Myant, the use of image recognition technology to identify models. Generative design, which we are gonna explore in some depth, using AI to build the design through widespread data sampling a bit reductionist, but basically that, and then AI driven usability research. So those are just some examples. And in our last workshop, for those of you who are with us we talked about idea couture as our exemplar Canadian design company. So they are a strategic foresight company that originated in Canada. They've now gone global and it was led by Idris Mootee and Idris after he used the term cashed out from idea, couture created a new company, which is called Urban Cool Lab, and he founded it with two of his children and it's an AI design company. While it's Larger mission is to advance computational creativity. So bringing together AI, cognitive science, psychology philosophy, and the arts it's focused right now on fashion and a sort of narrow bandwidth of fashion. So it's using machine learning to glean information from massive amounts of data on street culture, which their AI algorithms analyze. And then customers define keywords, which should suggest societal memes, celebrities, concepts. And the company's machines then spit out street clothing with designs that correspond to these ideas. Stylistically and also in the ways that logos and texts are used. They do, I've asked Idris about this, try and screen the data for bias and certainly for hate language and won't print those kinds of messages and in a recent panel with fashion designers and the fashion industry in Canada designers saw value in AI as a tool for kind of retrospective ideas for deep search that they could then use. Because fashion, is quite retrospective. It cites the past, but they didn't find it as a means to really look forward. So to think about ways of creating new physical forms and new styles, and, designers right now are really involved with interesting work with 3d printing and phase metal materials, et cetera. But they want to use their minds to think about how to use those materials. It was an interesting conversation. I just note that we do have a question for us from Sandy Lu, who's asking about, are there tools to support this process? And I think Sandy is asking about strategic foresight and speculative design. Am I correct Sandy? Yeah, it's correct. So fantastic question. And one of the areas of work right now is, what I would describe as a shift in speculative design, but even maybe more in thinking about strategic foresight, is a lot of use of data analytics. So using AI and machine learning to crunch through large data sets, to look at trends, to extract attitudes and opinions and social media to use all of those data scraping and analytic tools, at that with a caveat. And we'll talk at length about this later on to be really careful about bias that's embedded who is dominant in social media and certain kinds of forums, et cetera. But for example, to get generational opinion or highly segmented opinion. So there's a sort of marriage right now between what was very qualitative ways of working with foresight and AI, machine learning and some great work in Brazil, for example, in this space where there's really strong computational capacity and in the Canadian context. So that's a great question. So thank you. I'm going to go on to pick up a little bit here on this whole question of creativity and the use of AI in design. Like the endless cycling of social media, AI, and machine learning captures us in a remix of the past. So it relies on past formula an example the Globe and Mail who I've collaborated a lot within, in various guises, and this is a system I built with my team that does deep segmentation of readers to look at what their tastes are and what their interests are. So this would be work that would inform what journalists do, who work for the Globe and Mail . But other work that sort of on the other side is to develop standard formats for local reporting that can be filled by a very sophisticated natural language processing program. There's this sort of displacement of the human reporter happening, not just at the Globe, but at larger media outlets around the world. The challenge here or the question is what about inventing the future through a human lens? And what about breakthroughs in creativity and design? And are we willing to allocate that role to AI and machine learning? And the assumptions of human needs and interests built in there, or is it even, are these agents even capable of that? So I'll just give some examples. So Renaissance breakthroughs and perspective, modernism, there is Mondrian, Islamic geometries and tiles, buildings and mathematics formulas. A lot of this is based on deep observation of nature, but with this phenomenal human creativity, new forms of music, like Jazz, Hiphop, or New Wave that were really breakthroughs in practices. So I'm weighing this critique, I think AI is an exciting tool. I'm not the only one weighing, designers and artists are, to develop new combinations of the past, whether recipes, music, or designs, and potentially imagine forms from natural patterns. So it's certainly not just a tool for optimization, but for creative application, but it'll be interesting to see if we're willing to accept a breakthrough that's generated through a design process versus human new forms of cultural expression. And their weight may be hybrids, Google labs in Paris and the US is filled with artists who are working with deep learning tools to think through new creative expressions. So there you go. And we'll come back to these questions for sure.

