absorb.md

Surya Ganguli

Chronological feed of everything captured from Surya Ganguli.

paper / suryaganguli / May 25

LaNiSb₃ Single Crystals Reveal Anisotropic Multiband Transport and Topological Semimetal Candidacy

Single crystals of LaNiSb₃, grown via Sn flux and confirmed in the orthorhombic Pbcm space group, exhibit metallic resistivity from 3–300 K alongside anisotropic, positive magnetoresistance reaching ~8% for H∥b. Hall resistivity and Kohler's scaling violation together indicate multiband (electron- and hole-type) charge transport, well captured by a semiclassical two-band model. The pronounced twofold angular symmetry in MR across crystallographic planes and the quadratic-to-linear MR crossover position LaNiSb₃ as a structurally and electronically rich platform for probing topology–structure correlations in semimetals.

crystal-growthsemimetalsmagnetoresistancecondensed-matter-physicsmaterials-sciencetopological-materialsmultiband-transport
“LaNiSb₃ crystallizes in the orthorhombic Pbcm space group with lattice parameters a = 13.0970(2) Å, b = 6.1400(4) Å, and c = 12.1270(4) Å.”
paper / suryaganguli / May 25

Disentangled Generative Priors Enable Interpretable Uncertainty Separation in Bayesian Inverse Problems

Ganguli & Constantinescu propose a structured Bayesian prior built on a disentangled deep generative model whose latent space is explicitly partitioned into interpretable physical parameters and residual variability. By linearizing the generator, they derive conditions under which the posterior achieves approximate block-diagonal structure in latent space — meaning disentanglement at the representation level translates into genuine separation of uncertainty in the inverse problem. The method is evaluated on elliptic PDE inverse problems (conductivity and source identification), matching oracle Gaussian process priors under correct specification and outperforming them under misspecification, while yielding spatially calibrated uncertainty and interpretable parameter recovery.

bayesian-inferencedeep-generative-modelsinverse-problemsuncertainty-quantificationdisentangled-representationsmcmcprobabilistic-modeling
“Partitioning the latent space of a deep generative model into auxiliary (interpretable physical) and residual variables induces a hierarchical prior that separates domain-relevant uncertainty from unknown residual variability.”
paper / suryaganguli / May 25

TFGN Solves Catastrophic Forgetting at LLM Scale Without Replay, Task Labels, or Regularization

TFGN is an architectural overlay for transformer LLMs that enables continual pre-training across heterogeneous text domains by decomposing forward and backward passes: the forward pass remains fully dense, while cross-domain parameter updates are structured to avoid writing to prior-domain subspaces. Evaluated across six domains at 1B tokens per phase and three model scales (up to ~9B), TFGN achieves near-zero backward transfer (-0.007 at LLaMA 3.1 8B) and ≥99.59% L2-orthogonal gradient separation between domain pairs — all without replay buffers, task identifiers, or Fisher penalties. Notably, the architecture also yields positive forward transfer (e.g., 26.8% JavaScript perplexity drop from Python training alone) and supports two extensions: an autonomous meta-control layer and an operator-level latent planner.

continual-learningcatastrophic-forgettingllm-pretrainingtransformer-architecturecontinual-pretrainingdomain-adaptationparameter-efficient-training
“TFGN achieves catastrophic forgetting mitigation at LLM scale without replay buffers, task identifiers, or regularization penalties such as Fisher penalty.”