Micah Goldblum’s Survey Offers a Deeper Look Into Deep Learning
In an era where Twitter’s extremities often eclipse nuanced academic discourse, a way to dig beneath the loudest and most viral opinions is extremely valuable. Micah Goldblum, a Postdoctoral Researcher at CDS, has created exactly that, in a recent survey intended to capture the multifaceted views of influential figures in deep learning. The preprint outlining the results of that survey, co-edited by Anima Anandkumar, Richard Baraniuk, and Tom Goldstein, is titled “Perspectives on the State and Future of Deep Learning.”
Goldblum’s work, the first in a planned series, aims to document diverse opinions in the field, particularly those not amplified by social media platforms. “You really see the most extreme opinions [on Twitter]. And since you’re seeing the posts most likely to go viral, these posts often don’t reflect the consensus or the opinions of influential researchers, many of whom don’t even post on Twitter,” Goldblum explained. This initiative counters the narrow lens through which machine learning’s trajectory is often viewed, opening a window to a broader spectrum of thoughts and insights from key players in the industry.
Curating a mix of perspectives from various backgrounds, the project strives to represent the views of the field on a range of topics that matter to researchers. “We asked topical questions, things we thought the community would be interested in hearing about,” Goldblum said. This approach not only diversifies the range of opinions presented but also ensures that the dialogue remains relevant and engaging.
Goldblum’s selection process for participants was meticulous, balancing the need for diverse viewpoints with the depth of each contribution. Goldblum and his team faced a choice: “either have tons of participants and allow each one to say very little, or have a smaller number of participants and allow them to say a lot,” Goldblum noted. The decision to opt for depth over breadth reflects a commitment to providing meaningful insights rather than superficial soundbites.
The survey has been met with great applause — somewhat ironically, very much so on Twitter. This makes sense, though, since it was designed to have broad utility. Goldblum said he’s observed significant interest from those in industry, for example, indicating the relevance of these discussions in practical applications. Questions on deep learning moving beyond academic benchmarking and the importance of interpretability in models resonate with those in the industry, who grapple with these issues in real-world scenarios and have to make financially high-stakes decisions about, for example, what kinds of models to build products on.
By chronicling a range of opinions over time, Goldblum’s projected series aspires to serve as a historical record of the evolution of thought in the field. “I think that documenting thought over time is really useful,” Goldblum reflected. Such documentation could provide valuable insights for future researchers and practitioners, offering a window into the shifting landscape of machine learning. Goldblum notes that his own research group has found it helpful to go back and read papers from the eighties and nineties, to see what people had thought in the past, and potentially recycle good ideas that simply hadn’t been executable then given the comparatively limited state of technological progress.
The collaborative nature of the project brought together a diverse group of editors, each contributing their expertise to shape the direction and content of the periodical. This collective effort underscores the multifaceted nature of machine learning, a field that thrives on the confluence of varied perspectives and disciplines.
Goldblum and his co-editors’ work represents a significant step towards a more comprehensive and representative understanding of the machine learning community. By moving the conversation beyond the echo chambers of social media, this series promises to offer a richer, more nuanced view of the field’s future trajectory. We look forward to many more iterations in the future.
By Stephen Thomas
