A Summary of my activities and projects in AI since 2018

joepareti54@gmail.com

January 8, 2021

updated on January 29, February 15-19, March 7,19,20, April 6, 25, May 7, 31, June 6, July 9, 18,27, August 7, 23, 30 ;September 4, 7, November 29, December 16, January 2022, February 9, 2022 - April 2022- June, July, August 2022, October 2022 December 2022 February 2023, March 2023, April 2023, June 2023, July 2023, August 2023, October 2023, November 2023, December 2023 January 2024 February 2024 March 2024 April 2024 June 2024 June 2025 August 2026

​

‘Keep inventing, and don’t despair when at first the idea looks crazy.

Remember to wander. Let curiosity be your compass. It remains Day 1.’

• Jeff Bezos, founder of Amazon

​

​

​

​

Contents

Website & Brand Development

GitHub

Machine Learning Applications in Bio-science and Engineering

Key technology partners and industry trends in AI & Data Science

ML resources

Quantum

Exascale

Industrie4.0

Guidepoint HPC Cloud

Backup slides and Archived slides on website development and other projects

NLP training with attention models

Projects

​

​

Website & Brand Development

December 2021

current website

This is our current website

​

ad this is our linkedin group

​

traction and visibility in the marketplace should improve; the next slides in this paragraph are about ways to accelerate the business. Efforts in early 2021 led to the website launch in May 2021

​

group narrative

contents

1. go narrative workshop and narrative text - NEW
2. brand
3. commercial offers to improve our website

​

​

​

preliminary work, including a compass chart - June 2021 timeframe

GoNarrative workshop

Detailed slides

cut and paste from the mural file to make sure we do not lose the content

Voting

at the end of the workshop we identified a subset of bullet points

Results

how can we use what we learnt in the workshop to make a more effective value proposition

go Narrative - detailed slides

Transformation

Reasons to believe

Innovation

Problems

___________________

Desire

Difficulty

Denouement

​

Detailed: Transformation: technology changes

Detailed: Transformation: regulations & policies

Detailed: Transformation: trends affecting customer

Detailed: Transformation: how changes affect customers

Detailed: Transformation: how changes affect markets

Detailed: Transformation: how changes affect individuals

Detailed: Reasons to believe: facts&figures

Detailed: Reasons to believe: urgency

Detailed: Reasons to believe: voice of experts

Detailed: Reasons to believe: press

Detailed: Reasons to believe: aspiration

Detailed: Reasons to believe: cost of avoiding

Detailed:Innovation: killer feature

Detailed:Innovation: whole product

Detailed:Innovation: company differentiator

Detailed:Innovation: personal differentiator

Detailed:Innovation: results

Detailed:Innovation: why you

Detailed: Problems due to transformation

Detailed: Problems to get transformation going

Detailed: Problems: roadblocks

Detailed: additional Problems:

Detailed: Problems: what if not embarking

Detailed: Internal Problems:

Desire: purpose or mission

Desire: pain to be addressed

Desire: promised land

Desire: regulations and policies

Desire: awards and employees demands

Desire: social media image

Difficulty: blocker

Difficulty: roadblocks

Difficulty: villain

Difficulty: internal blockers

Difficulty: other factors

Denouement: overcoming roadblocks

Denouement: results

Denouement: transformation, how?

Denouement: 3rd party components

Denouement: take away

Denouement: how we help

GoNarrative: voting

GoNarrative: voting - transformation

GoNarrative: voting - reasons to believe

GoNarrative: voting - innovation

GoNarrative: voting - problems

go Narrative workshop results

Out of 3 voting participants, it is statistically impossible to draw any conclusions.

​

the 'number of unique voters ' is = 1 in 2 out of the 4 blocks which implies team members gave their own box a higher vote.

​

we are differing practitioners, each with their own agenda and priorities

​

How to implement? NEW here is a proposed narrative text

Brand

• Find your WHY
• I believe I have found my why: I want a piece of the action in the fast growing AI business
• How about your WHY?

​

commercial offers to improve our website

1. Florian Heinke provides web design but no programming
1. What are our objectives
2. Do we need a login
3. Do we need to design a brand, do we need to redesign the logo
4. The cost ranges from EUR 1K to 9K
2. Ritu Sharma makes an offer for EUR 500
• they propose a whole new design, but we want to stay on WordPress
3. Terry Fisher’s proposal
• professional design & programming : EUR 2500
• simpler WS for EUR500
4. hellophello.com : They coach you into selecting 300 Max key contacts and then emailing them periodically to get a phone conversation to sustain mind share.

​

offer to maintain the website

to be proposed

GitHub

GitHub

• While linkedin is great for generic business content, GitHub is THE platform for sharing software & building a community around an idea or software prototype
• Ideally, one could combine a blog with a github repo such as this blog on using Machine Learning to predict diabetes using patient data; the source code is available here.
• EVERYONE WORKING ON CODE IS ENCOURAGED TO USE GITHUB AND KEEP REPOS UPDATED

Creating a GitHub repo for project to Identify Customer Segments

• Identify customer segments to increase sales. It is based on a demographic file containing ~900K records, and a ~200K customers’ records file.
• This is an application of unsupervised learning techniques such as k-means clustering and of Principal Components Analysis.
• This project was part of the Udacity ML introduction training completed in 2020

​

Creating a GitHub repo for the Image Classification Project using PyTorch

This project was part of the Udacity ML introduction training in 2020

​

This GitHub repo and this are a preparation step for the full fledge project:

• specific code in Pytorch used for forward propagation: classes & methods
• Gradient calculation tutorial cases using Autograd are illustrated here.
• Additional Experimentation with PyTorch : trainloader, loss, SoftMax, Inference, Checkpoint

Checkpoint in PyTorch

• Repo in GitHub with code for explaining how to checkpoint & load a PyTorch model based on Udacity code for checkpointing
• model is the original model used for training
• model1 is loaded from the checkpoint file
• inference on the SAME image, out of the F_MNIST dataset gives the same results

​

Machine Learning Applications in Bio-science and Engineering

the slides on bio-sciences have been removed from this document: please refer to this deck which is maintained and updated as I learn more on ML for bio-science

NEW: workshop on AI for CFD - Introductory

Simple Content for an Introduction pass1

​

( pass2 )

playbook for this series

pass3 currently the latest release is in pass3

pass5

​

Project about ML libraries for CAE : pass0

NEW: workshop on AI for CFD - Advanced

​

slides and a short script

​

NEW - updated in June 2025 - German script , updated document and a linkedin summary-of-updates

​

(just for reference: content presented to a conference in June 2023 slides, script )

NEW - Exafoam workshop in June 2023

Exafoam : under construction, blog

​

Sample Benchmark

Open Source software for ML projects

presentation in June 2022

​

video

surrogate models integrating HPC and ML

here’s my linkedin article : covering a model developed by a startup collaborating with UberCloud, and advanced surrogate models for turbulent flow simulations at ISC2021

​

An integrated CAE Framework that includes generative models, Convolution Autoencoder, Design for Experiment and Transfer Learning to automate a CAD / CAE pipeline

​

Nvidia SimNet: calculation of the full spatial-temporal velocity field in a blood vessel aneurysm

RocketML workshop at SC21 on using PINNs

• Introduction on Scientific Machine Learning [211114_0647: min6-24]
• Azure HPC Cloud [211114_0647: min24-46]
• Building and deploying PINN based solvers [211114_0647: min46-end]
• Implementing PINNs and hands-on session [211114_0648:]
• Building and deploying CNN based solvers [211114_0649:] - Aditya Balu
• Slack channel
• Alternative implementation of the 1D PDE explained in the tutorial
• commercial providers?

Commercial CAE Applications using Machine Learning

Nvidia Simnet: uses PINNs, is supported in SolidWorks and has a roadmap. Target use case: intelligent CAD

Monolith AI : train a deep neural network to learn STAR CCM solutions of the transient flow in an engine combustion chamber

RocketML : use PINNs and provide a DevOps environment

CometML: Mlops/devops startup raising VC

Xitadel works with Beta Software on plastic parts models for Automotive

Nvidia Modulus: announced at GTC 2021

Applying Deep Learning to turbulent flow models

Prof Brunton’s review:

• Apply a Physics Informed Neural Network, or PINN to solve the Reynolds Averaged Navier-Stokes equations:
• the mathematical basis is work by Pope in 1975, augmented by equations that allow a PINN to be built that directly calculates the Reynolds stresses

ISC 2021 review

• LES and Deep Learning, also using a PINN
• Deep Reinforcement Learning applied to turbulent flow modeling

How AI/ML/DL make HPC applications run faster

• Surrogate Models, e.g. Deep Learning For Steady-State Fluid Flow Prediction In The Advania Data Centers Cloud
• Machine Learning for Fluid Dynamics: great youtube video, showing that you can apply a similar approach to solving for turbulent flow as you do in computer vision; it also shows autoencoders and PCA
• this Sorbonne white paper is about learning PDEs from data:
• Solving PDE with NNs ‐ Reduced models
• Dealing with partially observed data
• Combining physic models and NNs
• PINNs: the partial derivatives are calculated WITHIN the NN using the SGD algorithm

​

​

​

More use cases on CAE/HPC using ML

Machine Learning: practical use cases for CAE models and material database

​

Fugaku Leverages AI for Tsunami Flood Prediction

​

​

​

​

ML for nuclear fusion

ITER

ITER will confine a plasma almost 10 times greater in volume than today’s largest experimental magnetic confinement fusion device. Although ITER’s scale is expected to yield a higher fusion power output, it presents new technical challenges because plasma performance can be more difficult to predict at larger scales.