So we're now going to discuss generative design as it's a practice that's already in use at the NRC in your materials research, in aerospace we imagine in some other fields and data analytics, artificial intelligence, parametric, and geospatial visualization really do bring, when combined, a new set of possibilities to the design process. One in which patterns of data can be extracted and Danil Nagy and Lorenzo Villaggi , who worked a lot in this context as researchers. Describe generative design, and we'll just quote them as "the process of defining high-level goals and constraints and using the power of computation to automatically explore a wide design space and identify the best design options”. So here you go. And we're going to really go into detail here. So bear with me because I wanted to break down what the process is. So basically humans work with computers to apply meta heuristics. For those of you not in the computer science space, those are high-level procedures that coordinate other rules, and there are means to find, generate or select what then become optimization techniques. And the word design in this instance in part focuses on performance design. Although there are implications for aesthetic decisions as well. Humans are still, going back to our previous conversation, better than computers at creating bespoke products and AI and machine learning can calculate how constraints and conditions as well as changes to a product will influence different aspects of performance. So when coupled with aesthetic parameters or architectural aesthetics , AI can then generate prototypes of mass produced product. So that's sort of downstream after the generative design process. Generative design draws from specific computational approaches that will be familiar to some of you and not others. Two researchers who've been in this field for quite a while Vishal Singh and Ning Gu they've worked for over a decade and they've tested cellular automata. Genetic algorithms, where the principle is that the fittest designs, most aligned with design criteria survive and then merge with each other, swarm intelligence and multiagent societies, which are really helpful in modeling user behaviors in the space. So things like activity, clustering, or scheduling and then allow adjustments and shape grammars and also allowing adjustments and then also shape grammars and L-Systems have been deployed. And these are means to describe and write formal visual grammars step-by-step and in the latter instance with L-Systems by stringing symbols together, and we just have some examples here of both how these algorithms work, but how you could imagine their impact on a kind of physical geometry. So we're going to walk through the detailed generative design process right now. So first we define the design problem and we identify any objectives that are best addressed by the human designers, steering the generative design process. As it's often impractical to approach all design criteria in a competitional manner. So, I mean, I can testify I'm working on a large generative design project with a major developer and it's, absolutely research we had never worked at this scale before and we have to kind of reduce the sort of problem set and then the metrics we're going to define to six, right? And the story is more complex. It's also a lot about where you can find data of course, or generate data. And Ana Lisa Mayboom underscores. I think the problem quite neatly, she says part of the “Part of the process of parametric design is understanding what assumptions are embedded in the coding – what design is coding and what design is decision making outside of coding.” So using our design expertise then we choose the metrics. I just described that in a concrete story that describe the objectives or goals of a design problem. And then we build those metrics. We ensure that the chosen metrics sufficiently capture the priorities of the design problems. So we check this a lot before we start the generative design computation process and that they accurately describe the relative performance of each design, according to these metrics. So we're going to give you examples later, but some of the metrics will require the development of spatial analysis. For example, the layout of a building or the structure of an airplane, other analysis might involve financial calculations or statistical measures of diversity. The work I'm doing now with a large developer that wants to ensure that there's equitable housing, affordable, equitable housing. Greenhouse gas factors or materials capabilities, in some of the required algorithms, will be sourced from existing software toolkits, like Revit and Dynamo which is Autodesk or others will be implemented from the ground up. So we then applied back to metaheuristics search algorithms. So we apply a search algorithms such as a genetic algorithm to search through the design space and find a variety of high-performing design options based on the stated objectives. These are optimized and you look at the trade-offs of doing one versus another. For example, the multi-objective genetic algorithm combined sampling and cross-breeding in the use of the next stage to arrive at a diverse set of what are called Pareto optimal designs, sorry for all the tech talk, but we're showing them to you. So those are designs that improve the value of any of the criteria. Without deteriorating, at least one other criteria. So you're breeding. And then you start to eliminate and the set of designs reveal how to improve each metric and what tradeoffs must be made in the process. You use time plot graphs because they show the lineage of the designs as they evolve and they allow you to go back and edit to any point in the process and question and check it. And then you use a particular kind of algorithm, which is called a Principal Component Analysis algorithm. So it basically obtains lower-dimensional data specificity while maintaining the data's variation as much as possible. And this is incredibly important when you begin to talk about issues like inclusive design, where PCA is also used a lot to try and ensure that outlier datasets are included. And then you optimize the best designs in space, according to how they are weighted, given your metrics and you then cluster these again, according to your input metrics and you compare them to the output. And then you choose the designs, only 10 of them that represent the metrics and represent trade-offs. And then you take those and you undertake your consultation with your users. Your users are at the front when you're trying to figure out, what the priorities are, what the metrics are. And they're certainly at the end of this process as the designer. The stages here really are to generate a wide design space of possible solutions through a geometric system. So it's a geometry system. To evaluate each solution through measurable goals until we evolve generations of designs through evolutionary computation. And you iterate that process until you find your best designs. So we're going to give some examples in industrial design, and then we're going to talk about architecture and Urban planning and landscape, which is the space that I'm working in now.

