# Dec 26, 2024

## JaVAD - Just Another Voice Activity Detector

**JaVAD** is a **custom conv-based voice activity detector** that outperforms existing open-source VADs, utilizing a **sliding window approach** over mel-spectrograms for enhanced speed and efficiency.

## Fine-Tuning 175B Parameter Language Models on a Single Consumer GPU through Optimized Memory Management

**Fine-tuning 100B parameter models** on a single consumer GPU is now feasible due to **optimized memory management** and NVMe SSD utilization, significantly enhancing accessibility for individual researchers and smaller labs.

## Could "activation engineering" replace prompt engineering or fine-tuning as a technique for steering models?

**Activation engineering** manipulates activation vectors in LLMs, exemplified by the "Golden Gate Claude," which altered neuron activity to embed the Golden Gate Bridge concept into model responses.

## Teaching VLMs to Convert Handwritten Images into Digital Ink with Read and Write Tasks

**InkSight** effectively teaches Vision-Language Models to convert handwritten images into digital ink, achieving **87% validity** and **67% indistinguishability** from human-generated ink in blind studies.

## Discussion: SOTA for implicit feedback in recommender systems

**Recent advancements** in handling **implicit feedback** in recommender systems focus on **Logistic Matrix Factorization** and **Collaborative Filtering**, which enhance recommendation accuracy by leveraging vast amounts of user interaction data. [Logistic Matrix Factorization for Implicit Feedback Data](https://stanford.edu/~rezab/nips2014workshop/submits/logmat.pdf) | [Collaborative Filtering for Implicit Feedback Datasets](http://yifanhu.net/PUB/cf.pdf).

## Terabyte-Scale MoEs: A Learned On-Demand Expert Loading and Smart Caching Framework for Beyond-RAM Model Inference

**Terabyte-scale Mixture-of-Experts (MoEs)** leverage a _learned_ loading mechanism to dynamically manage expert models, enabling inference beyond traditional RAM limits while optimizing performance.

## Inverse Design of Complex Nanoparticle Heterostructures via DL on Graphs

**Deep learning** is utilized to optimize the **nonlinear optical properties** of core-shell upconverting nanoparticles (UCNPs), achieving a **6.5 times increase** in predicted emission under 800nm illumination compared to existing structures.