LogoNotes by Cole Gawin
v0
Back to LING 385

Lecture 6

Why is the Correlation Matrix Memory we developed not so great?
  1. it learns too fast
    • when you learn a foreign language, it takes a long time
    • we learn gradually/iteratively
    • in designing an AI, we need to get it to learn a bit at a time from what it observes
  2. the neurons are all synchronous
    • neural firing in brains is asynchronous
    • at some random time, some neuron updates, and at some other moment, a different one updates, etc
  3. lack of variety in training data
    • requires enormous amounts of varied data

Neural Activation
  • neurons are either on or off
  • however, when neurons fire, they can turn on and off a lot in some amount of time, or they can turn on and off a little bit
  • activation = frequency of on/off switching in a neuron
    • with each neuron we will associate a number a for its activation

Neural Activation Dynamics
  • How does a neuron's activation change with time?
    • that is, what is a/dt? (rate of change of a with respect to time)
  • if a/dt is a big positive number it says that a is growing by a lot
  • if a/dt is a small positive number it says that a is growing by a little
  • if a/dt is a big negtive number it says that a is decreasing by a lot
  • if a/dt is a small negative number it says that a is decreasing by a little

Improving Learning by making it gradual
  • humans learn in a process involving knowledge acquisition and adaptation
    • noticing something new and adjusting our existing knowledge accordingly
  • Hebbian learning focuses on adjusting the weight of a connection based on the activations of connected neurons
  • if too-fast learning (as in the correlational memory matrix), the weight is set immediately based on observed neuron activations
  • alternatively, Hebbian learning is a gradual process using current weight (w_t) and a small part of what needs to be learned
    • update in weight w_t+1 is based on the current weight w_t and a fraction of the learning requirement
  • gradient descent is the incremental adjustment in the learning process.

Improving Recognition
  • instead of updating all the neurons at once, we will do the following:
    • if, for instance we have 5 neurons
      1. Pick a random number between 1 and 5, let’s say we get 4
      2. Update the 4th neuron’s value
  • this leads to an "asynchronous algorithm"
    • closer match to how brains work