Steven Braun

PhD Student working on #tractable #probabilistic modeling at the AIML Lab, @CS_TUDarmstadt with @kerstingAIML, Mastodon:

[Martin Mundt @mundt_martin]

WOW - our work "Deep Classifier Mimicry without Data Access" was awarded student paper highlight @aistats_conf #AISTATS2024!🔥

tl;dr: we introduced model-agnostic knowledge distillation without data access!

Couldn't be ANY prouder of @sbraunmz & sharing the joy w/ @kerstingAIML

[Kristian Kersting @kerstingAIML]

🏆🎖️Super proud supervisor moment 🥳 @sbraunmz won an #AISTATS2024 Outstanding Student Paper Highlight Award🎉 His work shows how to mimic deep classifiers without access to the original data. Perfect for all the pre-trained models without data access!

https://t.co/xYEoyTiGyq

[Martin Mundt @mundt_martin]

I’m sadly not at #AISTATS2024, but *please* go see @sbraunmz present our work “Deep Classifier Mimicry without Data Access” on Saturday May 4th at 10:30-11:30 (oral session 8)

It’s a really fun paper that shows how to distill classifiers without data access across model types 🤩

[Steven Braun @sbraunmz]

@tetraduzione @ccanonne_ @zulip @Mattermost is amazing as well and comes for free with its self-hosting option. I’m maintaining our server and only ran into upgrading issues once in four years.

On the other hand @zulip even offers a free version, hosted by zulip, for research groups, so that’s a big plus.

[Martin Mundt @mundt_martin]

🎉 Our work “Deep Classifier Mimicry without Data Access” w/ @sbraunmz @kerstingAIML was accepted as an ✨oral✨at #AISTATS2024

https://t.co/3SpT6AlaXI

🍰: data-free & model-agnostic knowledge distillation!

ICLR spotlight, AIStats oral; this year is truly off to a great start

[Steven Braun @sbraunmz]

@Mattermost Awesome, thanks!

[Steven Braun @sbraunmz]

@jmmldotme @typstapp Sure, but that a) take a lot of time to get done right and b) risks a desk reject at the venue if I should get something wrong. As I said, for this to be adopted by authors, the venues themselves would have to provide templates.

[TU Darmstadt @TUDarmstadt]

Wie kann ich mich schon im Studium auf KI spezialisieren? Und wie finde ich später einen Job? #TUDa Informatik-Professor Kristian Kersting @kerstingAIML @CS_TUDarmstadt beantwortet im Interview @zeitcampus die wichtigsten Fragen 👇 https://t.co/rVK3TpsepE via

[Kristian Kersting @kerstingAIML]

🥳Super thrilled that my fantastic team has managed to get six #Neurips2023 papers accepted: XAI for multimodal transformer, interleaving inputs of multiple modalities & languages, semantic guidance for text2image, causality, neurosymbolic AI for RL & characteristic circuits 🙏

[Wolfgang Stammer @WolfStammer]

Finally our novel Learning by Self-eXplaining (LSX) approach is online: https://t.co/WFwm6GJYv0
For the first time a model is trained by explaining to and receiving feedback from an internal critic, going beyond standard accuracy benefits. Check it out! :) #kerstingAIML #hkrsnd

[Martin Mundt @mundt_martin]

I'm featured in this great article (in 🇩🇪) on unlearning + forgetting in AI, adding perspective to the challenges & highlighting the importance of our research @CS_TUDarmstadt @Hessian_AI & @ContinualAI

Critical to raise the public's awareness!
🙏 @handelsblatt @LarissaHolzki

[Steven Braun @sbraunmz]

@BlackHC If your code is encapsulated in a module you could do the following when your in pdb or some other interactive debugger:

import importlib
import my_module

importlib.reload(my_module)
# Continue execution with the modified code

[Steven Braun @sbraunmz]

Interested in how to tackle overconfidence and introduce uncertainty in #ProbabilisticCircuits? 🧠

Join our oral pres. on "Probabilistic Circuits That Know What They Don’t Know" Thu 3:15PM and chat with us at our poster (118) later at 4:30PM @ #UAI2023! 🎉

[Kristian Kersting @kerstingAIML]

🙏 für das Interview! Es gäbe noch so viel mehr zu sagen. Wir brauchen den Dialog, national & international. Das Hier& Jetzt bei #KI und ihren Chancen/Risiken sind entscheidend. Wir dürfen bei der Gestaltung von KI nicht nur passiv zuschauen

[Steven Braun @sbraunmz]

@Hessian_AI @ds_darmstadtium Were the talks recorded and if so, are they going to be uploaded somewhere?

[Martin Mundt @mundt_martin]

🚨Announcing Self-Expanding Neural Networks (SENN): https://t.co/QrYBkWyAV8 @kerstingAIML

SENN can start with 1 unit in 1 layer & use natural gradients to jointly answer when/where/what to add during training. Growing on-the-fly!

New cornerstone for resource-aware & adaptive ML

[Wolfgang Stammer @WolfStammer]

Excited to introduce our novel dataset for visual logical learning, allowing for challenging tasks across the spectrum of AI approaches:
https://t.co/jquP6yGwqR
@lukas_helff @hkrsnd @devendratweetin @kerstingAIML

[Martin Mundt @mundt_martin]

🚨Care to distill deep classifiers' knowledge but don't have access to data & model internals?

In "Deep Classifier Mimicry without Data Access" https://t.co/3SpT6AlaXI we show model-agnostic distillation without access to intermediate layers or data!
🧵👇 @sbraunmz @kerstingAIML