fixing shape-prior shortcuts in single-shot 3d scanning
a new network design prevents depth-from-boundary shortcuts in fringe projection profilometry by using a wrapped-phase representation and calibration layer.
topic
machine learning
12 ways to cut llm latency and inference costs
practical techniques to reduce large language model latency and inference costs in production by eliminating wasted work.
reward transport: steering molecular generation via noise alignment
a new method uses optimal transport coupling during flow matching training to align noise with molecular rewards, enabling controllable generation without extra models or gradients.
knowledge distillation works by sparsifying interactions in llms
a unified framework reveals that various knowledge distillation methods for large language models share a common mechanism: they force student models to use fewer interactions while zeroing out the rest.
outlines brings deterministic structure to llm outputs
outlines is an open-source library that enforces structured output from large language models by masking invalid tokens during generation.
five sql projects for a data portfolio
practical sql projects covering churn, warehousing, sales, banking, and healthcare to show real business problem-solving.
generative patient twin guides sepsis treatment at inference time
a new framework uses a generative ehr model as a patient digital twin to optimize sepsis treatment through model predictive control, adapting to changing goals without retraining.
signed symmetric quantization for few-bit integers
signed symmetric quantization uses the extra negative value in signed integers to reduce clipping error without runtime overhead.
fusing normalization into gemm and attention kernels
novel kernel fusion techniques hide up to 90% of normalization latency by integrating layer norm and rms norm with matrix multiplications and attention.
local video summarization with smolvlm2-2.2b
a pipeline that runs on consumer gpus to summarize videos by extracting frames and analyzing them with a small vision-language model.
profiling attention in pytorch
a walkthrough of profiling different attention implementations in pytorch, from naive to fused kernels, using torch.profiler to spot performance differences.
explaining temporal graph networks via memory backtracking
a new method explains temporal graph network predictions by tracing how past events shape node memory vectors through backtracking and topological attribution.
who gets missed in chest x-ray ai decisions
study shows how threshold choices in long-tailed chest x-ray models can miss rare-positive patients, especially in subgroups, and how tail-aware methods reduce underdiagnosis.
recolora keeps llm fine-tuning from forgetting past tasks
a new spectrum-aware method recursively consolidates low-rank adapters so each task builds on previous ones without overwriting them.
automate descriptive statistics in python
learn to build a repeatable pipeline for generating publication-ready summary tables from raw data using python tools.
run openclaw with ollama for private ai assistant
set up a private ai research assistant on telegram using openclaw and ollama, with web search and docker deployment.
conformal prediction fails minority classes in drug screening
standard conformal prediction under-covers rare classes in imbalanced drug discovery data, but a class-conditional fix restores coverage.
pytorch 2.13: flexattention on apple silicon, fused loss, new distributed backend
pytorch 2.13 brings flexattention to mps, a fused linear and cross-entropy loss, a new torchcomms backend, and more.