AiBro

PostDoc, ex researcher @KITKarlsruhe, @TU_Muenchen, @BoschGlobal.🧐 #AI #Neuromorphic systems and #uncertainty quantification

[AiBro @tuhinmailme]

@nurcantureen Hey 👋

[AiBro @tuhinmailme]

@sferik Top 10 things that never happened

[AiBro @tuhinmailme]

Beauty

[AiBro @tuhinmailme]

@sumitdotml You can add me

[AiBro @tuhinmailme]

@jubueche In terms of what?

[AiBro @tuhinmailme]

@FabrizioRomano It’s like saying. I never had any contact from Margot Robbie for becoming my wife. I remain focused on my wife. I am 100% focused.

[AiBro @tuhinmailme]

Visiting Bosnia has to the most shit decision ever. It’s like visiting North Korea.

My advice stay away.

[AiBro @tuhinmailme]

@thegautamkamath @NYU_Courant Congratulations. Are you taking postdocs?

[AiBro @tuhinmailme]

#uncertainy #Ensemble #deeplearning #machinelearning

[AiBro @tuhinmailme]

Latest Llama 3.2 offers lightweight models + multimodal large language models that can see, reason, and understand

Model Size:
• Lightweight: 1B and 3B
• Multimodal: 11B and 90B
- Outperforms most open and closed multimodal models on industry benchmarks!
-Takes Llama 3.1 to the next level with vision adapters and cross-attention layers.

Key Use Cases:
• Visual Question Answering (VQA): Ask questions about an image and get accurate answers!
• Document Visual QA: Understand complex docs like maps and contracts through images
• Image Captioning: Generate captions that tell a story from visual data
• Image-Text Retrieval: Match images and descriptions seamlessly, like an AI-powered search engine
• Visual Grounding: Pinpoint specific objects in images through natural language

[AiBro @tuhinmailme]

Man united fans right now

[AiBro @tuhinmailme]

> Show Sigmoid Attention matches Softmax Attention in various tasks and domains

Code: https://t.co/lAekHyO8Nl

[AiBro @tuhinmailme]

Flash Sigmoid Attention (SA) - 17% faster than Flash Attention (FA) 2

> Paper proved SA is a universal function approximator on seq-to-seq tasks
> Extend FA 2 with the sigmoid kernel, reducing kernel inference wall-clock time by 17% and real-world inference by 8%

[AiBro @tuhinmailme]

- MMLU test set available in 14 languages, translated by professional human translator
- Covers 57 categories from elementary to advanced professional subjects
- Evaluates AI models' general knowledge across diverse cultures, used in openai/simple-evals

[AiBro @tuhinmailme]

Open #Dataset release by OpenAI! #OpenAI just released a Multilingual Massive Multitask Language Understanding (#MMMLU) dataset to evaluate multilingual #LLMs more easily!

Dataset: https://t.co/gC3qayrkKl

[AiBro @tuhinmailme]

@Kaju_Nut There is no such thing as European 3 years PhD. Each country, every institution, and even group has their own rules.

[AiBro @tuhinmailme]

#uncertainy #Ensemble #deeplearning #machinelearning

[Artificial Intelligence Papers @SciFi]

Tiny Deep Ensemble: Uncertainty Estimation in Edge AI Accelerators via Ensembling Normalization Layers with Shared Weights. https://t.co/58JBKFdKpB

[AiBro @tuhinmailme]

@wellingmax Very Interested