Stanford University

Project Team: Christoph Lippert (HPI), Sanmi Koyejo (Stanford) 

In today's world, hundreds of daily decisions are supported by systems that have learned to make these decisions based on past data. This ranges from daily decisions about what to buy, whom to meet and talk with to the complex task of getting health advice and medical diagnosis. While modern methods of machine learning (ML) have reached human-level accuracy, how to make the predictions and decisions taken by an ML system explainable to the person affected is an open question.

We propose developing objective, machine-optimizable criteria for predictions and decisions, making them understandable to humans. We will start by gathering survey data and conducting in-situ Interviews with those affected. Using methods from machine learning and econometrics itself, the idea is to infer a measurement function of task- and user-specific “explainability" of a prediction and decision from the interview data. Since collecting such feedback is onerous and time-consuming, the methods considered need to be very data efficient and have causal strength in their prediction power to permit predictive accuracy and the use of the learned functions as objective function that can be optimized in a domain-specific ML task to jointly optimize for accurate and human-understandable predictions and decisions.

One testbed for this theme will be in explainable medical advice. As sensors for the human body and compute have evolved significantly over the past 10 years and are broadly being adopted, it is now feasible for a doctor to augment their decisions on preventive and symptomatic treatments with complex data from scans (e.g., X-ray, CT) or body-measurements (e.g., heart rate, blood oxygen, pace, vertical oscillation). Moreover, smart sensing and inference systems for in-hospital or in-home medical care are also becoming more commonplace today. However, decision support for doctors and patients based on such complex time-series data is often difficult to understand, in particular when patients are interested in simple cause-effect relationships for their personal health. The aim of the investigation is how to blend expert (medical doctor) knowledge with the advice that is based on patients' data to help both the doctor and patients gain trust. Additional areas for investigation include applications and interfaces for multiple types of users so as to meet the needs of doctors, nurses, patients, and their families. The goal is to also focus on how to gain trust when giving and receiving algorithmic health advice.

We further aim to more broadly re-think the ML pipeline from a user-centered design perspective. Today, ML explainability approaches base their explanations around properties of the item in question, aiming to communicate how an item's features or content led the model to make its decision. However, for ML tasks ranging from online comment toxicity to misinformation detection to medical diagnosis, different groups in society may have irreconcilable disagreements about ground truth labels, so which voices a model learned from during training plays a significant role in the decisions a model makes. We propose an approach that affords a new, complementary lens to machine learning explainability, in which we aim to explain a model's prediction as a function of the properties of its annotators. Second, we focus on model training. Today's technologies are driven by a culture of optimization, but we're increasingly aware that today's optimization functions are woefully out of sync with human wellbeing. We propose to reimagine the way that optimization functions, and hence ML models, are designed by taking inspiration from sketching: a powerful technique leveraged by designers to rapidly explore a design space and gather feedback on novel ideas. Our work will investigate how to develop sketch-like versions of objective functions that similarly allow ML practitioners to rapidly explore the human impacts of their models.

Project Team: Niclas Boehmer (HPI), Michael Bernstein (Stanford)

Social computing systems are the computational systems that mediate our social interactions with one another; examples include social networks such as Facebook (messenger), WhatsApp, Instagram, TikTok, dating sites such as Tinder, OkCupid, or eHarmony, or gaming platforms such as Xbox Live or PlayStation Network. All of these services have in common that the connection between people is mediated by means of algorithms: the selection of who you interact with, chat with, share with, play with, or even date is based on algorithmic choices that influence or recommend the connections based on predicted social interactions. However, there are three challenging problems:

How can we assess the quality and bias of such services before they are launched in the real world? By its very definition, it's impossible to run A/B tests or re-evaluate the effect of the algorithms when changing parameters unless there are computationally efficient and accurate ways to simulate the effect and design of (people) these algorithms. Designing such a simulation must be both data-driven (based on experience) as well as embody basic patterns of social interactions that are inherent in human-human interaction.

How can we design social algorithms that make intelligent tradeoffs between values, goals, and objectives? How do we encode our moral objectives alongside our practical objectives into these systems, to prevent them from increasing polarization, enmity, or disinformation? How might different algorithms and social designs produce different emergent or societal-level outcomes?

How can we identify non-human participants in a social network? While the advent of the internet made it much easier for people to connect with each other (independent of time and space), it also opened the possibility for Al algorithms to imitate a human in said networks often called bots. However, people are (often) seeking interactions with real people and the existence of bots not only harms the legitimacy of the social network but can also be used for nefarious purposes by adversaries. In this line of research, we plan to investigate algorithms that can predict “humanness" based on the social trace that individuals leave in networks.

Project Team: Patrick Baudisch (HPI), Maneesh Agrawala (Stanford)

Builds on Foundational Theme: Objective Functions for Human-Understandable Explainability

Fabrication tools (e.g., 3D printers, laser cutters, machine tools, sewing machines) have tremendous potential for speeding up rapid prototyping and creating personalized physical objects. However, there are four problems that hinder the wide-spread adoption of these tools:

Current fabrication tools require manual setup and tedious calibration before they can be used to make objects that are well-crafted and fit well. This is usually a long and often frustrating process, which makes it all but impossible for non-experts to operate these machines. Some of these machines require frequent recalibration as the Input materials are changed (e.g., the printing medium, the hardness of the wood) or as usage over time leads to misaligned parts. We will start by developing open-loop calibration Al that learns how to adjust a fabrication tool (e.g., laser power to cut, kerf, laser power to engrave) from examples of material features (e.g., material type, material thickness, hierarchical machine type, laser type, maximum nominal laser power, maximum nominal machine speed).

