Linear Pattern Associators
Readings
Recommended background reading (required if you haven't taken linear
algebra): Chapter 5.
Required reading: Chapters 6, 7.
Human memory systems
- Short-term versus long-term memory: states versus weights
- Episodic or "flashbulb" memory, versus gradual, incremental, assimilation
into an existing representation
- Characteristics of human episodic memory:
- huge capacity
- one-shot learning
- impaired in amnesics
- associated with the hippocampus & related brain structures
- probably only accounts for a small fraction of what is learned
- may be modelled well by associative memory models
Associative memory models
- Hebb's learning postulate
- neurobiological evidence: LTP, LTD
- the generalized Hebb-rule or outer-product rule
- a feed-forward linear pattern associator
- conditions for perfect recall
- hetero-associative (supervised) versus auto-associative (unsupervised) models
- recurrent versus feedforward auto-associative models
- performance measures:
- memorization (recall & recognition rates)
- generalization (pattern completion)
- Summed vector memories, matched filters
- Summary of major features:
- one-shot learning
- poor capacity