AI-900 정리노트 - Generative AI

How LLM Works

Tokenization

• Tokens are string with a known meaning
  

1. Start text want to tokenize
2. Split the words in the text based on a rule
3. Stop word removal
4. Assign a number to each unique token

Word Embedding

• Define the semantic relationship between words
• created during model training process
• represented as vectors
  • point in n-dimensional space

Machine Learning Model Architecture and RNNs

• Model architecture defines the structure for data processing, training, evaluation, and prediction generation.
• Recurrent Neural Networks (RNNs) were an early breakthrough in language model architecture.
• Context matters: The meaning of a word depends on the surrounding sentence context, not just the word itself.
• RNNs process sequential data:
  • Each step takes an input (a new word) and a hidden state (memory from the previous step).
  • Hidden states pass information forward, helping the model remember earlier parts of a sentence.
• Example: In a sentence about Vincent Van Gogh, remembering the painter’s name is essential to predicting the next word.
• [MASK] token: Used to indicate a missing word that the model needs to predict in NLP tasks.

Transformer Architecture Overview

• Transformer architecture has two main components:
  • Encoder: Processes input sequence and captures token context.
  • Decoder: Generates output sequence by attending to encoder output and predicting the next token.
• Key innovations:
  • Positional encoding
  • Multi-head attention
• In the encoder:
  • Input sequence → Positional encoding → Multi-head attention → Encoded representation.
• In the decoder:
  • Output sequence → Positional encoding → Multi-head attention → Combines with encoder output → Final output generation.

Understand Positional Encoding

• Position matters for meaning in text.
• Positional encoding adds position information to word embeddings.
• Combines word embedding vectors + positional vectors to encode both meaning and position.
• Simple index values (e.g., "The" = 0, "work" = 1) could be unstable; positional vectors encode position more meaningfully.

Understand Attention

• Attention replaces recurrence (RNNs) to process tokens in parallel, not sequentially.
• Attention mechanism maps new input (query) to learned knowledge (keys and values).
• Example:
  • Query: Vincent van Gogh
  • Key: Vincent van Gogh
  • Value: Painter
• In practice:
  • Query, Key, and Value are all encoded into vectors.
  • Scaled dot-product computes similarity between query and keys.
  • Softmax function produces a probability distribution over keys.
  • Highest-probability key determines the output value.

Multi-Head Attention and Efficiency

• Multi-head attention processes tokens multiple times in parallel.
• Enables different "views" of the same input for richer understanding.
• Parallel processing allows for faster and more efficient model training compared to sequential models like RNNs.

How to improve prompt

1. 어시스턴트에게 원하는 작업에 대해 구체적인 목표를 설정하세요.
2. 응답의 범위를 한정할 수 있는 출처를 제공하세요.
3. 적절성과 관련성을 극대화하기 위해 맥락(컨텍스트)을 추가하세요.
4. 응답에 대한 명확한 기대사항을 설정하세요.
5. 이전 프롬프트와 응답을 기반으로 반복 수정하여 결과를 개선하세요.

Key Mechanisms to Improve Generative AI Responses

• Grounding Data: Align outputs with factual, contextual, and reliable data sources.
  Methods include linking to databases, using search engines for real-time info, or incorporating domain-specific knowledge bases to enhance trustworthiness and relevance.

• Retrieval-Augmented Generation (RAG):
  Connects the model to an organization's proprietary database.
  Retrieves relevant information from curated datasets to generate contextually accurate and up-to-date responses.
  Useful for real-time applications like customer support or knowledge management.

• Fine-tuning:
  Further trains a pre-trained model on a task-specific dataset.
  Helps the model specialize in a domain, improve accuracy, and reduce irrelevant or inaccurate outputs.

• Security and Governance Controls:
  Manage access, authentication, and data usage.
  Helps prevent publication of incorrect or unauthorized information.

Metrics to Measure Response Quality

• Performance and quality evaluators:
  Measure accuracy, groundedness, and relevance of generated content.

• Risk and safety evaluators:
  Assess risks associated with content generation, ensuring the AI avoids harmful or inappropriate outputs.

• Custom evaluators:
  Industry-specific metrics tailored to meet specialized needs and goals.

• Modern services like Azure AI Foundry support these evaluation workflows and responsible AI practices.