Deep Learning Image & Voice

Overview

  • for applied AI four pillars
    • corresponding to unlocking unstructured data
      • AI “core” capabilities in language and vision
      • natural language processing (NLP)
  • the “core” deep learning breakthroughs
    • vision - image recognition
    • language - voice recognition

Fastest growing product in history

  • result
  • conclusion
    • smart speaker increases the most quickly to 50%
    • but smart speaker goes smoothly after 5 years, compared the continuous dramatical growth of smartphone and TV after reaching 50% (growth flatten)
      • may due to the loss of freshness
      • may due to the loss of practicality

Cases

  • vision
    • in 2015, the accuracy of human and AI are equal, which is 95% (human accuracy level)
    • self-driving is another example
  • language
    • in 2017, the accuracy of human and AI are equal, which is 95%
    • Amazon Alexa (Echo Dot) is an example
      • core components of AI/ML applications
        • data input - voice
        • data (pre)processing - language data for soundbites + NLP
        • predictive models - far field voice recognition + understanding of intent and context of voice question or command
        • decision rules (rule sets) - use the appropriate Alexa skill to address user intent
        • response/output - Alexa voice response
    • voice-enabled smart kitchen is another example

Deep Learning & Neural Networks

Artificial neuron & neuron network

  • from scorecards to decision trees
  • an artificial neuron
  • a neuron network model

Deep learning

  • feature
    • the more the layers, the “deeper” the network
  • convoluted neural network (CNN)
    • not connecting all inputs to all of the neurons in the first layer
  • recurrent neural network (RNN)
    • create feedback loops where the output of later layers act as inputs to earlier layers
    • time latency effect
  • basic machine learning paradigms
    • supervised learning
      • using labelled data
    • unsupervised learning
      • using unlabelled data
    • semi-supervised learning
      • using a small amount of labelled data
      • and unlabelled data
    • reinforcement learning
      • concerned with how intelligent agents ought to take actions in an environment in order to maximize the notion of cumulative reward

White Box & Black Box AI

Create a common sense baseline first

White box vs. black box AI

  • white box
    • simple models - scorecards, decision trees
    • easy to understand how a score and prediction about someone is arrived
    • which data items are important, which less
    • easy to code
    • still produce pretty good predictions
  • black box
    • complex models - neural networks, object recognition, language translation, game playing
    • requires multiple machine learning approaches - autonomous robots, cars, digital personal assistants

The ethics of applied AI & deep learning

  • can do with AI -> should do with AI

Additional Reading