douglasorr · HackerTrans

douglasorr

no profile record

Submissions

[untitled]

1 pointsby douglasorr5 miesięcy temu0 comments

[untitled]

1 pointsby douglasorr7 miesięcy temu0 comments

Five game jams: Lessons learned

douglasorr.github.io

1 pointsby douglasorr7 miesięcy temu0 comments

[untitled]

1 pointsby douglasorrw zeszłym roku0 comments

Procedural Textures with Hash Functions

douglasorr.github.io

32 pointsby douglasorrw zeszłym roku5 comments

Llama 3.2 Vision – A Deep Dive

graphcore-research.github.io

2 pointsby douglasorr2 lata temu0 comments

Benchmarking std:{string, string_view, reference_wrapper}

douglasorr.github.io

1 pointsby douglasorr2 lata temu0 comments

ML papers (June): Mamba-2 and Matmul-free Models

graphcore-research.github.io

3 pointsby douglasorr2 lata temu0 comments

[untitled]

1 pointsby douglasorr2 lata temu0 comments

[untitled]

1 pointsby douglasorr2 lata temu0 comments

Almost-scaled dot-product attention

graphcore-research.github.io

2 pointsby douglasorr3 lata temu0 comments

Metaphor Thinking in Deep Learning

douglasorr.github.io

2 pointsby douglasorr3 lata temu0 comments

Rasterized ray tracing of sound in JavaScript

douglasorr.github.io

2 pointsby douglasorr5 lat temu0 comments

GitHub Game Off – “Bug”

github.blog

3 pointsby douglasorr5 lat temu0 comments

Autodiff from scratch

douglasorr.github.io

59 pointsby douglasorr5 lat temu1 comments

How to train your classifier – deep learning training objectives

douglasorr.github.io

1 pointsby douglasorr5 lat temu0 comments

My phone case, the hidden state of an RNN

douglasorr.github.io

71 pointsby douglasorr5 lat temu36 comments

Write code for humans. Design data for machines

douglasorr.github.io

4 pointsby douglasorr6 lat temu0 comments

Write code for humans. Design data for machines

douglasorr.github.io

3 pointsby douglasorr6 lat temu0 comments

Finding unique items: hash vs. sort

douglasorr.github.io

137 pointsby douglasorr7 lat temu81 comments

comments

douglasorr

5 lat temudiscuss

I think at least some of the structure is down to the update equation, roughly 'new = sign(gate * old + ...)'. Since 'gate' is always positive, this means that even when the network is randomly initialized, the state is more likely to stay the same between two time steps than to change.

I tried "deep transitions" which repeats this equation a couple of times, and the results were indeed noisier.