1
|
- Thomas R. Shultz
- McGill University
- and
- Francois Rivest
- Université de Montréal
|
2
|
- Tabula rasa networks
- Humans use existing knowledge
- Accounts for the ease & speed of much human learning, & for
interference effects
|
3
|
- Recruits networks as well as single units
- Both are differentiable functions
- Trains weights to inputs of existing networks to determine whether their
outputs correlate with error
- Trains output weights from recruited network to incorporate it into
solution
|
4
|
|
5
|
|
6
|
|
7
|
|
8
|
|
9
|
|
10
|
|
11
|
- Discriminability-based transfer (Pratt, 1993)
- Multi-task learning (Caruana, 1997)
- Task-rehearsal method (Silver & Mercer, 1996)
- Lifelong learning (Thrun & Mitchell, 1993)
- Input re-coding (Clark & Thornton, 1997)
- But KBCC stores & searches for knowledge within a generative network
|
12
|
- Can KBCC find & combine source knowledge of components to learn a
more complex target comprised of these components?
- Does use of these knowledge components speed learning?
|
13
|
|
14
|
|
15
|
|
16
|
|
17
|
|
18
|
|
19
|
|
20
|
- KBCC finds, adapts, and uses existing knowledge in new learning,
significantly shortening the learning time
- When exact knowledge is present, it is recruited for a quick solution
- The more relevant the source knowledge is, the more likely it is
recruited & the faster the new learning
|
21
|
- Unlike previous techniques for which both inputs & outputs of the
source and target task must match precisely, KBCC can recruit any sort
of function
- Source network inputs & outputs can be
- Arranged in different orders
- Employ different coding methods
- Exist in different numbers than in the target network
|
22
|
|