Muhammad Firmansyah Kasim

Researcher at @UniofOxford | Lecturer in math at @TrinityOxford

[Muhammad Firmansyah Kasim @mfkasim]

Fully-differentiable Density Functional Theory written in @PyTorch. Pre-print paper here: https://t.co/ZvjZLn2Voo. I wonder what new applications this could bring.

[Muhammad Firmansyah Kasim @mfkasim]

@ibamarief How can you solve epidemic of loneliness by making an AI companion? Wouldn’t that make people more isolated to other people?

[Muhammad Firmansyah Kasim @mfkasim]

Really good work parallelising RNN!

[Scott Linderman @scott_linderman]

Did you know that you can parallelize *nonlinear* RNNs over their sequence length!? Our @NeurIPSConf paper "Towards Scalable and Stable Parallelization of nonlinear RNNs," which introduces quasi-DEER and ELK to parallelize ever larger and richer dynamical systems! 🧵 [1/11]

[Muhammad Firmansyah Kasim @mfkasim]

@crawlingladybug @AsahPolaPikir Pantasan bersin2 ka …

[Muhammad Firmansyah Kasim @mfkasim]

@_fikri_auliya Kalau diplot index vs time, bakalan jadi sawtooth signal dengan periode 2*(n-1), jadi T bisa direduksi jadi T=T%[2*(n-1)]. Habis itu tinggal dibagi aja kasusnya apakah T<=n-1 atau tidak.

[Muhammad Firmansyah Kasim @mfkasim]

@uqyauthor Pantesan notifikasi twitter tiba2 jadi banyak. Hahaha.. Thanks, qy!

[Muhammad Firmansyah Kasim @mfkasim]

@imrenagi … dan bisa bertahan hidup dengan “gaji” Indonesia di Swiss.

[Muhammad Firmansyah Kasim @mfkasim]

@_fikri_auliya Kalau tensor programming di deep learning, pakai for-loop (apapun variablenya) hukumnya makruh

[Muhammad Firmansyah Kasim @mfkasim]

@francoisfleuret You might want to check our paper on how to compute nonlinear RNN in parallel: https://t.co/B5S7J5bs2U

[Muhammad Firmansyah Kasim @mfkasim]

@_fikri_auliya Di rumah sering, bahkan kita mau bikin kata baru biar singkat: “GPT it”

[Muhammad Firmansyah Kasim @mfkasim]

@sonnylazuardi Kalau lagi di Oxford, ngopi2 yuk

[Muhammad Firmansyah Kasim @mfkasim]

@ZikriIsmal @aria_ghora @__nggih @nmonarizqa @rizalzaf @imcimon Tergantung jobdesc-nya. Kita biasanya nyebutnya ML research engineer, karena kerjaannya dekat ke riset. Skillsetnya juga beda dari SE: tensor programming, baca paper+implementasi, linear algebra

[Muhammad Firmansyah Kasim @mfkasim]

@austerralia @imrenagi Takut kena UU ITE :p

[Muhammad Firmansyah Kasim @mfkasim]

@imrenagi Seminarnya diadakan sama anak ITB, ngundang 2 alumni dan ternyata sama2 dari Oxford juga. Kata lainnya: "anak milenial terkenal". Hahaha

[Muhammad Firmansyah Kasim @mfkasim]

Can we parallelize RNN?

We can! Now you can see how on our repo: https://t.co/aWIdNWjN5J
This is the code implementation of the paper I shared last week: https://t.co/B5S7J5bs2U

[Muhammad Firmansyah Kasim @mfkasim]

Our work's post @Reddit is being featured on @slashML twitter!

[/MachineLearning @slashML]

Parallelizing RNN over its sequence length https://t.co/4AV8NC9QMk

[Muhammad Firmansyah Kasim @mfkasim]

@aria_ghora Non-convergence karena Newton's method memang ga ada jaminan bakalan converge, jadi kalau initial guess-nya jauh, bisa diverge bahkan. Paper-nya lumayan berat di matematikanya, jadi kalau ada pertanyaan, tanya aja.

[Muhammad Firmansyah Kasim @mfkasim]

Another drawback is scalability. Our method scales poorly: O(n^3) time complexity with the number of dimensions n. However, we can still achieve speed up using GRU with 64 dimensions (which is quite common value for experiments).

[Muhammad Firmansyah Kasim @mfkasim]

Really excited to share our new work: parallelizing RNN!

RNN is known to be non-parallelizable because its state depends on the previous state.
In our paper, we present a parallel algorithm to evaluate RNN, achieving orders of magnitude faster with a GPU!
https://t.co/cNdIOzcsXJ

[Muhammad Firmansyah Kasim @mfkasim]

The drawbacks: non-convergence and scalability. This is Newton's method and it has no global convergence guarantee. Although in our experiments the non-convergence is not that often, we expect that might occur more often in other experiments.

[Muhammad Firmansyah Kasim @mfkasim]

@aria_ghora Hahaha... Thank you! Gimana caranya bisa nyampe ke situ? Ini metodenya masih banyak bolong2nya, salah satunya non-convergence dan scalability, tapi insya Allah bisa jadi pembuka jalan untuk riset2 yang sejalan.