Valentina’s TNJC on A model of schema learning based on biological dimensionality reduction during sleep.
Main Ideas
- $\zz$: A fixed, sparse transformation of $\xx$ through $\WW_\text{ZX}$.
- Two pathways to compute value:
- Direct, rote pathway: $\WW_\text{rote} \zz$
- Low-dimensional, schema pathway:
- $\yy = \WW_\text{schema} \zz.$
- This is then centered and normalized to $\tilde \yy$
- Value computed as the sum: $$ \vv = \WW_\text{rote} \zz + \WW_\text{knowledge} \tilde \yy.$$
- The benefit of this system is that
- It avoids slow learning of new tasks
- Can perform transfer learning by aligning schemas
- Low-dimensionality of scheme representation also seems to speed up learning.