Machine Learning Guide

MLG 035 Large Language Models 2At inference, large language models use in-context learning with zero-, one-, or few-shot examples to perform new tasks without weight updates, and can be grounded with Retrieval Augmented Generation (RAG) by embedding documents into vector databases for real-time factual lookup using cosine similarity. LLM agents autonomously plan, act, and use external tools via orchestrated loops with persistent memory, while recent benchmarks like GPQA (STEM reasoning), SWE Bench (agentic coding), and MMMU (multimodal college-level tasks) test performance alongside prompt engineering techniques such as chain-of-thought reasoning, structured few-shot prompts, positive instruction framing, and iterative self-correction. Links • Notes and resources at ocdevel.com/mlg/mlg35 • Build the future of multi-agent software with AGNTCY Try a walking desk • stay healthy & sharp while you learn & code In-Context Learning (ICL) Definition: • LLMs can perform tasks by learning from examples provided directly in the prompt without updating their parameters. Types: • Zero-shot • : Direct query, no examples provided. One-shot • : Single example provided. Few-shot • : Multiple examples, balancing quantity with context window limitations. • Mechanism: • ICL works through analogy and Bayesian inference, using examples as semantic priors to activate relevant internal representations. Emergent Properties: • ICL is an "inference-time training" approach, leveraging the model’s pre-trained knowledge without gradient updates; its effectiveness can be enhanced with diverse, non-redundant examples. • Retrieval Augmented Generation (RAG) and Grounding Grounding: • Connecting LLMs with external knowledge bases to supplement or update static training data. Motivation • : LLMs’ training data becomes outdated or lacks proprietary/specialized knowledge. Benefit • : Reduces hallucinations and improves factual accuracy by incorporating current or domain-specific information. • RAG Workflow: • Embedding: 1. Documents are converted into vector embeddings (using sentence transformers or representation models). Storage: 2. Vectors are stored in a vector database (e.g., FAISS, ChromaDB, Qdrant). Retrieval: 3. When a query is made, relevant chunks are extracted based on similarity, possibly with re-ranking or additional query processing. Augmentation: 4. Retrieved chunks are added to the prompt to provide up-to-date context for generation. Generation: 5. The LLM generates responses informed by the augmented context. • Advanced RAG: • Includes agentic approaches—self-correction, aggregation, or multi-agent contribution to source ingestion, and can integrate external document sources (e.g., web search for real-time info, or custom datasets for private knowledge). • LLM Agents Overview: • Agents extend LLMs by providing goal-oriented, iterative problem-solving through interaction, memory, planning, and tool usage. Key Components: • Reasoning Engine (LLM Core): • Interprets goals, states, and makes decisions. Planning Module: • Breaks down complex tasks using strategies such as Chain of Thought or ReAct; can incorporate reflection and adjustment. Memory: • Short-term via context window; long-term via persistent storage like RAG-integrated databases or special memory systems. Tools and APIs: • Agents select and use external functions—file manipulation, browser control, code execution, database queries, or invoking smaller/fine-tuned models. • Capabilities: • Support self-evaluation, correction, and multi-step planning; allow integration with other agents (multi-agent systems); face limitations in memory continuity, adaptivity, and controllability. Current Trends: • Research and development are shifting toward these agentic paradigms as LLM core scaling saturates. Multimodal Large Language Models (MLLMs) Definition: • Models capable of ingesting and generating across different modalities (text, image, audio, video). Architecture: • Modality-Specific Encoders: • Convert raw modalities (text, image, audio) into numeric embeddings (e.g., vision transformers for images). Fusion/Alignment Layer: • Embeddings from different modalities are projected into a shared space, often via cross-attention or concatenation, allowing the model to jointly reason about their content. Unified Transformer Backbone: • Processes fused embeddings to allow cross-modal reasoning and generates outputs in the required format. • Recent Advances: • Unified architectures (e.g., GPT-4o) use a single model for all modalities rather than switching between separate sub-models. Functionality: • Enables actions such as image analysis via text prompts, visual Q&A, and integrated speech recognition/generation. Advanced LLM Architectures and Training Directions Predictive Abstract Representation: • Incorporating latent concept prediction alongside token prediction (e.g., via autoencoders). Patch-Level Training: • Predicting larger “patches” of tokens to reduce sequence lengths and computation. Concept-Centric Modeling: • Moving from next-token prediction to predicting sequences of semantic concepts (e.g., Meta’s Large Concept Model). Multi-Token Prediction: • Training models to predict multiple future tokens for broader context capture. Evaluation Benchmarks (as of 2025) Key Benchmarks Used for LLM Evaluation: • GPQA (Diamond): • Graduate-level STEM reasoning. SWE Bench Verified: • Real-world software engineering, verifying agentic code abilities. MMMU: • Multimodal, college-level cross-disciplinary reasoning. HumanEval: • Python coding correctness. HLE (Human’s Last Exam): • Extremely challenging, multimodal knowledge assessment. LiveCodeBench: • Coding with contamination-free, up-to-date problems. MLPerf Inference v5.0 Long Context: • Throughput/latency for processing long contexts. MultiChallenge Conversational AI: • Multiturn dialogue, in-context reasoning. TAUBench/PFCL: • Tool utilization in agentic tasks. TruthfulnessQA: • Measures tendency toward factual accuracy/robustness against misinformation. • Prompt Engineering: High-Impact Techniques Foundational Approaches: • Few-Shot Prompting: • Provide pairs of inputs and desired outputs to steer the LLM. Chain of Thought: • Instructing the LLM to think step-by-step, either explicitly or through internal self-reprompting, enhances reasoning and output quality. Clarity and Structure: • Use clear, detailed, and structured instructions—task definition, context, constraints, output format, use of delimiters or markdown structuring. Affirmative Directives: • Phrase instructions positively (“write a concise summary” instead of “don’t write a long summary”). Iterative Self-Refinement: • Prompt the LLM to review and improve its prior response for better completeness, clarity, and factuality. System Prompt/Role Assignment: • Assign a persona or role to the LLM for tailored behavior (e.g., “You are an expert Python programmer”). • Guideline: • Regularly consult official prompting guides from model developers as model capabilities evolve. Trends and Research Outlook Inference-time compute • is increasingly important for pushing the boundaries of LLM task performance. Agentic LLMs • and multimodal reasoning • represent the primary frontiers for innovation. Prompt engineering • and benchmarking • remain essential for extracting optimal performance and assessing progress. • Models are expected to continue evolving with research into new architectures, memory systems, and integration techniques.

