2024. 1. 7. 09:09ㆍ코딩 도구/LG Aimers
LG Aimers: AI전문가과정 4차
Module 3. 『Machine Learning 개론』
ㅇ 교수 : 서울대학교 김건희
ㅇ 학습목표
본 모듈은 Machine Learning의 기본 개념에 대한 학습 과정입니다. ML이란 무엇인지, Overfitting과 Underfitting의 개념, 최근 많은 관심을 받고 있는 초거대 언어모델에 대해 학습하게 됩니다.
Recent Progress of Large Language Models
-GPT-3
: OpenAI’s third-generation Generative Pretrained Transformer
The first commercial product of OpenAI
-InstructGPT
GPT-3.5, Self-supervised language models does not necessarily follow a user’s intent
Key idea: fine-tune GPT-3 using human feedback
Reinforcement learning from human feedback (RLHF)
-Training of InstructGPT
Supervised fine-tuning (SFT)
Reward model (RM) training
RL via PPO
-ChatGPT
A sibling model to InstructGPT with conversational UI
Released on November 30, 2022
Fine-tuned from a model in the GPT-3.5 series, which finished training in early 2022
100M monthly active users (MAU) in 2 months – fastest growing Internet App ever
-GPT-4
A large multimodal language model
•Accept image and text inputs, and generate text outputs
• Released on March 14, 2023
-GPT-4 –Test on Benchmarks
Exams that were originally designed for humans
-GPT-4 –Visual Input
Accept a prompt of text and images
• Similar performance on text and photographs, diagrams, or screenshots
-Limitation
GPT-4 has similar limitations as earlier GPT models
• Not fully reliable; hallucinate facts and make reasoning errors
• Sensitive to the input phrasing or different answers to the same prompt
• Various biases in its outputs
• Lacks knowledge after September 2021
• Not learn from its experience
• Confidently wrong in its predictions, not taking care to double-check work when it’s likely to make a mistake
• Labelers’ bias : biased towards the cultural values of English-speaking people
-hallucinate
사전에는 환각을 느끼게 하다 이지만 케임브리지 사전은 할루시네이트에 대해 생성형 AI 등장으로 '사용자의 의도에 반하는 거짓 정보를 생성해 진실인 것처럼 현혹하는 행위'라는 새로운 의미를 갖게 됐다고 했다.
-강의에서 언급하지 않은 여러 범용AI
구글은 차세대 범용AI '제미나이(Gemini),네이버 '하이버클로바X', 카카오 '칼로', SK텔레콤 '에이닷', KT '믿음', LG '엑사원'
-Anthropic Claude
Anthropic AI
• Founded by Ex-researchers of OpenAI in 2021
• Invested $400 million (+10% stake) by Google in 2022
• A formal partnership with Google Cloud
Claude (early access only)
• Carry out similar tasks to ChatGPT
• Claimed less likely to produce harmful outputs, and more steerable
• Constitutional AI: Letting AI respond using a simple set of principles as a guide
(started with around 10)
-Google Bard
Google issued 'Code Red’ over ChatGPT (Nov 2022)
-Google PaLM
Efficient scaling based on Google’s Pathways system
Great scaling and breakthrough reasoning capabilities
-Meta OPT & LLaMA
Open Pretrained Transformer (OPT-175B)
Large Language Model Meta AI
-Self-Instruct Tuning on LLaMA
Stanford Alpaca
LMsysVicuna
And many other models are coming…
• OpenAI’sTulu
'코딩 도구 > LG Aimers' 카테고리의 다른 글
LG Aimers 4기 그리고 Linear Regression (2) | 2024.01.10 |
---|---|
LG Aimers 4기 그리고 지도학습(Supervised Learning) (2) | 2024.01.09 |
LG Aimers 4기 그리고 Bias and Variance (6) | 2024.01.06 |
LG Aimers 4기 그리고 머신러닝의 목표는 ? (0) | 2024.01.05 |
LG Aimers 4기 그리고 머신러닝을 위한 수학, Optional Course (2) | 2024.01.05 |