@lorb said in #14:
> Also generally: humans don't overfit the same way machines do. So far all available actual data we have on the Woodpecker Method actually supports it, and besides speculation and analogies we don't have anything to proof the downsides mentioned.
All available data. not doubting that it might be true. but what is the extent of all available data.
A few book authors with apetizing titles, and how many books get sold, would not be a convincing argument.
The rationales about spaced or scheduled repetitions are themselves speculation and analogies, until proven.
So while in hypothesis land, stating alternative hypotehses, are also valid arguement to share.
The blog proposes a sound alternative, and possibly the opportunity to test it, not one author at a time.. The statistics would luckily be a nice side-effect of the group hypothesis as alternative motivator possible resulting in same diligence on hard work upon chess positions challenges.
I think i would agree that we need data. So where is that available data?
Last post about pre-existing learning, does point to more un controlled variables, not part of the woodpecker picth.
It also makes the dart throwing analogy possible more relevant. because the physical sports muscle memory that i think has been used in supporting analogy or perhaps even the method ideation, is not making explicit that the internal proprioception model that allow humans to throw things in some general direction with some effectiveness is usually learned in early childhood (i was said to have started early myself, ejecting food that had accumulated in my mouth, baby food not to my liking, i must have learned some basic baliistic then i bet, and locomotion trial and errors, rarely about repeating precise motions, more about trying in all direction of muscle tensions.. learning not to contract antagonist with agonist at the same time...
I agree that general arguments from other fields do not suffice. but that also applies to woodpecker.
music. we have innate musical perception or very early development of it, before even becoming musicians..
this is not about recognition more about execution precision and speed it seems. reproducible, and allowing to be part of a group so that hard thinking processes be about the whole coordination between musicians, as it can't be second natures what the other would do. or other conscious processes needed during performance.
bascially NN in machine are emulating basic visual cortex architecture, so I would say, that they might be model certain aspect of our ability to store and recognize pattern. In chess there is board feature pattern recognition, and mapping that to dynamic patterns.. The notions of generalization, is not an artefact of CPU implementation of the neural networks. it is a mathematical property of the basic cortex layer structure. It is about how a flexible function famiily can fit to data in parsimonious fashion and still extrapolate correctly to unseen new test input.
The functions do not have any prior knowledge about the reality to learn (exageration, we do have some connectivity that may be constraining or biased like all animals in their ecosystems, to some life relevant stimuli, gravity might have had some effect, maybe)...
As humans we can learn new things and adapt to new visual environment (think of the inverting goggles, if not a myth, that over the time scale of weeks, can be learned away and the brain would still make sense of where up and down is, in all it locomotion needs).
So, one the board feature recognition aspect alone, doing too much of one thing will keep adjusting the very flexible function familiy to the very details of the same position, even those that may not be relevant to the execution pattern being optimized for speed without any conscious interference.
So yes overfitting would apply in wet networks too.. we don't have mind control over our visual subconscious processes that actually implement such mathematical functions. So the only argument against over-fitting is that there are 1000 puzzle position challenges.. but we also need to know how different they span.