MIT's McDermott lab: humans infer physical properties from sound using internalized statistics of how objects make sound
JoshHMcDermott · x · 2026-11-10
A thread-length summary of research from Josh McDermott's MIT lab: perceiving physical variables (mass, material, damping) from impact sounds is an ill-posed scene analysis problem that humans solve with internalized knowledge of real-world object acoustics.
Key points:
- The team measured impulse responses from hundreds of everyday objects (glass, ceramic, metal, wood, plastic, cardboard), extracting resonant frequencies, decay rates, and amplitudes that vary systematically with material and size
- These measurements were used to fit probabilistic generative models that sample realistic impact sounds given physical variables
- In experiments, listeners judged the mass of balls dropped on surfaces (n=101/118) or whether an object was damped by being held (n=55); sampling resonances from "lesioned" generative models impaired performance only when the two surfaces differed
- Conditions causing the greatest impairment had the lowest likelihood under the intact generative model, consistent with humans constraining physical inferences with knowledge of the real-world distribution
More from Research
- Postdoc Opening in France on Theory of Large-Context Transformers — vaiter · 2026-11-10
- Postdoc openings in France on theoretical analysis of large-context Transformers — vaiter · 2026-11-10
- safeagents live RL runs: decision models cut false blocks toward reward-hacking-free training — yb2698 · 2026-11-10
- safeagents judge bench v2: 661 labeled agent tool calls to score safety judge models — yb2698 · 2026-11-10
- New ContactRich Benchmark Targets Objects That Look Alike but Behave Differently — leto__jean · 2026-11-10
- Using Formal Proofs to Guide AI Software Optimization — bigdata · 2026-11-10