45M

MLG 002 Difference Between Artificial Intelligence, Machine Learning, Data ScienceArtificial intelligence is the automation of tasks that require human intelligence, encompassing fields like natural language processing, perception, planning, and robotics, with machine learning emerging as the primary method to recognize patterns in data and make predictions. Data science serves as the overarching discipline that includes artificial intelligence and machine learning, focusing broadly on extracting knowledge and actionable insights from data using scientific and computational methods. Links • Notes and resources at ocdevel.com/mlg/2 Try a walking desk • stay healthy & sharp while you learn & code Data Science Overview • Data science encompasses any professional role that deals extensively with data, including but not limited to artificial intelligence and machine learning. • The data science pipeline includes data ingestion, storage, cleaning (feature engineering), and outputs in data analytics, business intelligence, or machine learning. • A data lake aggregates raw data from multiple sources, while a feature store holds cleaned and transformed data, prepared for analysis or model training. • Data analysts and business intelligence professionals work primarily with data warehouses to generate human-readable reports, while machine learning engineers use transformed data to build and deploy predictive models. • At smaller organizations, one person ("data scientist") may perform all data pipeline roles, whereas at large organizations, each phase may be specialized. Wikipedia: Data Science • describes data science as the interdisciplinary field for extracting knowledge and insights from structured and unstructured data. Artificial Intelligence: Definition and Sub-disciplines • Artificial intelligence (AI) refers to the theory and development of computer systems capable of performing tasks that typically require human intelligence, such as visual perception, speech recognition, decision-making, and language translation. ( Wikipedia: Artificial Intelligence • ) • The AI discipline is divided into subfields: • Reasoning and problem solving • Knowledge representation (such as using ontologies or knowledge graphs) • Planning (selecting actions in an environment, e.g., chess- or Go-playing bots, self-driving cars) • Learning • Natural language processing (simulated language, machine translation, chatbots, speech recognition, question answering, summarization) • Perception (AI perceives the world with sensors; e.g., cameras, microphones in self-driving cars) • Motion and manipulation (robotics, transforming decisions into physical actions via actuators) • Social intelligence (AI tuned to human emotions, sentiment analysis, emotion recognition) • General intelligence (Artificial General Intelligence, or AGI: a system that generalizes across all domains at or beyond human skill) • • Applications of AI include autonomous vehicles, medical diagnosis, creating art, proving theorems, playing strategy games, search engines, digital assistants, image recognition, spam filtering, judicial decision prediction, and targeted online advertising. • AI has both objective definitions (automation of intellectual tasks) and subjective debates around the threshold for "intelligence." • The Turing Test posits that if a human cannot distinguish an AI from another human through conversation, the AI can be considered intelligent. • Weak AI targets specific domains, while general AI aspires to domain-independent capability. AlphaGo Movie • depicts the use of AI planning and learning in the game of Go. Machine Learning: Within AI • Machine learning (ML) is a subdiscipline of AI focused on building models that learn patterns from data and make predictions or decisions. ( Wikipedia: Machine Learning • ) • Machine learning involves feeding data (such as spreadsheets of stock prices) into