《Embrace AIGC》Part 4: OpenAI and GPT

皓月光兮非自明

《Embrace AIGC》Part 4: OpenAI and GPT [Copy link]

OpenAI

In 2015, a research institute was founded by Elon Musk, Sam Altman, Greg Brockman, Ilya Sutskever, Wojciech Zaremba, and John Schulman. It focuses on deep reinforcement learning (DRL).

Institutional Creed

To Ensure That Artificial General Intelligence Benefits All Of Humanity

Deep Reinforcement Learning

Deep reinforcement learning DRL is a combination of reinforcement learning (RL) and deep neural network, and is a subset of machine learning.

Research Results

time	Results	describe
2016	OpenAI Gym	A toolkit for developing and testing reinforcement learning
2018	GPT-1	Generative Model Architecture
2019	GPT-2	1.2 billion reference parameters
2020	GPT-3	175 billion reference parameters
2023	GPT-4	Pass the Turing Test
2024	GPT-4o	Deploy cross-text, audio and video reasoning models

Model significance

1. Save training time and training costs
2. Easy to use by engineers without data science or machine learning skills

The mathematics behind the model

The structure of RNN (Recurrent Neural Network)

The output of the RNN layer at time step tn is passed as input to the next time step. The hidden state of the RNN is also passed as input to the next time step, allowing the network to save and propagate across different parts of the input sequence.

x is the input at time t

U is the weighted input of the hidden layer h

h is the hidden layer at time t

V is the weighted output of the hidden layer h

y is the output at time t

Main limitations of RNNs

(1) Gradient disappearance and gradient explosion

It is multiplied many times during the gradient back propagation process, causing the gradient to become very small or very large.

(2) Limited context

The input sequence can only be processed one element at a time, so only a limited context can be captured.

(3) Parallelization is difficult

RNN is essentially a sequential execution, which makes it difficult to parallelize the calculations, and therefore cannot make good use of GPU parallel acceleration (Graphical Processing Unit)

Jacktang

In a recurrent neural network (RNN), the output at each time step depends not only on the input at the current time step, but also on the hidden state at the previous time step. The output of the RNN layer at time step tn can be calculated by