Autodesk and Airbus teamed up to fundamentally change the way planes will be designed and manufactured and built in the future. And to look at not just the structure, but interior design, comfort and ways to maximize passenger numbers and generative design allowed the team to evaluate hundreds of design alternatives for multiple structured structural aircraft components, but also for interior design. Autodesk generative design algorithms developed lighter weight parts that exceeded performance and safety standards. And Airbus is in the process of using generative design to rethink other structural aircraft components, including the leading edge of the vertical tail plane of the A320. This research is relevant to the NRC's Aeronautical Product Development and Certification. You're focused on helping Canadian aerospace industry partners to reduce the cost of development and testing and evaluation, but also at looking at new kinds of products and technological readiness of these products. You may well be using, folks here from that research group, you may be using these tools already. We have another example which is Under Armour, which is really an innovator in performance, footwear, apparel, and equipment. And they use generative design technology with advanced additive manufacturing and they have created super hybrid trainers. So these are super flexible and they're stable for all types of athletic training. And the concept was originated by observing tree roots, which reminds us of design strategies and inspired by biomimicry, which is really about the imitation of natural processes. And generative design enables designers and engineers to explore a myriad of computer generated design possibilities, often using unconventional geometry, and it helps them arrive at the final product much faster. It allows for 3d printed prototypes to be tested by athletes very quickly in a fraction of the time that it would have taken in the past. And the company believes that new technology, such as generative design will help products to be made on a smaller local scale, improving efficiency and product quality and supply chain, and also allowing personalization and really reducing the time it takes to create new designs and get them to consumers. And again, back to workshop one, we emphasized how important personalization of mass products are in the 21st century. So, print your personally designed algorithmic generative runners.

So we're going to talk now a bit more about urban planning and generative design. So in this research, we use metrics from contexts as diverse as environmental data, users’ needs, architectural and building metrics, and by-laws, which can result in benefits to the physical, social and cultural environment. Generative design, again may be relevant to the NRC's research on housing for Northern communities and certainly for the High-performance Buildings program. So just a few examples here. So SUSTAIN is a project at Carleton University where generative design processes seek energy efficiency in building design and retrofitting, which is a lot of it. And they're trying to model how people move to the campus. And they're using actual data from buildings while wanting to pay attention to qualities of light, spatial characteristics, and heritage characteristics that they want to maintain for the campus. And the team is an engineer, architect, and computer scientist. Generative design can optimize urban landscape planning and will allow for example, the review of wall surface positions, heights, and textures of buildings. And you can vary these. And here's some research work done by the University of Toronto Daniel’s School of Architecture. Here are some images of generative design in relationship to the Autodesk research center in Toronto. So the factors that were considered were adjacency preference, so it's important to say that they surveyed and pull a lot of data and they considered six factors, adjacency preference, work style preference, buzz, productivity, daylight, and views to the outside. If you go to Toronto and you go to Autodesk, which is in the MaRS building, get yourself a tour and you'll get to see generative design in action. And people are quite happy with the design outcomes. Daniel Nagy and his team applied generative design to residential neighborhood of 7,000 square miles in Alkmaar in the Netherlands where environmental constraints and costs to the developer were two driving factors. Generative design models were used in this. It was meters not miles, sorry. That would be huge. In case studies with Bonava, which is a developer of affordable housing, generative design tools were used to create more detailed site-layout solutions in earlier stages, and measure both hard and soft values for each option. So again, using this for planning and then for actual design outcomes. And a growing interest in generative design right now are means to include stakeholder consultations whether by survey and other consulting tools or simulations. And the goal is to gather large data sets from current and potential users of the environment or their personas or demographic. Organize, categorize, and put these as design inputs.