​

Summit and Theta Supercomputers Power ML Models for Fusion Energy Research

In building any fusion reactor in the future, predicting the heat-load width is going to be critical to ensuring the divertor material maintains its integrity when faced with this exhaust heat

​

Applications of deep learning predict disruption

Deep Learning and GANs for High Energy Physics

CERN slides - 2019 on preparation for the high luminosity LHC

Initial Collaboration with UberCloud in 2018 on AI for CFD

Review of white papers advised by Nvidia

Key technology partners and industry trends in AI & Data Science

​

NEW - GTC March 2024

AI for Molecular Modeling: slides, document

STARCCM

Drug Discovery

Astellas Pharma BioNemo: Slides, Document

Nvidia Inference Microservices - LLM Slides Document

Omniverse

CERN

(Before GTC: LLMs, Prompt Engineering, LangChain): Slides, Document

​

NEW - GTC March 2024 - continued

​

Generative AI for Financial Services

RAG: Slides, Document

AI for med tech: document , linkedin post

Generative AI for medicine

AI Agents Code

Marketing

Keynote

Supercomputing

Lowe

​

AWS reInvent recap in 2024

Slides

​

​

document

​

​

AWS reInvent 2023

• CEO keynote
• Bedrock: A service to access, customize, and integrate foundation models (FMs)
• useful video about bedrock
• Nvidia collaboration, ultra cluster
• Amazon Q, to connect AI to YOUR business data
• -Data strategy: A set of tools and frameworks to simplify data management and use for generative AI
• ZERO- ETL: by eliminating the need to build ETL data pipelines
• code Whisperer
• ​
• AI keynote
• amplifying human creativity through multimodal LLMs, Bedrock, Amazon Q, etc.
• CTO keynote
• lab: using langChain and AWS Bedrock
• LangChain is a platform that aids developers in building context-aware, reasoning applications using large language models (LLMs), offering tools for proof-of-concept to production
• Amazon Bedrock is a fully managed service that enables the building and scaling of generative AI applications with foundation models from leading AI companies

SC23 1/3

Deep Learning Tutorial provided by the Ohio State University

Deep Learning at scale Tutorial provided by NERSC and Nvidia [1][2][3][4][5][6]

Digital Twin Workshop:

• A paper, presented by a company originating from the Barcelona Supercomputer Center, includes simulating the heart and vascular system through virtual humans, employing a unique blend of multiphysics and multiscale models and AI
• A paper that addresses uncertainty in biomedical modeling within the digital twin framework including AI AND models rooted in physics : it discusses in silico clinical trials, revolutionizing drug development and paving the way for personalized medicine based on individual responses:
• Nuclear fusion using tokamak
• Weather Modeling

SC23 2/3

• AI application for Vascular System: general, detailed
• Gordon Bell presentation on ab initio with quantum accuracy, and explanations
• Gordon Bell presentation on a deep equivariant model for biomolecular simulations
• Computer Continuum for Science
• Oracle Cloud
• Penguin - infrastructure service provider for AI and HPC
• Parallel Programming: history perspective

SC23 3/3

• One thing I learned at SC23 is about Visual Transformers, or ViT. These are similar to the NLP transformers that are used in BERT, GPT, etc but applied to high resolution images. They work on the same principle: self attention. The difference is that instead of being applied to words they are applied to patches that form an image. After vectorization, those patches can be processed similarly to words or tokens.
• The deep learning tutorial provided by NERSC included an application of ViT for weather modeling which was a simplified version of FourCastNet
• The model is trained on era5 data which includes decades of experiments and calculations results. Once trained, the model can predict the output variables at time= t+1 given the input at time=t
• This approach is gaining traction because it supports finer grid modeling, and hence greater accuracy while delivering a solution in much shorter time than numerical codes.
• The point I am making here is that ViTs are not limited to WMF, so any application that has high resolution images can benefit from a huge performance boost.
• For example a holy grail in engineering is about leveraging all data collected in the past to train a suitable model. ViT could be one such model.

