Complementary Learning Systems Flashcards
What is the goal of learning?
Developing an internal representation of allowing one to predict what will happen next, making inferences about unseen elements of environment
Simple Neocortical Model
After training, if a network is given information about some properties of an object, the network can make a guess about other properties of that thing
Sensory input (sees beak, hears tweeting, smells bird feather) → internal representation figures it is a bird
Hebbian Learning
“Neurons that fire together, wire together” - works effectively using neuroscience, this helps learn that “birds can fly” but has difficulty picking up on content such as “penguins cannot fly”
Error-driven
Learning by reconstruction
Allows you to contrast real world and internal presentation and tweak it accordingly. Memorizing shirt color example, looking down and realizing you missed some things, then you add some more inputs to your internal representation, more selective in the weight changes that it makes, learns from the errors made in your models
Gradual vs One-time Learning
Neocortex uses gradual learning, but we also have the ability to learn rapidly (episodic memory)
If we want to learn after one trial, we need to make larger weight changes
Catastrophic Interference
When you need to learn a new piece of information (that a penguin is a bird that doesn’t fly) it can lead to to an old piece of information being forgotten
Hippocampus doesn’t have this problem
Hippocampus’s function for memory consolidation
Can assign distinct, minimally overlapping representations, rapidly learning individual episodes
Helps with slow extraction of statistics and correlations across episodic memories
Can sound taxing and necessary for huge capacity but most of the episodic memory isn’t retained
Begins to consolidate episodic memories into semantic memories
Pattern Completion
The hippocampus can rapidly memorize patterns of cortical activity in entorhinal cortex so they can be recalled later, given a partial reminder
The cortex can do pattern completion but only after lots of exposure to information
Pattern Seperation
Assigning of a distinct set of neurons to each memory, only done by hippocampus and not cortex
How is sparse activity reinforced?
Inhibitory interneurons and sparse connecitvity
Describe CA3 region
Through trisynpatic pathway, there is sparse activity and connectivity
Limitation of Hebbian Learning
Trouble making fine distinctions, does not have any provisions for fixing these errors once they happen
Theory of Equipotentiality
Memories are not stored in one area of the brain; rather, the brain operates as a whole to store memories.
LTP:
A process in which synaptic transmission becomes more effective as a result of recent activity;