So we just wanted to pause for a second here and just, we were curious to see whether any of you are using generative design. And if you had any comments on it or questions.

Justine De Ridder: We had a question, as you were presenting, asking for generic design, what kind of tool are needed.

Dr. Sara Diamond: Yeah. Basically, just going back. So you know, the tool that's on the market where people are using a lot of research in partnership with them are the Autodesk tool sets. Not to sell Autodesk. But it's a combination of using those tools for modeling and for analytics. And then you have to basically apply the algorithms to that process as you are making your choices as you move through. The tools that exist in the software kit from Autodesk are Revit and Dynamo. So you can use those, for search and for modeling, but then they write the genetic algorithms. So that's why computer scientists are sitting side by side with the designers, unless they have that skill. So genetic algorithms are used to call and breed the design possibilities. So it's a combination of tools that are basically fast, efficient design tools, sampling and analysis tools. And then using genetic algorithms that you write yourself, Hope that's helpful. Okay.

Just moving on. So this discussion on user engagement leads us to the role that designers can play in making AI decisions understandable, to understand that databases and algorithms that underlie recommendation systems effectively represent diverse needs. And to see if these capture what we said, or weak signals that are behaviors and practices on the edge. And supporters of ethical expansion and experimentation in AI, also, I think we all recognize the challenge of bias that can be baked into data collection methods and existing databases, including AI driven data collection and algorithms. We know that when bias data sets are used for machine learning or deep learning algorithms, these can reinforce ethnic gender socioeconomic or racial stereotypes and preferences, and as algorithms process and display, relevant data for human use, they may select information then based on previous historic choices of a similar user or group of users and exclude the under-represented. And training data may not include real-world context or may miss updated information, discoveries or human rights codes and legislation. I referred to the issue of sample size where larger data samples are prioritized ,yet, we know for foresight, small data is often very meaningful. And we know that algorithmic bias can be engineered into machines by programmers who hold explicit or implicit biases. And the question that designers always ask of who are you designing for? How do you imagine the user? This is incredibly important when we think about bias and the same is true for algorithms for recommendation engines. And there's often technical biases, that assume that human behavior will obey machine logic. Although we can say in many instances, it increasingly does. We've been trained by our machines. Let's obsessively, just look at our mobiles. Cathy O’Neil, a mathematician and former wall street luminary wrote a very interesting book called Weapons of Math Destruction. She argues that we must hold algorithms to the same standards as we hold humans. There are practical ways that AI scientists and designers are working together while we continue to address policy issues. And there's great initiatives like the IEEE Global Initiative for Ethical Consideration in Artificial Intelligence and Autonomous Systems. But we can play a role in designing tools that disclose data and decision paths and how algorithms work and that growing field is called Explainable AI and Tim Miller is an important kind of figure in this world. He's the co-director of the Centre of AI and Digital Ethics from the Australian University of Melbourne. And basically he just argues that members of the public, who are subject to automatic decisions regarding employment opportunities, college and university entry acceptance into that university of college, parole, investment in your community or non-investment transit decisions. And those who administer decisions, whether they're judiciary, governments and companies, have the right to understand why decisions are made. So how does the recommendation algorithm work and a pragmatic argument is that society will resist AI adoption unless there is trust in its accuracy and fairness. And finally, the more humans interact with AI algorithms, those of you in this space totally know this, the more accurate and nuanced the data and explanations can be. So this sort of human in the loop argument brings together design thinking and other design tools like UI and UX and visualization. So interaction design, and user experience design. And I'm going to go on about visualization here. So in data visualization, an analyst starts with the data set. And then you may question the source of data, but not how the algorithm that analyzes and restructures the data makes the decisions. Computational visualization is different, it aims to understand and show how the algorithm works. And these are just examples of data visualization from something called Compara that we built in our lab that looks at different visualization strategies. But with computational visualization, you try and understand how the algorithm works. And why it produces the results it produces, you're not looking at the data. So there's some interesting leaders in this field. Catherine Griffiths is concerned with data fluency and misreadings of facial recognition. Data and surveillance algorithms. And she sees the algorithm. She is a designer by training, computationally literate. She argues that the algorithm is a design tool. And we need