Amir Habibian Profile Banner
Amir Habibian Profile
Amir Habibian

@amir_habibian

1,260
Followers
187
Following
13
Media
94
Statuses

Research Scientist (Director) @Qualcomm AI Research

Amsterdam, The Netherlands
Joined August 2014
Don't wanna be here? Send us removal request.
@amir_habibian
Amir Habibian
3 years
Do we need to process every single pixel in a video? TLDR: Compute the features only at the pixels that convey new information. Result: 4~5x less compute without any performance drop. Manuscript:
Tweet media one
7
39
244
@amir_habibian
Amir Habibian
9 months
Multiple PhD internships available in Qualcomm AI Research, Amsterdam Topic: - Image/video generation and editing - Efficiency of diffusion models - Efficient architectures - Adopting generative models to improve down-stream tasks DM me for any question!
5
25
132
@amir_habibian
Amir Habibian
9 months
Five fellowships (each 40,000$) available for PhD students in European universities. More info: DM if have any question.
Tweet media one
2
11
82
@amir_habibian
Amir Habibian
10 months
Do we need to run the whole UNet for all the diffusion steps? No! Accelerating your Diffusion Model with a simple trick, even without retraining ?! w/ Amir Ghodrati, @noor_fathima_ , @gsautiere , Fatih Porikli and @peterjensen_ Code: Stay tuned
Tweet media one
1
17
80
@amir_habibian
Amir Habibian
2 years
Several research internship positions are available in Qualcomm AI Research in Amsterdam. Interested in efficient computer vision? Shoot me your CV. #video_perception , #video_compression , #video_enhancement , #conditional_compute , #vision_transformers
1
9
76
@amir_habibian
Amir Habibian
8 months
Qualcomm launched AI-Hub, a comprehensive suite of models highly optimized to leverage all compute cores (CPU, GPU, NPU) available on your phone with a Snapdragon! Lots of cool models e.g, Segment anything, Stable Diffusion and Whisper.
1
10
59
@amir_habibian
Amir Habibian
8 months
Internship positions available in Qualcomm AI Research - Amsterdam. Topic: Generative image/video models Requirements: PhD student with relevant publications Application:
1
8
56
@amir_habibian
Amir Habibian
9 months
A summary of our recent research on efficient diffusion models for images, videos and 3D: arXivs: - - - - #GenerativeAI #diffusion #efficient
0
6
55
@amir_habibian
Amir Habibian
2 years
2/2 accepted to #ECCV2022
1
0
44
@amir_habibian
Amir Habibian
7 months
Are you interested in making foundation vision models faster at training, adaptation, and inference? Please checkout the @CVPR workshop on Efficient Large Vision Models
0
9
47
@amir_habibian
Amir Habibian
9 months
Are you interested in making foundation vision models faster at training, adaptation, and inference? Please checkout our @CVPR workshop on Efficient Large Vision Models
Tweet media one
1
8
39
@amir_habibian
Amir Habibian
9 months
Object Centric Diffusion for fast video editing TLDR: 1. Tokens can be merged a lot especially along the temporal axis and on the background regions 2. Different sampler for foreground and background: More sampling steps foreground/edited regions
Tweet media one
Tweet media two
1
4
34
@amir_habibian
Amir Habibian
8 months
We are looking for seasoned academic/industrial researchers joining us to direct Qualcomm AI Research in Amsterdam. A sizable team of talented research scientists with a huge diversity in research background and interests. DM me if you are interested!
0
6
34
@amir_habibian
Amir Habibian
9 months
We are looking for a couple of high profile senior scientists to join Qualcomm AI Research in Amsterdam. An excellent team with great diversity of topics, publishing dozens of papers each year in top ML and CV venues: Job ad:
Tweet media one
0
6
33
@amir_habibian
Amir Habibian
9 months
Glad to see Qualcomm CEO promoting the research from our team!
@cristianoamon
Cristiano R. Amon
9 months
Creating photos, videos, and 3D content with generative #AI requires immense computational power and large memory bandwidth. @Qualcomm ’s AI research is changing that by making #GenAI more efficient. Learn about the key techniques we’ve developed.
0
2
18
1
1
30
@amir_habibian
Amir Habibian
4 years
2/3 papers accepted #CVPR2021
2
0
27
@amir_habibian
Amir Habibian
9 months
@CVPR workshop on efficient training and inference for large vision model
Tweet media one
@CVPR
#CVPR2025
9 months
The list of #CVPR2024 workshops is now available!
Tweet media one
2
35
188
1
6
25
@amir_habibian
Amir Habibian
9 months
Interested in efficient diffusion models generating images, videos, and 3D objects? Join us in the next Qualcomm webinar on Jan 25
@QCOMResearch
