- 578
- 52 111
SupportVectors
United States
Приєднався 24 чер 2019
Welcome to our SupportVectors AI Lab channel! We are dedicated, as a community, to the learning of the broad field that encompasses the terms Machine-Learning, AI, Data-Science, Visualization, and Data-engineering.
Most of the videos here are recordings of lecture sessions from the various workshops offered at SupportVectors, where the speaker (and instructor) is Asif Qamar. (www.linkedin.com/in/asifqamar/)
The videos here are focused on understanding various concepts in the field, through discussion-sessions, code-walkthroughs, and careful readings of the various technical papers and articles.
Most of the videos here are recordings of lecture sessions from the various workshops offered at SupportVectors, where the speaker (and instructor) is Asif Qamar. (www.linkedin.com/in/asifqamar/)
The videos here are focused on understanding various concepts in the field, through discussion-sessions, code-walkthroughs, and careful readings of the various technical papers and articles.
[Paper Reading] Fastrack: Fast IO for Secure ML using GPU TEEs
As part of our weekly paper reading, we are going to cover the paper titled "Fast IO for Secure ML using GPU TEEs"
- Introduction to Secure Machine Learning: Emphasizes the need for end-to-end secure workflows using Trusted Execution Environments (TEEs) on CPUs (Intel Xeon) and GPUs (NVIDIA H100), leveraging AES-GCM encryption for secure communication.
- Motivation: Highlights the underutilization of GPU resources (15-20%) in enterprise AI and foundation models, despite high costs, driving the need for optimized GPU resource usage.
- Technical Innovations: Proposes three optimizations: direct GPU communication to eliminate redundant encryption, multi-chaining for parallel processing of data chunks, and an efficient decryption pipeline to overlap operations and reduce idle time.
- System Architecture: Details secure CPU-GPU communication, using shared encryption keys for data transmission, encryption, and Message Authentication Code (MAC) verification to ensure integrity.
- Performance Gains: Achieved up to 84.6% reduction in processing time for secure machine learning tasks with improved GPU utilization through parallelization and optimized data transfer.
- Evaluation Models: Tested on ResNet-50, GraphSAGE, and Two-Tower Neural Network models, demonstrating significant reductions in latency and improved throughput compared to baseline implementations.
- Impact and Future Work: Showcases practical benefits for enterprise AI, encourages further testing on next-generation GPUs like NVIDIA B100, and suggests improvements to default NVIDIA GPU drivers for broader adoption.
- Introduction to Secure Machine Learning: Emphasizes the need for end-to-end secure workflows using Trusted Execution Environments (TEEs) on CPUs (Intel Xeon) and GPUs (NVIDIA H100), leveraging AES-GCM encryption for secure communication.
- Motivation: Highlights the underutilization of GPU resources (15-20%) in enterprise AI and foundation models, despite high costs, driving the need for optimized GPU resource usage.
- Technical Innovations: Proposes three optimizations: direct GPU communication to eliminate redundant encryption, multi-chaining for parallel processing of data chunks, and an efficient decryption pipeline to overlap operations and reduce idle time.
- System Architecture: Details secure CPU-GPU communication, using shared encryption keys for data transmission, encryption, and Message Authentication Code (MAC) verification to ensure integrity.
- Performance Gains: Achieved up to 84.6% reduction in processing time for secure machine learning tasks with improved GPU utilization through parallelization and optimized data transfer.
- Evaluation Models: Tested on ResNet-50, GraphSAGE, and Two-Tower Neural Network models, demonstrating significant reductions in latency and improved throughput compared to baseline implementations.
- Impact and Future Work: Showcases practical benefits for enterprise AI, encourages further testing on next-generation GPUs like NVIDIA B100, and suggests improvements to default NVIDIA GPU drivers for broader adoption.
Переглядів: 46
Відео
![[Paper Reading] Do Mice Grok? Glimpses of Hidden Progress...](http://i.ytimg.com/vi/4TZRRrLKkvc/mqdefault.jpg)
![[Paper Reading] Do Mice Grok? Glimpses of Hidden Progress...](/img/tr.png)
[Paper Reading] Do Mice Grok? Glimpses of Hidden Progress...