to look at aesthetic and visual dimensions and explore alternate expressions and she's popularized the term slow computation. So reducing computational speed to human scale, to see processes in real time, being able to visualize the rule sets, not datasets. So you see structures of an algorithm. And pulling layers off of animations to disclose the execution of an algorithm. And she proposes using visualization tactics to reveal the logic of programming languages or a set of instructions or rule set . And we know that algorithms can produce their own data during the computational process. So just very quickly, examples of her work Visualizing Algorithms is one and it discloses the workings of a decision tree classifier you can manipulate and interact with it to see the frame of a machine-learning decision that usually occurs behind human perception. The Nexus Automata I, and it shows the techniques of machine vision to reveal image processing analysis, which are used in surveillance technologies and using a cellular automata logic. Her partner is often a fellow named Mike Bostock who founded a company called Observable. But most importantly, you may know him because he created the open source visualization programming language D3.js and Bostock believes that if we look at the structure of algorithmic processes and rules, we will study how a system is composed and then simulate the system to execute with slightly different roles or with a different structure and think through what alternate outcomes are possible or how results can be manipulated. So here's just a few examples of his work. And I think what's important here is. He showed graph theory. Is that, both of them are really trying to help us be computationally, literate.

So, Explainable AI builds on the work of Griffith and Bostock and it spans interpretation. Explainable systems explain the relationships between inputs and the outputs of the system, but also transparency. So the need to describe inspect and reproduce the mechanisms through which a system makes decisions or learns to adapt to its environment and to the governance of the data used and created and how the decision was made. So like, kind of building on what we've just seen. And designers ask questions, like what information do human decision-makers using the AI models need? What are the demands for and types of explainability and how are these different across different kinds of applications? How can humans guide the process to imagine and improve the explainability and the accountability of the model? So there's a whole set of technical questions about whether the algorithm itself produces a model that is easily explained or whether you're going back like with deep learning and adding in an explanation after the recommendation or prediction is made, and this is called post-processing and it's much less accurate also, explaining local specific decisions tend to be more accurate than a general decision. So, visualization is really important here. And we're gonna again, run through a series of visualizations and hope not to torture those of you who this is new for or familiar with. But we just thought it would be interesting to see how important it is to use visualization. Saliency maps. are amongst some of the first models for deep learning. The brightness of a pixel is directed, is proportional to its saliency. So the class to which that object belongs. And this is work by Cynthia Wang and their image classifiers that she sieging. Partial dependence plots, and plots of individual conditional explanations are another way to look at this. You can see how this is applied. The field of Convolutional Neural Network or CNN is a way of analyzing visual imagery for image and video recognition. And classification and segmentation of images for medical image analysis to build recommender systems, natural language processing, financial time series and brain-computer interfaces. So these are really important applications and we're just showing you. This is include Luisa M. Zintgraf’s heat maps, John Stack’s heat maps. So heat maps are used a lot in this context. Guijin Wang used blurring strategy to look at outputs, to look at where to focus your attention andPaulo E. Rauber used reduction of the filters and the maps. Another one is Kanit Wongsuphasawat and his team, showing the model structure. So in all of these, I'll just underscore that the interaction from the user, really helps to rectify and adjust the explanation and retrain the model. If the user finds the explanation to be opaque and long. You're actually training the accuracy and you're training the coherence of the models. So humans remain really important. There's an example. We're just going to briefly show by two researchers Meena Devii Muralikumar and Mathhew J. Bietz which showed algorithmic selection in social media. Using visualization as a mean to facilitate user's awareness of how filter bubbles in Facebook and Twitter work. These are very simple visualizations showing different potential outcomes depending what filter was used. And just a point that many of the visualizations we shown were designed by computational experts and designers immersed in this space for other experts and are potentially impenetrable by those impacted by recommendations. Now, heat maps, not so much, if we look back on what we showed you, some of the system mapping, I think, people can understand with some sophistication. But we really need visualization research that can address domain experts and lay communities. Because you want feedback through interactivity, not just on the quality of the explanation, but the content of recommendations so the model can adapt itself. So there's sort of two ways of thinking about that machine learning.

