gost

[gost @smoke88__]

@ariiikato don't choose, hot mama 555

[gost @smoke88__]

@ariiikato hot

[gost @smoke88__]

@ariiikato i like your bangs

[gost @smoke88__]

@ariiikato happy valentines day

[gost @smoke88__]

@ariiikato i lied

[gost @smoke88__]

@ariiikato make an only fans already

[gost @smoke88__]

@ariiikato 💚

[gost @smoke88__]

@ariiikato "how tite is ur love"

[gost @smoke88__]

@ariiikato are you me

[gost @smoke88__]

@ariiikato i could eat it for you

[gost @smoke88__]

@ariiikato this crop is so 😈😈😈

[gost @smoke88__]

@ariiikato Bible Black

https://t.co/yn3Ywnkvfo

[gost @smoke88__]

@ariiikato yo

[gost @smoke88__]

This implies a different sort of optimization process where you spend most of your time computing higher derivatives from where you stand instead of computing first order derivatives at many points.

[gost @smoke88__]

What if instead of doing deep learning optimization through a series of approximate steps you took a single exact step?

[gost @smoke88__]

You would simply compute the derivative, the derivative's derivative, the derivative's derivative's derivative, ... from the point at which you stand -- the curvature of the Earth beneath your feet.

[gost @smoke88__]

What if you take two batches A and B, and alternate between training on them 100 times? ABABABAB.. (100 times). A 1.33% absolute improvement over baseline in the first epoch of MNIST.

[gost @smoke88__]

What if you taught a neural network how to remember before teaching it arbitrary things?

The initial configuration of a neural network is not guaranteed to be optimal for remembering.

[gost @smoke88__]

By arriving at zero loss while alternating between batches A and B, we hypothesize that the parameter configuration of the network becomes initialized to a state that is better suited for remembering new data and internal states.