ALLPCB
OpenAI's Defense Against LLM Attacks
OpenAI's study unveils an instruction hierarchy to boost LLM security against attacks like prompt injections, enhancing model safety.
Mixture-of-Experts (MoE) Models: Structure and Trade-offs
Overview of Mixture-of-Experts (MoE) transformers: sparse routing with gating networks and experts, training and inference trade-offs, and recent Mistral-8x7B-MoE.
FPGA Applications in Machine Learning
Overview of FPGA applications in machine learning: accelerating neural network inference, hardware quantization, algorithm optimization, and efficiency for edge AI deployments.
Latest Trends in Machine Learning Hardware Performance
Analysis of ML hardware trends across GPUs and accelerators, quantifying compute performance, interconnects, cost-performance, and energy efficiency.
Key Steps for Edge AI: Hardware Selection and Model Deployment
Edge AI technical overview covering distributed architecture, model lightweighting, data preprocessing, and model deployment to edge devices for real-time inference.
AI Applications in Wargaming: Case Studies
Analysis of AI applications in wargaming: case studies, benefits like enhanced scenario realism and decision support, and limitations such as black-box effects and cost.
Dynamic Memory Compression for KV Cache During Inference
Dynamic Memory Compression (DMC) compresses the Transformer KV cache online during autoregressive inference, improving throughput and enabling longer context windows.
Change Detection in Remote Sensing Using Foundation Models
Time-Traveling Pixels integrates SAM into remote sensing change detection, using low-rank fine-tuning and a Time-Travel Activation Gate to mitigate spatial-semantic domain shift.
What Is Automatic Speech Recognition and GPU-Accelerated ASR
Overview of ASR (speech-to-text): pipeline, acoustic and language models, CTC training, decoding strategies, and GPU acceleration using NVIDIA NeMo and toolkits.
GEAR: Efficient KV Cache Compression for LLM Inference
GEAR: hybrid KV cache compression combining 4-bit quantization, low-rank residual approximation, and sparse corrections to cut peak memory and boost inference throughput