We've been talking about designing for and with human users and just wanted to talk now about the ways that human computer interaction, HCI design, and A I can enable user testing and some of the really big change there. So you've heard again and again, through these workshops that designers routinely used data collection tools. We do that to understand users, to identify needs and markets and to test prototypes. These are increasingly automated and some tools are not driven by machine learning and certainly have gained a lot of traction during COVID-19 because we're doing so much of this work online, not face to face. So some of these simply support online interviews, stakeholder group interviews, usability testing, and then natural language processing tools and other kinds of tools we'll talk about in a minute, have been added to address the collected data and offer descriptive and thematic analysis. So coding and summary sentiment analysis and qualitative content analysis, typology building, these visual tools for mapping thematic relationships, showing keywords in context, something that is really important for sentiment analysis, for example, or effective computing, or to understand emerging themes, changes in subject or terminology over time. Just going back to what I said. And chart flow between team members undertaking analysis. So all of this is being automated. The most advanced applications of AI are in market research. The use of CNN, which we just talked about, for a facial effective expression coding of video interviews, to understand consumer responses when shown a product. So it's focus is on why some product concepts are more successful and others not. So of course, designers want to know this and market researchers do. And Delvinia is a Canadian company that has created a whole system, which is called Methodify. So it automates market research environments, and it supports large scale survey collection and it has a series of tools. So Chris is one of them which basically uses AI and machine learning to conduct one-to-one chat-based interviews at scale and their AI and machine learning tools are pretty good here, better than many chat bots you're going to encounter. But again, it's, trying to use this for qualitative and quantitative analysis and it replicates a moderated discussion. So it gets cues for followups, keeps the participants focused by going back to topic. Asks probing questions in the moment. Of course, pre-programmed, well, responsive to texts, and encourages participants to stay with the survey. They've added a virtual panel capacity. So these are models that are trained to mimic human behaviors and interests of real consumers. So they're aggregate video representations, and it has things like product concepts or ad copy. The personas are smart. They build their identity from bots that crawl the internet to pull together content to build their profile. And the pitch for using these tools has been that these kinds of tools won't steal a company's product ideas. They don't get tired, they don't have response bias and they have no desire to please the questioners. This is a really important space. We saw it in week two how important personas are to the design thinking process. Obviously automation requires, tests against bias and training data like GBA+ tests that look at the ways that data review can interact with different interactions between aspects of a social identity. So, complexity of humans. It's an interesting question because automation, again of panelists. They look diverse, but they don't have the lived experience of diversity that humans might own. Again, very interesting important development, somewhat of a a threat in fact, a transformation to traditional design practices. So we're just going to move on now to talk about it in our final segment, which is really looking at oh, we do have a moment here for questions. There's some good input here of different kinds of generative design tools. Thank you. And any comments that you might want to make here as we're moving through the space and we know that there's people who kind of work at the NRC very much in AI and machine learning and looking at human interactions. So before we move on to our grand finale, are there any comments or questions here. What are your thoughts on the strengths and weaknesses and limits of automating user consultation processes to choose new products and to test usability? I have a feeling the questions are going to come in as we start the next section, but we'd love to hear your thoughts on this. Because we know that there's real capacity around the virtual table in this space. And, your thoughts on visualization too and how to understand making those tools. It's the balance between simplifying visualization outputs and representing complexity. And always in visualization conversations, there's a debate about how much does the user need to know, how educated should they be around the tools and the visual interpretation of data. Those are ongoing big questions in this space.

