evjrob
evjrob on Hacker News
The model is implemented in PyMC3 and runs daily as a batch job. The model state and predictions are stored in DynamoDB.
Early each morning this batch job loads the previous day's posterior as priors, fetches the results for the games played that day, and then fits the Bayesian network to the game results to get new posteriors. I make predictions with these posteriors and then write all the results back to DynamoDB.
The frontend is essentially D3, Django, and Zappa.
This rings true from my experience as a data scientist for a non-tech compay. I spend more time doing expectation management than ML work. This seems to hold true whether we're trying to use ML on a new and novel problem statement, or trying to reproduce some published application or supposed solution a competitor claims to have.
I think it's clear that humans at least try to pick out the causal factors, and reason causally about the outcomes of specific actions. Non human animals do too; they can learn to take specific actions for food rewards in a Skinner Box for example. Now I agree humans get it wrong a lot of the time, and other animals might get it wrong even more than we do.
I don't think our fallibility in causal reasoning makes it useless to pursue as a goal in artificially intelligent systems. It doesn't need to be perfect, just useful and better than not having it. Afterall, our perception systems are pretty fallible too, otherwise things like optical illusions wouldn't exist.
True, but tight oil production that was previously though to be impossible to extract economically has more than offset losses in conventional production. Peak oil predictions of decades past presumed this could only happen when oil prices rose so high that alternative energy sources would be competitive. Therefore these reservoirs would remain uneconomic. Instead fracking got really cheap.
Extracting oil from shale reservoirs was though to be too difficult and expensive to ever really flood the market like this. Turns out it wasn't.
The point lalaland1125 is making is that the equations of motion are the existing superior model in this scenario.
I wouldn't be surprised if these speech to text models are susceptible to black box attacks in the same way that image classifiers are:
https://arxiv.org/abs/1605.07277