Scribd Logo
Upload

Welcome to Scribd!

  • 10 views

    COMP9491 Week1 Background

    Uploaded by

    ryj740447138rj
    The document discusses the current state-of-the-art in artificial intelligence including deep learning models for computer vision and natural language processing, machine learning models for tabular data analysis, and reinforcement learning for robotics and games. It also discusses research questions regarding developing AI methods that balance accuracy, efficiency, robustness, generalizability, explainability and trustworthiness as well as questions around commonsense reasoning and integrating knowledge and learning.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd

    COMP9491 Week1 Background

    Uploaded by

    ryj740447138rj
    10 views21 pages
    The document discusses the current state-of-the-art in artificial intelligence including deep learning models for computer vision and natural language processing, machine learning models for tabular data analysis, and reinforcement learning for robotics and games. It also discusses research questions regarding developing AI methods that balance accuracy, efficiency, robustness, generalizability, explainability and trustworthiness as well as questions around commonsense reasoning and integrating knowledge and learning.

    Original Title

    COMP9491-week1-background

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    The document discusses the current state-of-the-art in artificial intelligence including deep learning models for computer vision and natural language processing, machine learning models for tabular data analysis, and reinforcement learning for robotics and games. It also discusses research questions regarding developing AI methods that balance accuracy, efficiency, robustness, generalizability, explainability and trustworthiness as well as questions around commonsense reasoning and integrating knowledge and learning.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    10 views21 pages

    COMP9491 Week1 Background

    Uploaded by

    ryj740447138rj
    The document discusses the current state-of-the-art in artificial intelligence including deep learning models for computer vision and natural language processing, machine learning models for tabular data analysis, and reinforcement learning for robotics and games. It also discusses research questions regarding developing AI methods that balance accuracy, efficiency, robustness, generalizability, explainability and trustworthiness as well as questions around commonsense reasoning and integrating knowledge and learning.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    You are on page 1of 21

    Background

    COMP9491 Applied AI
    Term 2, 2023
    Outline

    ▪ Current state-of-the-art

    ▪ Research questions

    COMP9491 T2, 2023 1


    Current state-of-the-art

    ▪ Deep learning models for computer vision and natural


    language processing

    ▪ Machine learning models for tabular data analysis

    ▪ Reinforcement learning and logical reasoning for


    robotics and games

    ▪ Integration of multi-modal, or hybrid-AI, for enhancing


    method performance

    COMP9491 T2, 2023 2


    Current state-of-the-art (CV)

    ▪ ImageNet classification

    Source:
    AlexNet (NeurIPS’12)

    COMP9491 T2, 2023 3


    Current state-of-the-art (CV)

    ▪ ImageNet classification

    Source:
    https://paperswithcode.com
    /sota/image-classification-
    on-imagenet

    COMP9491 T2, 2023 4


    Current state-of-the-art (CV)

    ▪ Instance segmentation on COCO test-dev

    Source:
    Mask R-CNN
    (ICCV’17)

    COMP9491 T2, 2023 5


    Current state-of-the-art (CV)

    ▪ Instance segmentation on COCO test-dev

    Source:
    https://paperswithcode.com
    /sota/instance-
    segmentation-on-coco

    COMP9491 T2, 2023 6


    Current state-of-the-art (CV)

    ▪ DeepFake

    Source:
    https://www.youtube.com/w
    atch?v=l_6Tumd8EQI

    COMP9491 T2, 2023 7


    Current state-of-the-art (CV)

    ▪ Autonomous driving – Waymo dataset

    2D bounding boxes on camera images;


    and 3D bounding boxes in Lidar data

    Source:
    https://waymo.com/open/data/perception/

    COMP9491 T2, 2023 8


    Current state-of-the-art (NLP)

    ▪ Transformers – Attention Is All You Need (NeurIPS’17)

    Source:
    https://jalammar.github.io/illustrated
    -transformer/

    COMP9491 T2, 2023 9


    Current state-of-the-art (NLP)

    ▪ BERT marks a new era of NLP (Google’s search engine)

    COMP9491 T2, 2023 10


    Current state-of-the-art (NLP)

    ▪ GPT-2 / GPT-3: huge model with very impressive


    performance even without fine-tuning

    Source:https://towardsdatascien
    ce.com/choosing-the-right-
    language-model-for-your-
    nlp-use-case-
    1288ef3c4929

    COMP9491 T2, 2023 11


    Current state-of-the-art (NLP)

    ▪ Generative AI

    Source: https://towardsdatascience.com/choosing-the-right-language-model-for-your-nlp-use-case-1288ef3c4929

    COMP9491 T2, 2023 12


    Current state-of-the-art (NLP)

    ▪ IBM’s CodeNet dataset


    ▪ Project CodeNet is a large dataset aimed at teaching AI to
    code that consists of some 14M code samples, about 500M
    lines of code, in 55+ different programming languages.

    Source:
    https://github.com/IBM/Project_CodeNet

    COMP9491 T2, 2023 13


    Current state-of-the-art (ML)

    ▪ Decision trees, such as Gradient Boosted Trees,


    typically provide the best results for tabular data

    Source:
    https://towardsdatascience.com/the-
    unreasonable-ineffectiveness-of-deep-
    learning-on-tabular-data-fd784ea29c33

    COMP9491 T2, 2023 14


    Current state-of-the-art (RL)

    ▪ Reinforcement Learning in robotics and games:


    ▪ Humanoid walk – DeepMind
    ▪ Puppy walk and play fetch – Unity
    ▪ Alpha Go – DeepMind
    ▪ DOTA - OpenAI

    Source:
    https://medium.com/machinevision/
    overview-of-reinforcement-
    learning-58fbb905dbe0

    COMP9491 T2, 2023 15


    Current state-of-the-art (Logic)

    ▪ Logical reasoning in robotics:


    ▪ Robot navigation
    ▪ Task planning

    Learning and Reasoning for Robot Dialog and Navigation Tasks (SIGdial’20)

    COMP9491 T2, 2023 16


    Current state-of-the-art (Logic)

    ▪ Logical reasoning in games:


    ▪ Sudoku puzzles
    ▪ Crossword puzzles

    The championship puzzle


    (Dr. Fill completed it in 49 seconds).

    Source: https://slate.com/technology/2021/04/american-crossword-puzzle-
    tournament-dr-fill-artificial-intelligence.html

    COMP9491 T2, 2023 17


    Current state-of-the-art (Hybrid)

    ▪ Hybrid AI approaches for various applications:


    ▪ Object recognition and semantic segmentation in images
    ▪ Solving algebraic equations and complex algorithmic problems
    ▪ Search engine, graph data analysis

    Source: L. Adilova et al., Plants Don’t Walk on the Street: Common-Sense Reasoning for Reliable Semantic Segmentation, CVPRW 2021.

    COMP9491 T2, 2023 18


    Research questions

    ▪ How to develop AI methods that can meet the following


    criteria, often with a certain degree of balance:
    ▪ Accuracy
    ▪ Efficiency
    ▪ Robustness
    ▪ Generalisability
    ▪ Explainability
    ▪ Trustworthiness

    COMP9491 T2, 2023 19


    Research questions

    ▪ Commonsense reasoning
    ▪ Implicit knowledge in commonsense
    ▪ Important for many real-world applications

    ▪ Integration of knowledge and learning


    ▪ Knowledge: rich expressiveness but difficulty in representation
    ▪ Learning: relative simplicity in model design but huge requirement
    of training data

    ▪ Interesting article:
    ▪ G. Marcus, The Decade in AI (https://arxiv.org/abs/2002.06177)

    COMP9491 T2, 2023 20

    You might also like