So we're going to end with a consideration of the third wave of AI. One that pushes deep learning to its next level, Artificial General Intelligence or super intelligence. This is in the realm of speculative fiction. I'm currently reading Machinehood by S.B. Divya, which is a sort of dystopian version of that world, or complex version of this world because there's different options and scenarios around human machine intelligence. And it's also, very practical research in Canada. There's companies like Sanctuary AI led by Canadians, Suzanne Gildert and Geordie Rose. They're really quite advanced in terms of the kind of robotics capacity that they're building. And they do see this modeled on the human agency. And then there's deep learning research at University of Alberta, the Vector Institute, University de Montreal and also at the NRC and other places that pushes the boundaries of this capacity and AGI is often understood as human like intelligence, the reasoning and effective capabilities of humans, we've said this quite a while ago, and the processing speed of machines and physical strength way beyond the human. Although things like robotic grasping is still a hard problem. So the field of research is increasingly called upon to establish the standards of Artificial General Intelligence and deep learning and argues for values being at the center of design and relationships with super-intelligence. There's some really interesting literature here. Brian Christian, he wrote a very smart book called the alignment problem. So the gap between human values and machine learning. Jeff Hawkins, who's a major inventor has now focused all of his work on simulating the way that the neurocortex works in human intelligence. And he talks about the old brain, which sort of has emotion in it and the new brain, and how to segregate those in building predictive models. So trying to push the old brain away. You can model reference frames that are super smart, but free of human emotion, if we want that. In Human Compatible: Artificial Intelligence and the Problem of Control, Stuart Russell, again, major researcher and a computer scientist, programer, searches for appropriate and what he describes as negotiated containment for non-human super-intelligence, but argues that it's inevitable and could also save the world. And Mark Carney's book that just came out about a month ago. His book is called Value(s) Building a Better World for All. He proposes that we need to build values into AI and machine learning that reflect a sort of more human positive stakeholder capitalism in our society and culture. The big question is how do we agree on these values, these are design questions. And then of course, what does this mean computationally? Like how does one create this capacity? We've just got a few things to say. But we wanted to ask you what challenges and opportunities, if you're working in the space, you see in the design of artificial general intelligence. And if you're not working in this space. Okay we'll let you meditate upon that. We're going to come back at the end and talk some more about this.

We started with a recognition of indigenous knowledge and culture and some of the work in speculative and indigenous futures. And we're going to talk very briefly now about the work that indigenous designers, philosophers, and scientists are engaged with in thinking about the ways that we could create an intelligent machine that would also embody indigenous values. And then how to think about other relationships with forms of consciousness that others might build. There's a group that is led by Canadian and Hawaiian designers and artificial intelligence researchers, ML researchers. Through a series of workshops that are looking at indigenous protocol to AI. It that was initiated by Jason Lewis and funded by the Canadian Institute for Advanced Research and the US government and others. And they have a series of think tanks to look at the question of designing a kind of indigenous super-intelligence. In their article, Making Kin with the Machine, which was really early on in the process. Sort of before this grouping was created. Lewis, Noelani Arista, Archer Pechawis, and Suzanne Kite began to lay out an indigenous protocol to AI. In which humans self-discipline, as they described it, to other forms of life, including creating communications and covenants. And network theories, you know, really are at the core Indigenous philosophies. So viewing technologies as part of natural systems and then asking humans to be in an integrated relationship to these, not necessarily differentiating between a living thing and another natural thing. Drew Hayden Taylor, who's a really important indigenous writer in Canada and has also started to write science fiction. He said: believing that AI has a spirit does not necessarily