AIR Seminar with Jianyang Gu | Hariri Institute | Boston University

AIR Seminar with Jianyang Gu | Hariri Institute | Boston University

by Jianyang Gu · View the channel on PSI

Applied Sciences & Engineering Computing & Cybersecurity UComputing and Information TechnologyUYComputer science
🎙 Jianyang Gu 👥 1K 📅 October 13, 2026 ⏱ 57 min 👁 3 📄 expert opinion 🧭 2026-10-13
Available in: English (current) Français

Keywords

conceptrepresentationmachine learningscientific AIdata-centric

Summary

Jianyang Gu presents a seminar on shifting from sample-centric to concept-centric data representation in machine learning. He argues that treating samples as atomic units underuses the conceptual information they contain. The talk spans two main areas: scientific AI, where he introduces BioCLIP 2, a vision-language model trained on the Tree of Life 200M dataset covering ~1 million species, demonstrating emergent concept organization; and data-centric AI, where he examines how concept composition in datasets influences model behavior. He discusses method injection via generated captions to make representations more interpretable, and uses visual autoregressive models to analyze how different concept scales (coarse, mid, fine) affect classification. The talk emphasizes the importance of concept structure for interpretable scientific learning and highlights real-world applications in ecology and medicine.

124 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides substantial value by addressing a fundamental limitation in current ML—ignoring intra-sample concepts—and proposes concrete alternatives. The argumentation is coherent, moving from motivation (sexual dimorphism) to methodological innovations (BioCLIP 2, caption injection, VAR models) with supporting experiments. The speaker clearly articulates the trade-offs and presents quantitative results, such as accuracy improvements from amplifying coarse-scale features. The reasoning is logical and well-supported by examples, though some claims about emergent concepts could benefit from more direct comparison with prior work.

Scientific Rigor, Source Quality, Title Accuracy

The talk demonstrates high scientific rigor: datasets (Tree of Life 200M) are open-sourced, and methods are described in detail. The sources mentioned (e.g., Wikipedia, existing datasets) are credible, and the speaker acknowledges limitations such as label insufficiency. The title accurately reflects the content, focusing on the conceptual shift. However, as a seminar, it lacks formal citations and peer-review, and the single provided link (Harvard Institute events) is not central to the research. Overall, the presentation is rigorous for its format, with clear alignment between title and content.

182 words

Quality & Reliability

8/10

The presentation is well-structured, provides clear arguments, and references specific datasets and models. However, as a seminar talk, it lacks formal peer review.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Concurring Sources

  • Hariri Institute events — Link to the Hariri Institute events page, which may list future seminars and related information.

Contribution & Novelties

For more information, explore:

  • BioCLIP 2: An extension of BioCLIP trained on the Tree of Life 200M dataset, showing emergent concept organization.
  • Tree of Life 200M: A dataset covering ~1 million species for training vision-language models.
  • Visual autoregressive models: Used to examine concept composition and its effect on downstream classification.
  • Hariri Institute events page: https://www.bu.edu/hic/events/ for more seminars.

58 words

Radar Profile

The radar profile shows high scores across information quantity, quality, and reliability, with slightly lower technical level, suggesting a balanced presentation that is informative and trustworthy but accessible to a broader audience. The strength lies in the novelty and applicability of the concepts, reflecting a strong contribution to the field of data-centric AI.

Reliability 8/10