Qualcomm Research & Technologies
9 months
With a text prompt and a click of a button you can now generate brand new 2D images, 3D images, and even videos. But at what cost? While this technology has great potential, it faces challenges which we will address in this #AI webinar. Sign up here:
Tweet media one
0
4
13
0
2
14
@amir_habibian
Amir Habibian
3 years
An interview discussing our recent research on efficient video processing
@QCOMResearch
Qualcomm Research & Technologies
3 years
Checkout this @twimlai podcast featuring @Qualcomm #AI Research's very own @amir_habibian where he discusses his #CVPR2021 accepted papers, efficient video perception, and a neural video decoder demo.
0
7
15
0
0
14
@amir_habibian
Amir Habibian
9 months
Simple trick to accelerate your transformer by skipping the correlated self-attention computations across adjacent blocks. Our new paper to be presented in #ICLR2024
@shawshank_v
Shashank
9 months
Delighted to share that SkipAT has been accepted to #ICLR2024 @iclr_conf This work was done as part of my internship at @QCOMResearch together with some amazing researchers - Amir Ghodrati, @y_m_asano , Fatih and @amir_habibian . Please check the thread 👇 for more details.
2
2
33
0
1
13
@amir_habibian
Amir Habibian
10 months
Folks attending #CES24 , please checkout our demo on using diffusion models to generate data for automotive perception at Qualcomm booth! #CES #qualcomm #Automotive #GenerativeAI
Tweet media one
Tweet media two
0
1
10
@amir_habibian
Amir Habibian
3 years
Stop processing frames as early as possible ... How? Train binary gates jointly with the classifier
@BabakEht
Babak Ehteshami Bejnordi
3 years
How many frames are needed to reliably recognize an action? #FrameExit uses self-supervised gates to adjust the computation to the difficulty of the input video. Check out our #Oral #CVPR paper: #CVPR2021 Great collaboration with @ghodrati & @amir_habibian
Tweet media one
4
5
37
0
0
10
@amir_habibian
Amir Habibian
9 months
How to keep up reading the 11K papers came out in top CV and ML conferences in 2023? It will be even more in 2024 Great tips and tool posted by @ProfJasonCorso
0
0
9
@amir_habibian
Amir Habibian
4 years
@CSProfKGD @TheBMVA Still no update?!
0
0
5
@amir_habibian
Amir Habibian
8 months
Models also available on #huggingface
0
0
5
@amir_habibian
Amir Habibian
3 years
@wellingmax I hope this be the case for everywhere, including the wars started in the middle east :)
1
0
3
@amir_habibian
Amir Habibian
3 years
@wellingmax Thanks for your enormous contributions. We will miss you and hope to see you around 🙂
0
0
3
@amir_habibian
Amir Habibian
3 years
A big step towards neural video codecs
@QCOMResearch
Qualcomm Research & Technologies
3 years
Check out @Qualcomm #AI Research's latest breakthrough: the world’s first software-based neural video decoder running HD format in real-time on a commercial smartphone. Learn more:
0
25
81
0
0
3
@amir_habibian
Amir Habibian
3 years
@LeopolisDream Indeed, people have tried to sparsify convolutions on images (e.g. ). But learning what to process is way easier for videos because of strong motion clues.
1
0
3
@amir_habibian
Amir Habibian
2 years
@LucaAmb @jmtomczak Thanks guys👍
0
0
2
@amir_habibian
Amir Habibian
9 months
@ma_jauhar Suggest sending out your CV to habibian @qti .qualcomm.com and I will respond. In some cases, we can arrange internship for highly talented master students
2
0
2
@amir_habibian
Amir Habibian
9 months
0
0
2
@amir_habibian
Amir Habibian
3 years
@benno_krojer Good point! In human brain the information is processed only when the amount of activations exceeds some threshold as modeled by Spiking Neural Networks. The residual processing and gating functions mimic a similar behaviour for feedforwars Networks.
0
0
2
@amir_habibian
Amir Habibian
9 months
@asaad_balla Yes, it is
1
0
1
@amir_habibian
Amir Habibian
4 years
@mentzer_f Congrats Dr. Mentzer!
0
0
1
@amir_habibian
Amir Habibian
3 years
@yperugachidiaz Congrats Yura👍👍
0
0
1
@amir_habibian
Amir Habibian
9 months
0
0
1
@amir_habibian
Amir Habibian
4 years
@3scorciav Indeed, let us hope the next conference be physical! Sure, pls ping me if you come around😊 Cheers
0
0
1
@amir_habibian
Amir Habibian
3 years
@jmtomczak Thanks Jacub 👍
0
0
1
@amir_habibian
Amir Habibian
9 months
@jaom7 Thanks for your interest! Are students affiliated with KUL?
1
0
1
@amir_habibian
Amir Habibian
9 months
@kkahatapitiy @Qualcomm Likewise, fun and fruitful collaboration! Best of luck with the rest of your PhD :)
1
0
1
@amir_habibian
Amir Habibian
7 months
1
0
1
@amir_habibian
Amir Habibian
3 years
@BabakEht @ghodrati FrExit? Is France following UK?
1
0
1