mean anthropomorphizing it, since being alive and having a soul does not necessarily equate to being human Indigenous cultures. So a fundamental principle within these kind of considerations of AI by these Indigenous thinkers and designers is the diversity of AI itself and this related principle of empathy and kindness to those that are other, that are not of your own community. For the Hawaiians, there's a term which is kanaka maoli and it means attending carefully to power. So it's about power between oneself and another, and the need to think about the power between AI and humans and to consider and integrate AI into community, into kind of relationality and treat it with kindness, which is an interesting concept. If you think of training databases of what you would train intelligence on. And then Suzanne Kite, who is a Plains Cree describes the importance of appropriate kinship relationships. And she talks about an Indigenous development environment. So the creation of programming languages that are grounded in nēhiyaw nisitohtamowin, which is the language of the Cree people, including its sort of cultural philosophies, or the framework of other Indigenous peoples. So the goal is that Indigenous culture values are a fundamental aspect of programming choices. And Loretta Todd , who's a very interesting filmmaker, media artists, thinker, and really involved again with next generation technology. She founded the, IM4 VR/AR in Vancouver, in partnership Emily Carr University of Art and Design. And it's working on building an Indigenous development environment and tech ecosystem.

Through our research and connection with the NRC and putting this series together, we've had the honor of seeing the exemplary work, that the NRC undertakes with Indigenous communities, including the development of language systems. You also have your very Artificial Intelligence for Design Challenge program.. And we just thought that Lewis and his colleagues' compelling initiative is interesting to look at a time when Canada is reckoning with their own history in relation to Indigenous people. And Indigenous people are looking at a future where they are empowered as inventors, but also can bring those values to a kind of kinder and more inclusive Canada. But one that is technologically at the avant-garde. So this is our final workshop. And we've done a lot in the 360 minutes that we've spent together. We've shown that Canada leads in design capacity, there's major design industries and design research in universities and colleges. We've shown that in this century, designers have developed many strategies to address needs and solutions. And we reviewed lots of design examples, and we looked at the role of designers and adding sustainable development goals to constraints. We gave you tools for design thinking in imagining and planning solutions. And we're leaving you with a design thinking toolkit to empathize, define ideate, prototype and test. We've talked again about strategic foresight tools to imagine future possibilities that are grounded in data, and to understand the value of that for managing risks. And we've also talked about design as fundamental to effective communication. We talked a lot about data visualization as a means to ensure data fluency, critical thinking and explainability. And we've noted that design is ubiquitous, and excellent design is often invisible. We've talked about the disruption of design by AI and ML, yet the designer and our expertise remains an important means to keep human imagination and strategy in the loop. We hope we've convinced you that there are many points in your research for designers to integrate and methods to enrich planning, process and outcomes.

Fichiers à consulter
FichierFormatTailleDernière mise à jour
PDF8 Mo2021-08-19
MP4792 Mo2021-08-26
TXT875 octets2021-08-20
DOITrouver le DOI : https://doi.org/10.4224/40002673
AuteurRechercher : ; Rechercher : ; Rechercher : ; Rechercher : 1
Affiliation du nom
  1. Conseil national de recherches du Canada. Technologies numériques
FormatTexte, Object d'apprentissage
Sujetconception spéculative; vieillir chez soi; futurisme autochtone; intelligence artificielle; apprentissage par machine; design génératif; design industriel; planification urbaine et de l’environnement; biais algorithmique; IA explicable; collecte de données; visualization; créativité
Résumé
Date de publication
Date de création2021-06-30
Maison d’éditionNational Research Council of Canada. Digital Technologies Research Centre
Publication connexe
Voir les objets (3)
Traduction de
Voir les objets (1)
Langueanglais
Exporter la noticeExporter en format RIS
Signaler une correctionSignaler une correction
Identificateur de l’enregistrement89ed13ef-fc84-476c-9233-a2d852f69a5e
Enregistrement créé2021-08-18
Enregistrement modifié2021-09-03
Date de modification: