AI Brief

•

Artificial Intelligence(AI) 的演進, 應用, 影響, 未來.

AI的演進

三分鐘看懂人工智慧

Evolution of Deep Learning models

Types of Machine Learning

Neural model

Biological Neuron model

Dendrites (樹突), Soma(細胞體), , Axon (軸突), Synapse (突觸)

Mathematical Neuron model (Perceptron)

Y = X * w + b


Multi-Layer Perceptron

AI Key Researchers

Turing Award 2018 : Yann LeCun, Geoffrey Hinton , Yoshua Bengio

Yann LeCun : father of CNN, Facebook Chief AI Scientist

Geoffrey Hinton : father of Deep Learning, Google Brain

Yoshua Bengio : RNN, GAN, pre-training, theano

Andrew Ng (吳恩達) : Stanford AI Lab, founder of Coursera, deeplearning.ai

李宏毅 : Machine Learning Lectures, 台大電機系教授

Machine Learning (Hung-yi Lee, NTU)
李宏毅老師機器學習課程筆記

Sam Altman, CEO of OpenAI

AI 模型的應用

CNN ~ Image Classification (影像分類)

CNN ~ Object Detection (物件偵測)

CNN ~ Pose Estimation (姿態估計)

CNN ~ Face Recognition (人臉識別)

GAN ~ Image Inpainting (修圖)

RNN ~ Text-to-Speech (語音轉文字)

RNN ~ Text-to-Text : Translation (翻譯), Text Generation(文本產生)

GPT ~ Q&A, Exam, Chatbot (問答考試，聊天機器人)

Generative AI (生成式人工智慧)

ChatGPT

Summary/Code/Image

Text-to-Video

OpenAI Sora

[Runway](https://runwayml.com/）

Text-to-Image

Midjourney

Stable Diffusion

Text-to-Voice/Speech

ElevenLabs

Meta SeamlessM4T

Facebook SeamlessM4T covers 101 languages for speech input, 96 Languages for text input/output, 35 languages for speech output.

Meta Audiobox

Text-to-Song

Suno

Humanoid Robot

只用 13 天，OpenAI 做出了能聽、能說、能自主決策的機器人大模型

News

AI的影響

50 Most Visited AI Tools and Their 24B+ Traffic Behavior

  • The top 50 AI tools attracted over 24 billion visits between September 2022 and August 2023.
  • ChatGPT led with 14 billion visits, making up over 60% of the analyzed traffic.

ChatGPT user demographic

More than one quarter of ChatGPT’s audience falls into the 18-24 age bracket in Similarweb’s website demographics model (which does not include children under 18). The chart below is for the US, but the worldwide numbers are similar.

AI的未來

AGI - Artificial General Intelligence (通用人工智慧)

Paper: Levels of AGI: Operationalizing Progress on the Path to AGI

GAIA - General AI Assistants

Paper: GAIA: a benchmark for General AI Assistants
Created by researchers from Meta, Hugging Face, AutoGPT and GenAI, the benchmark “proposes real-world questions that require a set of fundamental abilities such as reasoning, multi-modality handling, web browsing, and generally tool-use proficiency,”

GNoME: Harnessing graph networks for materials exploration

GNoME uses two pipelines to discover low-energy (stable) materials. The structural pipeline creates candidates with structures similar to known crystals, while the compositional pipeline follows a more randomized approach based on chemical formulas. The outputs of both pipelines are evaluated using established Density Functional Theory calculations and those results are added to the GNoME database, informing the next round of active learning.

A-Lab

A-Lab, a facility at Berkeley Lab where artificial intelligence guides robots in making new materials. Photo credit: Marilyn Sargent/Berkeley Lab

DeepMind Cultural Transmission

Paper Learning few-shot imitation as cultural transmission
Blog: DeepMind智慧體訓練引入GoalCycle3D
以模仿開始，然後深度強化學習繼續最佳化甚至找到超越前者的實驗，顯示AI智慧體能觀察別的智慧體學習並模仿。
這從零樣本開始，即時取得利用資訊的能力，非常接近人類積累和提煉知識的方式。

LiquidAI

Paper: Liquid Time-constant Networks
Paper: Closed-form continuous-time neural networks
News: Liquid AI, a new MIT spinoff, wants to build an entirely new type of AI

Blog: 受線蟲啓發開發液態網路
液體 神經網路的想法，源於多年前奧地利維也納工業大學（Vienna University of Technology） Radu Grosu 教授的實驗室。
蟲僅 １ 毫米長，神經系統只有 302 個 神經元（人類有大約 860 億個 神經元），位於食物鏈底層，卻能進行一系列高階行為：移動、覓食、睡覺、交配，甚至從經驗中學習。
他們意識到，研究線蟲的大腦實際上如何工作，也許有助於製造能適應意外情況的彈性 神經網路。

  • 所謂液體神經網路，首先是指其架構像液體一樣，是動態的，具有高度的靈活性和適應性。
  • 和動輒數十億引數規模的生成 AI 模型相比，液體神經網路的另一個特點是規模小得多 。
    比如，GPT-3 包含約 1750 億個 引數和約 50,000 個 神經元。 而針對諸如在室外環境中駕駛 無人機等任務進行訓練的液體 神經網路可以包含少至 20,000 個 引數和不到 20 個 神經元。
  • 由於尺寸小，架構也簡單，液體神經網路也有可解釋性方面的優勢。
    液體神經網路也有侷限性。和其他 神經網路不同，液態 神經網路青睞「時間序列」資料。
    Rus 和 Liquid AI 團隊成員聲稱，該架構適合分析隨時間波動的任何現象，包括視訊處理、自動駕駛、大腦和心臟監測、金融交易（股票報價）和天氣預報等。

A Tutorial on Liquid Neural Networks including Liquid CfCs

Code: Closed-form Continuous-time Models
Two Neuron Models: Lliquid Time-Constant (LTC) model and Closed-form continuous-time(CfC)
Wiring:
Diagram:

AI competitions

Kaggle Competitions

AI Cup 教育部全國大專校院人工智慧競賽



This site was last updated June 27, 2024.