Once the open-loop calibration works, we plan to build a closed-loop calibration Al that uses sensors (e.g., cameras, pressure sensors, audio) to dynamically maintain precise calibration throughout a fabrication session.

Designers typically develop designs using general purpose visual design tools such as Adobe Illustrator (or, in 30, SolidWorks). While numerous such general-purpose applications are readily available, their large number of tools and functions tends to limit their learnability. Our objective is to improve learnability by reducing tool sets to what is actually necessary and to replace complex tools with simpler tools, while, at the same time, still be sufficiently general to create unseen objects. Our goal is to develop Al-based program induction techniques that use sample documents as input and generate highly learnable special-purpose tool sets as output.

Fabrication tools cannot generally execute drawings made using Adobe Illustrator or 30 models made using SolidWorks. Instead, such designs first must be converted into a program such as g-code or a different tool-specific programming language that specifies the operations the tool must perform to generate the object, e.g. the cutting program for a laser cutter. We seek to develop Al-based program induction techniques that incorporate a complete specification of the capabilities and constraints of the fabrication tool and train on the interaction patterns experts use in vector and CAD drawing tools to automate the translation into a fabrication-ready program as the designer is creating the visual design.

Fabrication tools usually are not fully automated and instead require that a human perform various tasks such as preparing the materials, assembling fabricated parts, adjusting machine settings. To ensure that the human operator can perform these tasks in tight synchrony with the fabrication tool we will aspire to develop Al-based methods for generating human-readable plans that orchestrate the entire fabrication process. The plans or instructions will teil the human how to efficiently operate the fabrication tool(s) and how to assemble the parts. Together, the program induction for the fabrication tools (see above) and the human instructions will provide an end-to-end specification for rapidly prototyping personalized objects.

Project Team: Gerard de Melo (HPI), Monica Lam (Stanford), James Landay (Stanford) 
Builds on Foundational Theme: Design of Social Computing Systems

In our society, many needs in domains like mental health and education go unmet because of the limited availability of human experts like therapists or teachers. For instance, 1 in 44 children has been diagnosed with autism spectrum disorder in the United States, according to the CDC. Two-thirds of high-functioning and well-educated adults with ASO in Germany show high rates of unemployment. Pivotal Response Training has shown to be effective in improving social skills, but there is a lack of trained professionals. Can Al assistants be used to treat more individuals with ASO? In the US and Europe, over 27% of individuals over 60 live alone, and social isolation is associated with about 50%percent increased risk of dementia. Can an Al assistant provide the elderly with companionship and encourage them to stay active in order to improve their mental and physical health?

To handle such tasks, the virtual assistants need to be socially intelligent Besides being capable of following verbal orders, they must also have the ability to hold conversations about general topics, show empathy to the speaker, understand social norms and human values, reason with common sense, adapt to the user, commit relevant Information to memory, and plan ahead in order to form a meaningful long-term relationship with the user.

Our approach is inspired by the role of the prefrontal cortex {PFC) in the human brain: it performs high cognitive functions by sending top-down control signals to the perceptual, motor, and language cortices. We will develop a programmable executive controller that lets developers’ program high-level executive control over neural large language models (LLMs) to create socially intelligent virtual assistants. Our focus will include research on the fundamental concepts in socially intelligent agents. We plan to create a high-level programming language, called GenieScript, that lets developers all around the world create socially intelligent, conversational agents easily. We will investigate techniques to efficiently extend the linguistic capabilities of such agents to additional human languages such as German. We further aim to perform select. impactful, interdisciplinary research that also informs the development of the theory for such agents. Specifically, we plan to develop a virtual coach to teach people with autism spectrum disorder social skills and an assistant to provide companionship to the elderly.

Project Team: Ralf Herbrich (HPI), James Landay (Stanford)

Builds on Foundational Theme: Objective Functions for Human-Understandable Explainability

Today, most homes and offices are equipped with simple controls for light and temperature that only allow switching them on/off or setting a target room temperature. This leads to a significant waste of energy due to rooms being lit too long and kept too warm at times when people living the home are not present. Methods of Al are increasingly being considered to learn the usage patterns of people in homes and offices (both using mobile and flexed sensors for presence and identification) to make better predictions. Controls for homes need to also be advanced in lockstep with the sensors that perceive their presence (e.g., WIFI signal strength of their mobile phones) to ensure that people will continue to trust them as well as provide effective Inputs to the automated lighting and heating systems instead of overwriting using on/off switches or target temperature gauges. Further, these in-home/office sensing systems will soon sense our wellness states (e.g., stress) and activities (e.g., exercise). Using this information in these systems can help nudge us (e.g., ambient art displays to reward physical activity) or change our environment (e.g., music or lighting) in ways that improve our wellness. All of these intelligent sensing systems, especially in homes, must both be trusted to do the right thing and trusted to not violate our privacy. How do we create these systems in ways that private information is always maintained safely, and users will trust and use these systems without fear?

Similar issues will occur as we add similar types of sensing into health care domains, e.g., hospital rooms, eldercare facilities, and even for aging in place in private homes. These issues become even more acute as these systems rely on what is commonly called "ambient intelligence” (e.g., computer vision) to infer human activities in these spaces (e.g., "A patient is trying to get out of a hospital bed and might fall."). We plan to apply some of our early learnings on Al in home and office settings to the more complex healthcare domain.