algorithms that detect patterns (learning phase) and generate models, which are then used to predict future outcomes. • Although ML started as a distinct subfield, in recent years it has subsumed many of the original AI subdisciplines, becoming the primary approach in areas like natural language processing, computer vision, reasoning, and planning. • Deep learning has driven this shift, employing techniques such as neural networks, convolutional networks (image processing), and transformers (language tasks), allowing generalizable solutions across multiple domains. • Reinforcement learning, a form of machine learning, enables AI systems to learn sequences of actions in complex environments, such as games or real-world robotics, by maximizing cumulative rewards. • Modern unified ML models, such as Google’s Pathways and transformer architectures, can now tackle tasks in multiple subdomains (vision, language, decision-making) with a single framework. Data Pipeline and Roles in Data Science • Data engineering covers obtaining and storing raw data from various data sources (datasets, databases, streams), aggregating into data lakes, and applying schema or permissions. • Feature engineering cleans and transforms raw data (imputation, feature transformation, selection) for machine learning or analytics. • Data warehouses store column-oriented, recent slices of data optimized for fast querying and are used by analysts and business intelligence professionals. • The analytics branch (data analysts, BI professionals) uses cleaned, curated data to generate human insights and reports. • Data analysts apply technical and coding skills, while BI professionals often use specialized tools (e.g., Tableau, Power BI). • • The machine learning branch uses feature data to train predictive models, automate decisions, and in some cases, trigger actions (robots, recommender systems). • The role of a "data scientist" can range from specialist to generalist, depending on team size and industry focus. Historical Context of Artificial Intelligence • Early concepts of artificial intelligence appear in Greek mythology (automatons) and Jewish mythology (Golems). • Ramon Lull in the 13th century and Leonardo da Vinci constructed early automatons. • Contributions: • Thomas Bayes (probability inference, 1700s) • George Boole (logical reasoning, binary algebra) • Gottlob Frege (propositional logic) • Charles Babbage and Ada Byron/Lovelace (Analytical Engine, 1832) • Alan Turing (Universal Turing Machine, 1936; foundational ideas on computing and AI) • John von Neumann (Universal Computing Machine, 1946) • Warren McCulloch, Walter Pitts, Frank Rosenblatt (artificial neurons, perceptron, foundation of connectionist/neural net models) • John McCarthy, Marvin Minsky, Arthur Samuel, Oliver Selfridge, Ray Solomonoff, Allen Newell, Herbert Simon (Dartmouth Workshop, 1956: "AI" coined) • Newell and Simon (Heuristics, General Problem Solver) • Feigenbaum (expert systems) • GOFAI/symbolism (logic- and knowledge-based systems) • • The “AI winter” followed the Lighthill report (1970s) due to overpromising and slow real-world progress. • AI resurgence in the 1990s was fueled by advances in computation, increased availability of data (the era of "big data"), and improvements in neural network methodologies (notably Geoffrey Hinton's optimization of backpropagation in 2006). • The 2010s saw dramatic progress, with companies such as DeepMind (acquired by Google in 2014) achieving state-of-the-art results in reinforcement learning and general AI research. • The Sub-disciplines of AI • and other resources: AI on Wikipedia Machine Learning on Wikipedia Data Science on Wikipedia • Further Learning Resources Artificial Intelligence (Wikipedia) Machine Learning (Wikipedia) Data Science (Wikipedia) AlphaGo Movie AI Sub-disciplines

1T 4M

Episodes

Lyt når som helst, hvor som helst

Andre har også lyttet til ...