Second understanding: Illustrating artificial intelligence training and working process (Chinese flowchart)

In order to facilitate everyone to understand the process of AI training from the whole and avoid too many concepts affecting the overall understanding, we drew this flowchart.

Diagram of the AI training process

1. Data collection: Collect data for training and output raw data (unprocessed dataset)
2. Data preprocessing: Cleansing and formatting data, and outputting preprocessed data (data suitable for model training)
3. Model initialization: set the initial parameters of the model, output the initialization model (the model with the initial parameters)
4. Model training: Train a model with data and output the trained model (a model trained after one round)
5. Model evaluation: evaluate the performance of the model and output the evaluation results (performance indicators of the model)
6. Model optimization: Optimize the model according to the evaluation results, output the optimized model (optimized model)
7. Model deployment: Apply the model to the actual environment and output the final model (the model that meets the requirements)

The following figure is a simple diagram of the ChatGPT workflow:

ChatGPT workflow diagram

1. User input: The original input provided by the user.

Second, input processing: process the user's input and prepare it for feeding into the model.

1. Tokenizer: Breaks down the input text into small pieces (tokens) that can be understood by the model.
2. Tokenization: The process of converting input text into tokens.

ChatGPT model: Receive the processed input and start generating a reply.

1. Transformer model: The core part of the ChatGPT model for understanding input and generating replies.
2. Self-attention mechanism: A key part of the Transformer model for understanding the relationship between the various parts in the input.
3. Multi-head attention: Part of the self-attention mechanism that allows the model to focus on multiple parts of the input at the same time.
4. Zoom dot attention: a portion of the bull's attention, calculating the relationship between the input parts.
5. Generate replies: Generate responses based on understood input and learned patterns.

4. Output processing: Process the output generated by the model and prepare it for display to users.

1. Inverse tokenizer: Reassemble the tokens generated by the model into human-readable text.
2. Detokenization: The process of converting tokens back to text.

5. The user sees the reply: The user sees the final reply.

Sample token