Переглядів 67День тому
As part of our weekly paper reading, we are going to cover the paper titled "Do Mice Grok? Glimpses of Hidden Progress... " - Generalization vs. Memorization: The distinction between generalization (learning abstract patterns, e.g., "cow-ness") and memorization (recalling specific examples) in machine learning was emphasized, with grokking representing the shift from memorization to generalizat...
![[Paper Reading] Differential Transformer](http://i.ytimg.com/vi/i0pHvb7aSaA/mqdefault.jpg)
[Paper Reading] Differential Transformer
Переглядів 6714 днів тому
As part of our weekly paper reading, we are going to cover the paper titled "Differential Transformer "
![[Paper Reading] DRAGIN: Dynamic RAG Retrieval](http://i.ytimg.com/vi/K5dacRVjTF8/mqdefault.jpg)
[Paper Reading] DRAGIN: Dynamic RAG Retrieval
Переглядів 214Місяць тому
As part of our weekly paper reading, we are going to cover the paper titled "DRAGIN: Dynamic RAG Retrieval" Paper link - arxiv.org/pdf/2403.1008
![[Paper Reading] Astute RAG: Overcoming Imperfect Retrieval Augmentation and Knowledge Conflicts](http://i.ytimg.com/vi/1dMaLjs5c8A/mqdefault.jpg)
[Paper Reading] Astute RAG: Overcoming Imperfect Retrieval Augmentation and Knowledge Conflicts
Переглядів 89Місяць тому
As part of our weekly paper reading, we are going to cover the paper titled "Astute RAG: Overcoming Imperfect Retrieval Augmentation and Knowledge Conflicts" Here are a few key points from the paper reading session: - Inference-Time Optimization: The paper focuses on enhancing model response quality during inference, dynamically resolving knowledge conflicts without retraining or altering the m...
![[Paper Reading] LoRA vs Full Fine-Tuning: An Illusion of Equivalence](http://i.ytimg.com/vi/dLC5RwhUls0/mqdefault.jpg)
[Paper Reading] LoRA vs Full Fine-Tuning: An Illusion of Equivalence
Переглядів 122Місяць тому
As part of our weekly paper reading, we are going to cover the paper titled "LoRA vs Full Fine-Tuning: An Illusion of Equivalence" Here are a few key points from the paper reading session: - LoRA’s Advantage: Low-Rank Adaptation (LoRA) allows fine-tuning of large models with minimal resource requirements, targeting specific model areas to avoid costly full-model adjustments. - Orthogonality and...
Unlock Your AI Potential: Enroll in Our Cutting-Edge Lab-Based Courses!
Переглядів 4452 місяці тому
Unlock Your AI Potential: Enroll in Our Cutting-Edge Lab-Based Courses!
![[Paper Reading] The Unreasonable Ineffectiveness of the Deeper Layers](http://i.ytimg.com/vi/eUygWRw-Rg0/mqdefault.jpg)
[Paper Reading] The Unreasonable Ineffectiveness of the Deeper Layers
Переглядів 2438 місяців тому
[Paper Reading] The Unreasonable Ineffectiveness of the Deeper Layers
NLP with Transformers: Day 4 (April 08 2023)
Переглядів 209Рік тому
NLP with Transformers: Day 4 (April 08 2023)
NLP with Transformers: Day 3 (April 01 2023)
Переглядів 116Рік тому
NLP with Transformers: Day 3 (April 01 2023)
NLP with Transformers: Day 5 (April 15 2023)
Переглядів 97Рік тому
NLP with Transformers: Day 5 (April 15 2023)
NLP with Transformers: Day7 (May 6 2023)
Переглядів 79Рік тому
NLP with Transformers: Day7 (May 6 2023)
NLP with Transformers: Day 2 (March 25 2023 )
Переглядів 109Рік тому
NLP with Transformers: Day 2 (March 25 2023 )
NLP With Transformers: Day 1 (March 2023)
Переглядів 584Рік тому
NLP With Transformers: Day 1 (March 2023)
Lecture 1: Introduction to Survival Analysis
Переглядів 47Рік тому
Lecture 1: Introduction to Survival Analysis
Lecture 2: Introduction to Survival Analysis
Переглядів 26Рік тому
Lecture 2: Introduction to Survival Analysis
On Writing A Technical Blog: Sharing from personal experience
Переглядів 63Рік тому
On Writing A Technical Blog: Sharing from personal experience
ML100Sping2020 Lecture Session 6: Classifiers-intro
Переглядів 50Рік тому
ML100Sping2020 Lecture Session 6: Classifiers-intro
ML-100, Lecture 2 : Covariance, Correlation, Causality, Sequentiality & Regression towards the Mean
Переглядів 158Рік тому
ML-100, Lecture 2 : Covariance, Correlation, Causality, Sequentiality & Regression towards the Mean
The Lab on Data Clustering Saturday May 1
Переглядів 23Рік тому
The Lab on Data Clustering Saturday May 1
Week 0 Quiz Review (ML400: Deep Learning Workshop)
Переглядів 32Рік тому
Week 0 Quiz Review (ML400: Deep Learning Workshop)
Olah's Blog: Neural Networks, Manifolds and Topology
Переглядів 77Рік тому
Olah's Blog: Neural Networks, Manifolds and Topology
Thank you in advance for this wonderful course! I can not thank you enough!
Absolute masterpiece in explaination
Thank you so much for the lectures on the current trends.
Great lecture sir. What are your thoughts on transformer based OCRs (specifically; Paddle OCR)? Do you think they have the capability to compete with DONUT architecture?
PaddleOCR is still going strong because of its immediate practical application. However, I think over time a direct visual understanding models like DONUT, ColPali, etc. will evolve in due time to something more powerful, and the OCR approach may gradually fade out. We are not there yet -- perhaps another couple of years before that happens.
I watched your starter lectures, they were amazing, but after watching this, I think useless talk is more than valuable talk. Please don't take it wrong, it's just feedback.
This is very true -- these are raw recordings of the lecture sessions. Since those sessions happen with a lot of participants having a conversation in the lobby, the break-room, and there being jokes and digressions, these raw recordings need a lot of post-editing to reduce them to half the size. Unfortunately, currently we do not have a video-grapher on staff to clean out the videos and reduce the distractions, but perhaps one day!
Thank Dr.
Can you share the colab notebook
Very useful explain of the optimization
This is a goldmine of knowledge!
I really enjoyed this ❤
Thank you for this perfect explanation
attention is selective concentration on sub-strata with an intent.
Is there a GitHub repository where I can get all these things in one place
❤❤❤
Very well explanation 👏🏼
awesome lectures ! where can i get the jyupter notebooks ? is the full course available -- with jyupter notebooks ? thanks
Thank you for your time and efforts. This is really good explanation with intuition . ELI5 level .
Thank you sir very well explained 👍
very nice explanation of the Buzz words and force us to think what actually AI is, and how can we relate it human understanding and intelligence. Basic terms of ML are described very well . Its a introductory video, excited to see the actual content ad future videos in this course.
The case studies presented give a nice idea on how AI can be implemented in the core engineering design. Even graduate students (at least myself) are only exposed to only analytical and numerical methods for solving design and analysis problems. Very often it takes huge amount of time before we get to see crude results of analysis to verify hypotheses. This lecture arouses curiosity to apply AI methods to get results faster enabling quicker learning cycles!
Very informative session sir 😊
Thank you, Rishab
yawning .....yawning....yawning....very annoying....either sleep or give lecture
Very good articulation and presentation of the problem and roadmap towards ML/AI solutions in engineering design. Thanks Prof Kannan.
Grateful for your comments. Look forward to working closely with you.
Very insightful. The video is perfect for a layman (like me :-)) as well as tech geeks, stimulating interest in some of the things we take for granted.
Thank you, Karthik. Of course, it's another matter that I know you are a good geek!
i AM CURRENTLY DOING PHD FROM MALAYSIA...THIS IS THE BEST CONTENT I HAVE EVER SEEN....TEACHER IS GOOD...LOTS OF LOVE PLEASE UPLOAD ALL VIDEOS....AS FEW VIDEOS ARE PRIVATE....PLEASE MAKE THEM AVAILABLE .THANKS LOTS OF LOVE...AND GOD BLESS YOU
Thank you for the kind words! I am new to UA-cam, and just getting started with posting the lectures here. Will make all the remaining videos public (was hoping to edit them a bit to remove 15-minute tea-break in the middle of the sessions)
amazinglecture sir
Can you have a code please?