Day 3-Research in An Automated Future

—  Thank you everyone for inviting me to Advancing Research again

 

— This is my dip into the future, and discuss the birth of AI/ML Design research

 

— I’ve lead teams responsible for AI in platforms, products, and services
  • My work includes building tools for model developers

 

— I’ll begin with a design ethics statement: “To amplify the beauty of humanity with design while avoiding practices that exploit its fragility”

 

— I’ll tell you about myself as well:
  • Prior to CapitalOne, I was the lead design researcher at IDEO. I was both a researcher and designer and I consider design and research as a symbiotic relationship, not knowing where one begins or ends
  • I also teach Ethical AI at DePaul University

 

— My background includes 30 years of working with data and people
  • I’m a former journalist, and have been designing intelligent systems using IOT, automating healthcare appointments, and working in finance and banking

 

— I believe that in automated world, we need to determine what not to design, to preserve human culture and values, in our future world
  • We are very close to world where everything can be automated
  • There will be a shift to valuing more abstract things like trust and culture, in order to make the tradeoffs about what technology should not do

 

— But first some definitions
  • Design for me is an effort to impose order upon chaos and I consider design as a verb

 

— Research is design in my view, and I have strong opinion about research and design, as they are two sides of the same coin

 

— When I lead design teams, its difficult to understand where research ends and design begins, as they are so symbiotic
  • Research is an active form of design

 

— Now I’ll go through Machine Learning, AI, and MLXD, to provide an overview of what it is and what it’s not
  • Hopefully you’ll learn something new.
  • I also do workshops on ML + AI and how it fits with design

 

— Machine learning is a category of computer science where computers learn to achieved desired outcomes, through applying problem solving rules automatically
  • Some ML networks are ’neural’, mimicking the human brain, and have outcomes that are achieved without human programmers
  • Machine learning is limited to past observations though, rather than actively interacting with certain environments
    • It depends on data already collected

 

— Algorithms are series of unambiguous rules to solve problems
  • We create algorithms every day— such as grabbing an umbrella after it rains

 

— If you’ve taken HCI, you’ve seen the visual above
  • To summarize: There is input of sensory information, which you run through memory, and determine your behavior from there
— This is how we make decisions of how we want to act, and process the world around us

 

— Data science models work in the same way
  • There is input of historical/environmental data
  • The model is trained on the data, and runs it through validation to see if training accurate
  • Fine-tuning goes on every time the model is trained in tasks, such as recognizing a cat from a dog
  • Model developer will tune model, and keep training until it passes validation results
— We will then look at results of the model, with aim of providing a high-level accuracy for developers

 

— So what does machine learning look like in practice?
  • Detecting anomalies that stand out (seeing 100 degree temperatures in January)
  • Classifying family photos and placing them in same album
  • Recommendation models such as those on Netflix or Amazon

 

 

— To bring machine learning models to life, I’ll use example of a dog I took care of for seven days, Silver
  • Supervised Learning: Teaching  the dog to pick-up bones and giving positive reinforcement
    • I gave the outcome that I wanted and trained to give outcome all the time
  • Unsupervised Learning: Letting the dog find the pattern and bring it back
    • This was analogous to reviewing data-sets and surfacing a pattern
  • Reinforcement: I would leave the dog crate open and drop treats in there to encourage dog to stay in the crate
    • This would train model to respond to right cues, and negative rewards
    • This model is used a lot in self-driving cars

 

— We do AI/ML in real-world as people
  • Computer vision corresponds to sight
  • Auditory sensors corresponds to hearing
  • Natural Language Processing  corresponding to speaking
  • Automation corresponds to how we act
  • Recommendation Models corresponds to how we make decisions

 

— But AI/ML does have limitations
  • It needs data to function
  • Bad data leads to bad outcomes
  • No rules/No action, at least not yet
  • Nuance is the enemy of AI
  • The obstacle of human irrationality, which ML/AI can’t replicate

 

 

— As user researchers, designing for AI is a little bit different than designing for software/enterprise/interaction
  • Ways for how things are processed is different
— Current tech
  • Current technology is static, and doesn’t change over time
  • Interactions are contained between user and device
  • The user controls the device, and that user has total control over device
  • Interaction is one way
  • Technology is task-based, like clicking on a link
  • There are affordances where people interact in the same consistent way
  • Technology is static and performs task in same way every time
— For AI/ML
  • The user or the machine can be in control
  • Multi-agency context of use, where both machine and I can act
  • Things are decision-based, not task-based, which makes things harder
    • Affordance can change over time as machine can learn
  • Technology is dynamic and ML’s performed differently with new knowledge
    • Google searches evolve based on information you feed into it

 

— Challenges for research in AI system.
  • It’s future oriented instead of present oriented
  • Speculative research is hard with current UXR methods
    • Sometimes it is hard to realize all  the things machine can do for UXRs
    • Technology isn’t static
    • It goes beyond simple agency framework

 

— Pre-software focuses on how a product is made, while post-software focuses on how product behaves
  • Need to seek unexpressed rituals, cultures and values

 

— I use the framework of design anthropology for an automated world
  • Its a hybrid mode of investigation, to overcome speculative objects

 

— Design anthropology puts both ethnography and anthropology together and get design anthropology
  • Focuses on how design translates human values into tangible experiences
  • There is trust between machine and human interaction

 

— Characteristics of design anthropology include
  • Trans-disciplinary work
  • Multi-agency and requiring co-participatory design to put people in speculative scenarios and situations to figure out how it works
  • Research-led in order for us to focus on what needs are and go backwards

 

— Methods
  • Provocation prototypes to capture what people think about tools, rather than surveying
  • Perpetually synthesizing things
  • Future-oriented
  • Highly considerate of value orientation needs

 

 

— So I’ll conclude for now and will take questions

 

Q&A
  1. What’s the way forward for ML algorithms becoming more transparent? What skillset expansion is needed? 
—> This is key, which is why I explained how ML works at a conceptual level
  • As UXR you need to have high-level data and AI + ML literacy, as you can’t impact model once it’s made
    • Need to impact model in the data collection phase of the process
—> No model runs well without data and you need to be literate about how data is used in models
  • There are many kinds of cognitive biases and those that are not mathematical, which you can’t tweak in model development process
—> Need to look out for, and think about what goes into the hood
  • i.e. What’s missing in the data-set like FICO scores, and alternative ways to assess creditworthiness
    • Consider sins of omissions, and who is excluded from the model
    • Start at the beginning to see if what goes into a model is fair, and mitigate bias as much as possible
—> As far as transparency, have clarity at every level
  • It’s hard to explain how model works in-depth for some people, and hard to trace and audit it
  • You need stage gates to make model building process a glass-box versus a black-box
  • Having model governance to have someone audit the model development process to make sure the model adheres to our values
    • Explainability is constantly worked on at Capital One