Feather Softwares’ Post

Detecting misusage of the SLUB allocator can be detected by machine learning algorithms and PoC is shown below that 3D visualized kernel allocation. Basically usage of the slab relies on patterns when it comes to the memory corruption exploitations.

Speculative decoding, a technique to accelerate LLM inference by employing a smaller approximation model in parallel to guess ahead of the main target model. https://lnkd.in/g_KU-tcJ

The first physical system to learn nonlinear tasks without a traditional computer processor https://buff.ly/3WSS6Ll

When you first hear of tensors, they sound pretty confusing - but they make more sense once you pick something and play a little. This example just loads an image of our pet gecko Leo (PNG) as a rank-3 tensor (height/width/rgba) and does some basic operations to lighten and rotate it. I thought it would be a nice intro for anyone who hasn't played with them before. Obviously they're built for advanced operations and GPU execution - but the basic data model is very relatable to working with multidimensional arrays/matrices. Note - I'm learning this myself; so take nothing here as best practice. :)

This illustrates well the modern approach to do unsupervised ML for tasks which were heavily dependent on labeled data before: - use a powerful general AI service to create synthetic labels - train a fast, efficient supervised ML model with those labels

Going to NeurIPS and want to trade notes? Some topics I am paying attention to: - Cost-optimizing large batch inference - Cross-platform model optimization methods - Image-gen observability - Computer vision & document processing Send me a DM if interested.

The approach to utilize #DeepSeek's reasoning capabilities for generating #synthetic labels to #finetune a #ModernBERT model: --- 1️⃣ Define the #Task and #Scope - Objective: Clarify the downstream task for ModernBERT (e.g., text classification, named entity recognition, question answering). - Data Requirements: Identify the unlabeled text corpus to process and the type of labels needed (e.g., sentiment labels, entity spans, QA pairs). --- 2️⃣ Leverage DeepSeek's #Reasoning for Label Generation a. Zero/Few-Shot #Prompting - Prompt #Design: Use structured prompts to guide DeepSeek in generating labels. - Output #Parsing: Extract labels from DeepSeek's responses using regex or keyword matching. b. Chain-of-Thought (#CoT) Reasoning - Request step-by-step reasoning to improve label reliability. - Use structured outputs (JSON, XML) for easy parsing. --- 3️⃣ Post-Processing Synthetic #Labels Noise #Reduction: - Remove samples with inconsistent/ambiguous labels. - Use confidence thresholds (e.g., discard labels with <80% self-consistency). Data #Augmentation**: - Apply paraphrasing (via DeepSeek) to expand the labeled dataset. - Example: "Rewrite this sentence in two different ways while preserving the label: {text}." --- 4️⃣ #FineTune ModernBERT #Curriculum Learning: 1. Train ModernBERT on high-confidence synthetic labels. 2. Gradually introduce lower-confidence samples. #NoiseRobust Techniques: - Use loss correction (e.g., [CleanLab](https://cleanlab.ai/)) to handle label noise. - Apply label smoothing or dropout to prevent #overfitting. #Hybrid Training: - Combine synthetic labels with a small #humanannotated dataset for semi-supervised learning. --- 5️⃣ #Validation and #Iteration #Evaluation: - Test ModernBERT on a gold-standard validation set to measure synthetic data quality. - Compare performance against baselines (e.g., training on human labels only). Iterative #Refinement: - Use ModernBERT's predictions to identify systematic errors. - Refine DeepSeek’s prompts or post-processing rules to address gaps. --- #Technical Considerations 1️⃣ Prompt #Engineering: - Optimize prompts for task specificity (e.g., domain-specific instructions). - Include examples in few-shot prompts to align DeepSeek with the target task. 2️⃣ #Scalability: - Parallelize label generation using batch inference APIs. - Cache responses to avoid reprocessing. 3️⃣ #Bias Mitigation: - Audit synthetic labels for demographic/cultural biases. - Add #fairness constraints during fine-tuning (e.g., adversarial debiasing). #Tools and #Frameworks - #Label Generation: DeepSeek API, OpenAI API, or self-hosted LLMs. - #Pipeline Automation: Apache Airflow or Prefect for workflow orchestration. - #Active Learning: Use ModernBERT’s uncertain predictions to prioritize human annotation.

11x faster and 11x less memory intensive LLM inference algorithm - preserving 95% of accuracy - from NVidia https://lnkd.in/gDBNy5R3

🎉 Course Completion! 🎉 I’m thrilled to share that I’ve completed the Computer Vision Onramp course! 🎓 This course has equipped me with valuable skills in computer vision, including image processing, object detection, and feature extraction. It’s been an exciting journey, and I can’t wait to apply these skills to future projects! Big thanks to the course creators and everyone involved. Onward to more learning and growth! 🌟 #ComputerVision #MachineLearning #AI #TechSkills #LearningJourney #ArtificialIntelligence #ContinuousLearning

An Introduction to Object Detection Part 1: Object Detection with Tiny Yolo V3 Object detection is a critical task in computer vision, enabling systems to identify and locate objects within an image or video. Among the popular object detection algorithms, YOLO (You Only Look Once) stands out for its real-time performance and accuracy. However, the standard YOLO models can be computationally intensive, making them less suitable for embedded devices with limited resources. This is where Tiny YOLO v3 comes into play. This is part one of a two part article in which we examine the Yolo approach to objection detection in detail. https://lnkd.in/ec65mgYF