- Just wanted to update him on some of the differences in SBF models, specifically the EIO-SBF model
- And how that effects my current methods
- We should not change our current methods
- We are applying it more in the automata scope
- Thus no predictive stimuli is fine
- Also decided to keep the scope to discrete behavior
- However if we are going to have unexpected changes to RP, we need the ability to [[Temporal Rescaling|rescale]] oscillators
- I think this will be the cherry on top of my first paper
-
However, because we are restraining the scope to discrete activity,
- with oscillators only voting at discrete timesteps
-
rescaling the oscillator periodicity does not map one-to-one with the biological model
-
What rules will decide how oscillator periods may Pre/Re-cess in their periodicity
- As I can see it there are two options:
- Move forward: Period stays the same, but is “bumped” ahead by some time frames
- Increase / Decrease Period:
- I forgot what these are actually called
-
Instead we will attempt to borrow some methods from the automata field
- See automata section below
-
This is essentially a second layer of learning in addition to the weights, so we need to be careful to avoid chaotic behavior
-
The learning rate of the rescaling will need to be much slower than the weight learning rate
- We can trace number of timesteps between individual oscillator periods with respect to:
- When reward was issued / not issued
- When oscillators of other periodicities received reward / punishment
- I like this as it creates a denser network of relations between neurons
- Lateral Inhibition?
- Potentially as some form of regenerating decay?
- This would basically end up meaning a degenerative threshold value for the vote
- Anis provided a paper which uses the example of web crawling with some periodicity where changes to web pages happens with unknown frequency
- [[@kolobovStayingDateOnline2019|Kolobov et al. (2019)]] [See online] (https://papers.nips.cc/paper_files/paper/2019/file/ad13a2a07ca4b7642959dc0c4c740ab6-Paper.pdf)
- Also provided an older paper: [[@wolfOptimalCrawlingStrategies2002|Wolf et al. (2002)]] See Online
- Look for “Adaptive RL” and “Restless MAB”
- See also the old papers you found in automata when you were looking for relevant literature - [[SBF - Automata Experiment#Exploring Literature]]