Large Language Models(LLMs) Concepts

1、Introduction to Large Language Models(LLM)

1.1、Definition of LLMs

• Large: Training data and resources.
• Language: Human-like text.
• Models: Learn complex patterns using text data.

The LLM is considered the defining moment in the history of AI.

Some applications:

• Sentiment analysis
• Identifying themes
• Translating text or speech
• Generating code
• Next-word prediction

1.2、Real-world application

• Transforming finance industry: 

  [Investment outlook] | [Annual reports] | [News articles] | [Social media posts]
  
  --> LLM
  
  [Market analysis] | [Portfolio management] [Investment opportunities]

• Revolutionizing healthcare sector:

  - Analyze patient data to offer personalized recommendations.
  
  - Must adhere to privacy laws.

• Education:

  - Personalized coaching and feedback.
  
  - Interactive learning experience.
  
  - AI-powered tutor:
    - Ask questions.
    - Receive guidance.
    - Discuss ideas.

• Visual question answering:

  Defining multimodel:
  
  Multimodel:
  - Many types of processing or generation
  
  Nun-multimodel:
  - One type of processing or generation
  
  
  
  Visual question answering:
  - Answers to questions about visual content
  - Object identification & relationships
  - Scene description

1.3、Challenges of language modeling

• Sequence matters
• Context modeling
• Long-range dependency
• Single-task learning

2、Building Blocks of LLMs

2.1、Novelty of LLMs

• Overcome data's unstructured nature
• Outperform traditional models
• Understand linguistic subteties

The bulding blocks show below:

2.2、Generalized overview of NLP

2.2.1、Text Pre-processing

Can be done in a different order as they are independent.

• Tokenization: Splits text into individual words, or tokens.

• Stop word removal: Stop words do not add meaning.

• Lemmatization: Group slightly different words with similar meaning so we can reduce words to their basic form. For example, we can map them to their root word.

2.2.2、Text Representation

• Text data into numerical form.
• Bag-of-words:
  
   

  Limitation:
  
  - Does not capture the order or context.
  
  - Does not capture the semantics between the words.

• Word embeddings:

  

2.3、Fine-tuning

Fine-tuning:
- Addresses some of these challenges.
- Adapts a pre-trained model.


Pre-trained model:
- Learned from general-purpose datasets.
- Not optimized for specific-tasks.
- Can be fine-tuned for a specific problem.

2.4、Learning techniques

N-shot learning: zero-shot, few-show, and multi-shot.

2.4.1、Zero-shot learning

• No explicit training.
• Uses language understanding and context.
• Generalizes without any prior examples.

2.4.2、Few-shot learning

• Learn a new task with a few examples.

2.4.3、Multi-shot learning

• Requires more examples than few-shot.

3、Training Methodology and Techniques

3.1、Building blocks to train LLMs

3.1.1、Generative pre-training

Trained using generative pre-training

- Input data of text tokens.

- Trained to predict the tokens within the dataset.



Types:

- Next word prediction.

- Masked language modeling.

3.1.2、Next word prediction

• Supervised learning technique.
• Predicts next word and generates coherent text.
• Captures the dependencies between words.
• Training data consist of pairs of input and output examples.

3.1.3、Masked language modeling

• Hides a selective word.
• Trained model predicts the masked word.

3.2、Introducing the transformer

3.2.1、Transformer architecture

• Relationship between words.
• Components: Pre-processing, Positional Encoding, Encoders, and Decoders.

3.2.2、Inside the transformer

(1) Text pre-processing and representation:

• Text preprocessing: tokenization, stop word removal, lemmatization.
• Text representation: word embedding.

(2) Positional encoding:

• Information on the position of each word.
• Understand distant words.

(3) Encoders:

• Attention mechanism: directs attention to specific words and relationships.
• Neural network: process specific features.

(4) Decoders:

• Includes attention and neural networks.
• Generates the output.

3.2.3、Transformers and long-range dependencies

• Initial challenge: lone-range dependency.
• Attention: focus on different parts of the input.

3.2.4、Processes multiple parts simultaneously

• Limitation of traditional language models: Sequential - one word at a time.
• Transformers: Process multiple parts simultaneously (Faster processing).

3.3、Attention mechanisms

3.3.1、Attention mechanisms

• Understand complex structures.
• Focus on important words.

3.3.2、Two primary types: Slef-attention and multi-head attention

For example:

3.4、Advanced fine-tuning

3.4.1、LLM training three steps:

• Pre-training：
  
• Fine-tuning:
  
• RLHF:
  （1）Why RLHF?
  
  （2）Starts with the need to fine-tune
  

3.4.2、Simplifying RLHF

• Model output reviewed by human.
• Updates model based on the feedback.

Step1:

• Receives a prompt.
• Generates multiple responses.

Step2:

• Human expert checks these responses.
• Ranks the responses based on quality: Accuracy、Relevance、Coherence.

Step3:

• Learns from expert's ranking.
• To align its response in future with their preferences.

And it goes on:

• Continues to generate responses.
• Receives expert's rankings.
• Adjusts the learning.

3.4.3、Recap

4、Concerns and Considerations

4.1、Data concerns and considerations

• Data volume and compute power.
• Data quality.
• Labeling.
• Bias.
• Privacy.

4.1.1、Data volume and compute power

• LLMs need a lot of data.
• Extensive computing power.
• Can cost millions of dollars.

4.1.2、Data quality

• Quality data is essential.

4.1.3、Labeled data

• Correct data label.
• Labor-intensive.
• Incorrect labels impact model performance.
• Address errors: identify >>> analyze >>> iterate.

4.1.4、Data bias

• Influenced by societal stereotypes.
• Lack of diversity in training data.
• Discrimination and unfair outcomes.

Spot and deal with the biased data:

• Evaluate data imbalances.
• Promote diversity.
• Bias mitigation techniques: more diverse examples.

4.1.5、Data privacy

• Compliance with data protection and privacy regulations.
• Sensitive or personally identifiable information (PII).
• Privacy is a concern.
• Get permission.

4.2、Ethical and environmental concerns

4.2.1、Ethical concerns

• Transparency risk - Challenging to understand the output.
• Accountavility risk - Responsibility of LLMs' actions.
• Information hazards - Disseminating harmful information.

4.2.2、Environmental concerns

• Ecological footprint of LLMs.
• Substantial energy resources to train.
• Impact through carbon emissions.

4.3、Where are LLMs heading?

• Model explainability.
• Efficiency.
• Unsupervised bias handling.
• Enhanced creativity.