jkwiki

 AI

* AIN (Artificial Narrow Intelligence)

* AGI (Artificial General Intelligence) - “1 trick ponies!”

* AI transformation playbooks

* AI and Society / AI bias

* Types of Machine Learning

• Supervised Learning → Input → Output (A → B)
  • Input (e-mail) → Output (Spam or not Spam?) .. In this case the ML “application” is “spam filtering”
  • Audio → Text Transcript (Speech Recognition)
  • English → Chineese (machine translation)
  • Ad, User Info → Click (yes/no) (online advertising)
  • image, radar info → position of other cars (self driving cars)
  • Input → Output (A to B mapping.. supervised learning)
  • DATA, and lots of it, makes Supervised Learning work great.
  • image(like cat or bike)/label(cat, or not cat)
    • Methods for Acquiring Data
      • Get by manual labeling, get a bunch of pics and manually label them as “cat” or “not cat”, gives you a starting data set for building a “cat detector”
      • Get from observing behaviors (like users, physical machines)
      • Download from websites
    • I.T. Team → feed data to → A.I. team (interplay between IT and AI teams)
    • Just because you have a lot of data, does not always mean an AI team can make your data valuable.
    • Data is messy , garbage in/ garbage out, incorrect data, or data with missing values, multiple types of data (images vs text)..unstructured data. structured data (spreadsheet for example, or csv file)
  • TERMINOLOGY
    • Output of a Data Science project is a set of INSIGHTS that can help you make better decisions.
    • Machine Learning and Data Science boundaries are fuzzy somewhat
    • ML project will often result in a piece of software
    • But output of a Data Science project is Insights, summaries of conclusions etc.
    • Deep Learning / Artificial Neural Network (both terms mean same thing) .. still A to > B
      • Deep Learning is a type of Machine Learning
    • ANN's not like the biological brain really, at all
  • WHAT MAKES AN AI COMPANY? WHAT DOES IT TAKE FOR A COMPANY TO BE GOOD AT AI?
    • A/B Testing, short iterations, decisions pushed down to product owners, engineers
    • Are you doing the things that AI lets you do really well? If so, you may be an AI company.
      • Strategic Data Acquisition
      • Thinking through how to get data
      • Unified data warehouse
      • Pervasive Automation
      • New roles like MLE (Machine Learning Engineer)
      • AI Transformation Steps
        • pilot projects to kick tires
        • build in-house AI team
        • provide broad AI training
        • develop AI strategy
        • develop internal/external communications
  • What can AI do, and what can it NOT DO?
    • almost anything you can do with 1 second of thought.
    • Deep Learning
      • Price → (Neuron) → Demand (a prediction of demand)
      • Different neurons in the network can figure out different things, you might
      • have 1 type of neuron that can figure out “affordability” of a product, and another neuron that can determine a consumer's “awareness” (based mostly on marketing).
      • (price + shipping cost) → feeds into → neuron → output is “affordability”
      • (marketing level) → feeds into a different neuron type → output is “consumer awareness”
      • and different “neuron” types can take different or even the same types of inputs.
      • so, you can mix and match inputs into different neuron types
      • the neural network's JOB IS TO MATCH INPUT A TO OUTPUT B
      • but the attributes that get computed are actually learned/created by the neural network itself! so things like “affordability” or “consumer awareness” are things that you don't have to program, the network figures it out on it's own via training on large data sets.
      • So 1 possible use case for a neural network is predicting demand of a product based on factors like price, shipping cost, consumer awareness etc. but there are other applications of deep learning/neural nets as well, like facial recognition.