DL Seminar | "Asynchronous Workflows for Maintaining Hardware" &"Algorithmic Monoculture: Modeling Risks and Benefits"

Individual reflection by Hongjin Quan (scroll below).

By Hongjin Quan

Cornell Tech

In the seminar on Feb 29, 2024, we’re very excited to have two guest lecturers, both of whom are DLI Doctoral Fellows, to share with us their exciting works in the field of computer science and information science.

Algorithmic Monoculture: Modeling Risks and Benefits – Kenny Lucien Peng [1]

In the realm of decision-making, the reliance on a single algorithm by multiple evaluators introduces the concept of algorithmic monoculture. Kenny Lucien Peng’s discussion sheds light on this phenomenon, highlighting both the potential pitfalls and advantages of such uniformity in evaluating applicants. Leveraging the Gale-Shapley stable matching model [2], Peng’s research offers a theoretical lens to explore the dynamics of a two-sided matching market, comparing the outcomes of a singular evaluation approach (monoculture) against diverse, independent methods (polyculture). Peng’s findings reveal three critical insights:

(1) Choose less desirable applicants when errors in the evaluation process are minimal.

In environments with many firms and polyculture, where each firm independently evaluates applicants, the market dynamic tends to favor the hiring of the most preferred applicants, particularly when the noise in their evaluation processes has lighter-than-exponential tails. Despite individual firms operating under noisy conditions, collectively, they achieve outcomes akin to having perfect knowledge of applicants without sharing information. Hence, polyculture can lead to superior outcomes for firms compared to monoculture, even when the latter relies on a shared, less noisy algorithm.

(2) Increase the likelihood of applicants being matched with their preferred choice, although this may not benefit all individuals depending on their perceived value and risk tolerance.

Studies find that, on average, applicants fare better under monoculture, being matched with firms they prefer more often. However, the suitability of monoculture or polyculture for an individual applicant depends on their specific value to firms. Some applicants may have a higher chance of matching with their top choice in a monoculture setting but face a greater risk of not matching at all, with their preference potentially influenced by their aversion to risk.

(3) Offer better resilience against the effects of varying application volumes.

By considering variations in application opportunities among applicants, a factor particularly relevant in contexts like college admissions where application costs can be prohibitive. Peng and his team discovered that applicants who submit a higher number of applications tend to benefit more from polyculture. However, disparities in the number of applications submitted can exacerbate welfare issues more significantly under polyculture than monoculture, making the latter more resilient to variations in application behavior.

Reflections on Digital Life

The discussion of monoculture’s impact on decision-making in Peng’s talk invites us to reconsider our digital infrastructures. It prompts questions about the algorithms that mediate our digital lives and the value of incorporating multiple perspectives to avoid homogenization of content, opportunities, and ideas. As we navigate the complexities of the digital age, Peng’s insights offer a crucial lens through which to evaluate the algorithms that increasingly govern our choices, interactions, and information landscapes.

In essence, Peng’s work not only contributes to the academic discourse on algorithmic decision-making but also serves as a pivotal reflection on our digital society’s structure. It highlights the need for a balanced approach that values diversity and resilience in the face of rapid technological advancements and changing digital behaviors.

Asynchronous Workflows for Maintaining Hardware – Amritansh Kwatra [3]

Do we fix the hardware or just dispose of it when it’s not working? This question arises to many people when their digital devices go broken. Unlike “stackoverflow” or other online programming platforms, where people can find code and solutions just by inputting the error message, it’s hard for users to describe the broken parts of digital devices, not to mention get accurate instructions on such platforms. The journey to acquiring the necessary skills for effectively maintaining these machines involves a complex blend of hands-on experience, access to specific equipment, and the availability of materials. Inspired by “stackoverflow” and the 3D Gaussian Splatting technique, Amritansh Kwatra introduced us to his work — “SplatOverflow”, an asynchronous, context-rich, and demonstrative platform that allows users to maintain or fix hardware. By performing the following steps, SlpatOverflow can help users maintain their devices as easily as on stackoverflow:

(1) Add Tags

The first step in using “SplatOverflow” for device maintenance involves placing specialized tags around the device in question. These tags are designed for the platform’s cameras to accurately locate and identify the modeling positions on your device. It’s crucial to position these tags near the problem areas or across the device to cover various angles. These tags act as markers for the platform’s imaging and modeling tools, enabling a detailed and precise capture of the device’s current state.

(2) Scene Capture

Utilize your device’s camera to capture images or videos of the problem area. “SplatOverflow” encourages users to provide a visual context to their queries. This can be especially helpful when it’s challenging to describe the issue in words. For more complex problems, multiple angles or a video demonstrating the malfunction may be necessary. This visual documentation makes it easier for community experts to diagnose the issue and suggest accurate solutions.

(3) Gaussian Splatting with CAD

After uploading your visuals, “SplatOverflow” employs a unique feature called Gaussian Splatting combined with Computer-Aided Design (CAD) models. This process involves overlaying your captured scene with relevant CAD models to highlight discrepancies or damages in a more detailed and technical view. It helps in pinpointing the exact location and extent of the physical damage or wear. Users can interact with these models, adjusting them to match their hardware’s specific configuration and problem areas, providing a precise and tailored troubleshooting experience.

(4) Workflow

Finally, based on the information provided through tags, scene captures, and the Gaussian Splatting technique, “SplatOverflow” generates a step-by-step workflow for repairing or maintaining the device. This workflow includes detailed instructions, recommended tools, necessary replacement parts, and safety precautions. The platform also allows users to ask follow-up questions or request further clarification from the community.

Reflections on Digital Life

Kwatra’s work on “SplatOverflow” not only addresses a practical need but also reflects a broader trend towards self-sufficiency and sustainability in our digital lives. As digital fabrication technologies like 3D printing become more accessible to hobbyists and home users, the skills for maintaining and repairing these devices grow in importance. “SplatOverflow” embodies this shift, empowering users to extend the lifespan of their devices and reduce electronic waste. Moreover, by democratizing the maintenance process, it fosters a community of knowledge sharing and innovation, reminiscent of the open-source movement that has transformed software development.

References:

[1] Peng, Kenny, and Nikhil Garg. “Monoculture in Matching Markets.” arXiv.Org, 15 Dec. 2023, arxiv.org/abs/2312.09841

[2] “Gale–Shapley Algorithm.” Wikipedia, Wikimedia Foundation, 30 Jan. 2024, en.wikipedia.org/wiki/Gale–Shapley_algorithm

[3] Subbaraman, Blair, and Nadya Peek. “3D printers don’t fix themselves: How maintenance is part of Digital Fabrication.” Proceedings of the 2023 ACM Designing Interactive Systems Conference, 10 July 2023, https://doi.org/10.1145/3563657.359599