Internet of things

Thing + Computer intelligence + connection to internet

Architecture

  • Embedded system (Things)
  • Node devices (THings) collect data through sensors
  • Data is sent to the internet through the gateway
  • Data analytics takes place in cloud or data center

These things may not have connection to the internet, thus it needs to connect to a computer system (Gateway) whcih are considered as edge (of internet)

Understanding IOT Data

Important:

  • Power management
  • security
  • Data conversion

Data:

  • Too big
  • Too fast: Data streaming and often processed in real-time
  • Too diverse: Different types of data types (Image data, heart rate data)

Ways to understand data:

  • Descriptive: “What happen”, summarises/presents the raw IoT data
  • Diagnostics: Look and understand the process that is causing the data (ML)
  • Discovery: Find patterns that we did not notice earlier, explain those patterns (Event)

Marketing found that epoeple in the eastern part of singapore prefer a particular brand of shampoo compared to those of the west. (Pattern)

  • Predictive: Use data and answer the question of what is going happen next
  • Prescriptive: Find a solution to the predictive analysis

Unlike the 3 prev tech, these are foresight techniques as we try to guess what to do in the future

CS3237-1-1.PNG

CS3237-1-2.PNG

Machine learning

  • Present the machine with alot of data
  • Machine learns on its own to recognise patterns in the data

Neural Networks

  • Dense networks
  • “Neurons” are computational unit called nodes
  • Neurons fires a 1 if the sum exceeds a treshold through an activation function (Step function)

CS3237-1-3.PNG

Learning laws: To give labels to figure out the meaning of the weights

Deep learning

Problems with neural

  • Adjust the parameters according to a learning law based on data given
  • Need large number of parameters to learn highly complex tasks
  • Curse of dimensionality
    • The number of parameters increases, the number of training data required grows exponentially
    • Neural network “Overfits” - it memroises the data it is presented instead of learning from it. Good results for data that is use for training but bad for any other data

What about deep?

  • Simplify the data
  • Extract features
  • Make data invariant
    • Shift invariance
    • Scale invariance

      To remove unneccessary data

Note that deep learning is part of neural

Convolutional neural networks

  • Kernal layer: Small window scan over image
  • Pooling layer: Summarise previous convolution layer to ease learning
  • Dropout layer: randomly drop info frm prev layer to prevent overfitting of training data
  • Dense layer: Classify based on simpify data

Kernal

Take many filters and scan over the image to produce activation maps

Apply fucntion to introduce non-linearity

In a neural networks, linear results are very bad (ax+b), thus we insert function to make it not linear

Pooling

  • Pick the max number within the window CS3237-1-4.PNG

  • Reduce the data produce
  • Remove the unnecessary data to remove noise

The distance the pooling window moves is called a stride

Dropout

  • Prevent CNN from memorising the data

CS3237-1-5.PNG

  • Choose some to deactivate = Remove weights = require less data to train
  • Nodes might be reinserted again

Dense layer

  • Look at the summarised maps

  • Make a classifcation CS3237-1-6.PNG

  • Each layer we will get a more abstract view of what is in the image. This help us simplify the image such that we can learn the data without many parameters

CS3237-1-7.PNG

We can see a more abstract version of the image, the abstract view is the thing that enables it to recognise the image without too many data

How deep learning is used in IoT

  1. Sensor gather data
  2. Deep learning and analytics can be put on this data
  3. Try to predict the future base on the data

Slides