Research Engineering at EdgeImpulse
@mat_kelcey
- 12 Followers
- 84 Following
- 29 Posts
Solal Nathan📚 What are you top general Machine Learning books ?
I would say:
📘 Deep Learning - Goodfellow, Bengio and Courville
📗 Artificial Intelligence: A Modern Approach - Russel & Norvig
📙 Pattern Recognition and Machine Learning - Bishop
#AI #ML #MachineLearning #book #books #recommendations #academia #question #discussion #media
oh boy, that point where you've got everything into a single sweet sweet jitted update_fn. while True pump it!!
Olivier Grisel@mat_kelcey this is very interesting, thanks for the heads up. We might even have a story to (optionally) use jax at training time for some specific estimators:
- either via the Array API spec that JAX might want to target at some point: https://scikit-learn.org/dev/modules/array_api.html#array-api
- or for Cython powered estimators, via a new plugin system: https://github.com/scikit-learn/scikit-learn/pull/24826
porting some code to sample audio triplets, cross fade & generate spectrograms from a mix of numpy/tensorflow to jax.
being able to move the batching all the way out with vmap makes sooo much less code.. it'll be even cleaner when it's bolted onto the training loop
being able to move the batching all the way out with vmap makes sooo much less code.. it'll be even cleaner when it's bolted onto the training loop
"solving cartpole... by evolving the raw bytes of a 1.4KB tflite microcontroller serialised model"
microcontroller models are so small you can just run a genetic algorithm directly against the bytes of the serialised model! :D
who needs gradient descent anways?
the clever elephant asks mr dizzy "what's small and furry and likes cheese?" time stretched from 2sec to 23 min.
whats small and furry and likes cheese?
first _in person_ ML/AI meetup coming up soon! it's been awhile... have a pile of #jax laptop stickers to give out....
the main use case for dropout is as a form of regulariser. but what if we used it instead to a make a model that's robust to having different combos / forms of the input? we could then use a genetic algorithm to trade off input complexity to performance, all with a single model...
