AI, ML, LLM, and Agent

Introduction

Embedding

• google embeddings

Embedding is a vector representation of data in embedding space.

• one-hot encoding: A method of representing categorical variables as binary vectors.

  • Each category is represented by a vector with a single high (1) value and all other values low (0).
  • Example: For three categories A, B, C, the one-hot encoding would be:
    • A: [1, 0, 0]
    • B: [0, 1, 0]
    • C: [0, 0, 1]

• Static Embedding: A fixed representation of data that does not change based on context (Word2Vec, GloVe).

• Dynamic Embedding: A representation of data that can change based on context or input (BERT, GPT).

• Contextual Embedding: A type of dynamic embedding that captures the meaning of words or phrases based on their surrounding context (BERT, GPT).

Similarity

Transformer

Transformer consists of an encoder and a decoder.

• The encoder processes the input data and generates a representation of it.
• The decoder takes the encoded representation and generates the output data.

Architectures:

• Encoder-Decoder architecture is used in tasks like machine translation, where the input sequence is encoded into a representation, and the decoder generates the output sequence.
• Encoder-only architecture is used in tasks like text classification, where the input sequence is encoded into a representation, and a classifier predicts the output label based on that representation.
• Decoder-only architecture is used in tasks like text generation, where the model generates output sequences based on a given input prompt.

Self-attention mechanism allows the model to weigh the importance of different words in a sequence, enabling it to capture long-range dependencies and relationships between words.

encoders use bidirectional self-attention, while decoders use unidirectional.

• Foundation (Base/Pre-trained) LLM:
  • Pre-trained on large amounts of data and can be fine-tuned for specific tasks.
  • Examples: BERT, GPT-3, T5.
• Fine-tuned LLM:
  • A pre-trained model that has been further trained on a specific task or dataset to improve performance on that task.
• Distillation LLM:
  • A smaller model trained to mimic the behavior of a larger model, often used to reduce computational requirements while maintaining performance.