Abduallah Mohamed

Applied Research Scientist @Meta, Solving machine perception problems. PhD, @UTAustin Alumni, Motion and action prediction, graph NNs

[Abduallah Mohamed @abduallahaom]

@KL_Div I used IMLE in my 2022 work, Social-Implicit where it proved its able to model the motion distribution covering multiple modalities. https://t.co/3VGWD7VpLL

[Emile van Krieken @EmilevanKrieken]

@abduallahaom @karpathy @simonw Another good to is to write down variable names with tensor shapes

[Abduallah Mohamed @abduallahaom]

@karpathy @simonw Have you tried typing a comment of what you intend to do before writing a code. Most of times it works fine if I did this

[Abduallah Mohamed @abduallahaom]

@2ayasalama I’d cluster the data set in unsupervised manner. Sample (weighted) from these clusters.

[Abduallah Mohamed @abduallahaom]

6) Apparently, the link I posted earlier is not working: https://t.co/hnQ1aNPYsR with paper link here: https://t.co/3VGWD8csNL and interactive demo here: https://t.co/xfpgQn1DhN

[Abduallah Mohamed @abduallahaom]

Our paper "Social-Implicit: Rethinking Trajectory Prediction Evaluation and The Effectiveness of Implicit Maximum Likelihood Estimation" is accepted @ #ECCV2022 It redefines SOTA in motion prediction and redefines the metrics used in this area. Details: https://t.co/wDU8FLiYxu

[Abduallah Mohamed @abduallahaom]

5) Thanks to my collaborators @cclaudel70 @moElhoseiny @ Deyao Zhu & @ Warren Vu.

[Abduallah Mohamed @abduallahaom]

4) Beside showing the core issue of the previous metrics BoN ADE/FDE and introducing the new metrics AMD & AMV we introduce Social-Implicit, a light weight deep graph models that achieves SOTA in motion prediction tasks with real-runtime inference and low memory footprint.

[Abduallah Mohamed @abduallahaom]

3) These new metrics directed us to use IMLE (Implicit Maximum Likelihood Estimation) which is suitable to control the behavior of individually generated samples.

[Abduallah Mohamed @abduallahaom]

2) Form this core issue we introduce two new metrics AMD( Average Mahalanobis Distance) & AMV (Average Maximum Eigenvalue) to evaluate how close a predicted distribution to the ground truth and measure the spread of predictions around this ground truth

[Abduallah Mohamed @abduallahaom]

1) We target a major issue in current evaluation metrics Best-of-N ADE/FDE and show that they are insensitive to predicted distributions and that current/previous reported models depends on outliers or lucky predictions to report a SOTA.

[Learning on Graphs Conference 2026 @LogConference]

Here it is: the first Learning on Graphs Conference! 🎊
We think this new venue will be valuable for the Graph/Geometric Machine Learning community.

What makes it so important+unique? See our blog post!
https://t.co/WPPbxxLjrE
1/6

[Abduallah Mohamed @abduallahaom]

Test

[Sara El-Ateif 🇲🇦 @el_ateifSara]

I will present "Fundamentals of Deep Learning" at the #NVIDIA #GTC22 conference this year as Lead of @TensorflowC part of the #NVIDIA Emerging Chapters.
Tune in and you might win $100-500 USD in DLI AI course coupons.
Link: https://t.co/9EakrypYsU

[Julien Chaumond @julien_c]

Karpathy dropped a new blog and as usual it's 🤯😮

https://t.co/LS6FkxXI7b

[Abduallah Mohamed @abduallahaom]

Excited to share our work: "HAR-GCNN" wish will appear IEEE PerCom CoMoRea 2022. We show that humans activities have an implicit chronology that can be leveraged to predict their future activities also identifying the older activities. More details: https://t.co/kaRXPjvbRM

[Abduallah Mohamed @abduallahaom]

@el_ateifSara @gusthema 😂

[Abduallah Mohamed @abduallahaom]

Just a friendly reminder that tomorrow’s the 31st by midnight PST is the due date of @CVPR rebuttal.

[Michael Black @Michael_J_Black]

I get a lot of reviews that say my work is not novel and I bet I'm not alone. It's always frustrating because I see novelty where the reviewer doesn't. Rather than rebut every critique, I've written a blog post to help reviewers think about novelty. https://t.co/UXLabOkYcn

[Abduallah Mohamed @abduallahaom]

Something about @CVPR rebuttals bugs me. I can understand why 1 page of rebuttal, but why not 1 page of text rebuttal and another one only for figures and tables? Usually, tables and figures consume space in latex and are difficult to organize … #CVPR2022