Altair presentation at Ferrari, Maranello - Italy

slides

​

Other presentations:

• optimize the design axial flux electric motor: digital twin, multi-physics, AI surrogate models
• iconsulting slides on anomaly detection

Google IO keynote - Mountain View, CA — May 2023

• Generative AI in Google products
• video
• Slides: part 1 till 53 min
• Gmail help me write
• Google Maps
• PaLM 2 and fine tuning, Med PaLM 2, Gemini
• Bard, programming, tools, integration with Google sheets
• Google Labs and search
• Slides: part 2
• Google Cloud, Vertex to develop Generative AI applications
• Responsible AI, tools to address inaccuracy, copyrights, etc

AWS SageMaker - June 2023

• Slides and Explanations : about using Huggingface on SageMaker and using SageMaker Jumpstart for pre-trained ML models
• Intuition about feature store and pipeline : my research to make sense of things. I used LLMs to compile the information
• Workshop done in June 2023 and Explanations
• This requires an AWS account
• The memory size has to be set to ~ 3000 , or else it fails:
• ValueError: {'Message': "'MemorySize' value failed to satisfy constraint: Member must have value less than or equal to 3008", 'Code': 'ValidationException'}
• There are 2 events in 2022 that also cover Sagemaker, and one in 2021 - see next slides
• The AWS re-invent conference of 2022 also covers Lambda

International Supercomputing May 2023

Summary Presentation

​

Linkedin Blog : final report

​

Linkedin Blog related to Dan Reed’s Keynote and to Argonne work on AI and HPC for health science

​

AWS : HPC and AI for the food challenge

GTC March 2023

Here is my summary with links to all tracks I focused on, slides capture and more in-depth analysis

​

Highlights:

keynote

PINN

Monai

Efficient Inference of Large Language Models

Deploying Huggingface models to Production with GPUs

Keynote

​

​

AWS Sagemaker tutorials

https://aws.amazon.com/it/getting-started/hands-on/machine-learning-tutorial-mlops-automate-ml-workflows/?nc1=h_ls

AWS re-invent - November 2022 - slides

​

• CTO

​

• Machine Learning

​

• de santis presentation

​

• data science and data management

​

• AWS storage

NLP Summit 2022

Here are my notes about the conference provided by John Snow Lab:

​

• widespread usage of NLP and transformers and BERT-based models across verticals
• I would like to call out a paper showing how to augment BERT with new vocabulary, one on antigens for cancer research, and one by booking.com again using BERT to optimize the travel experience based on customers’ interactions
• The next slides elaborate on the topics I focused on. Full conference coverage is here.
• Verticals covered in this slides deck: healthcare, retail, finance, travel
• Remarkable Nvidia contribution to accelerate AI

GTC September 2022

Besides the content explained in my notes, I would like to also call out:

• a linkedin blog publishing the notes
• A panel discussion on accelerating Health-sciences
• Jensen Huang’s keynote
• How to build a Digital Twin

​

In addition, here is an experiment to start using the Nvidia SDKs and to take advantage of the low/no-code options.

vbtransform - July 2022

Transform workshop and datascience

​

​

second week at #vbtransform : datascience, artificialintelligence and financialservices

aws conference on 6-29-2022

SageMaker demo on using ML for credit

​

AI services: Kendra for search , comprehend for smart doc, quicksight for BI

​

summary

ISC2022 - Hamburg

Keynote: Nvidia and BMW Digital Twin

​

My notes with links to presentations, abstracts, sound tracks: only covering the topics I am interested in

Global Azure Torino 2022

conference announcement in linkedin

​

​

my notes

GTC 2022 - Predictive Maintenance Workshop -more details here -

Part 1: Predict the exact failure of disk drives using XGBoost

Part 2: Model the time series using LSTM in Keras to predict a failure at a point in time before the failure

Part 3: Autoencoder for anomaly detection

​

• Differences compared to my work are explained in this blog: XGBoost seems better than random forest, but LSTM is more appropriate for time series problems.

​

• Anomaly detection was discussed here, in context of a Siemens presentation at SPS 2020

GTC 2022 - Digital Twins and Nvidia Omniverse

​

• compare Omniverse and Metaverse: many more useful applications in Omniverse: FourCastNet for climate modeling, Siemens, BMW, Pepsi, Amazon, Omniverse in cloud
• Some of the highlights are presented in the keynote.
• Modeling granular materials motion in industrial devices

GTC 2022 - Large Language Models and Megatron

​

• The pervasive nature and implications of large language models are summarized in this blog
• A French and Spanish language model based on Megatron

GTC 2022 - my notes

​

• my notes on the presentations I studied during the conference, including a link to the voice recordings
• ​

HPC conferences in March 2022 - 1 of 2

• Swiss HPC Conference: lots of information on security, multi tenancy in supercomputers, hpc storage security and an interesting software tool from Gent University for software installation automation. Useful for cloud providers
• Giuseppe Lo Re describes the storage stack in Alps which is the next supercomputer in Lugano
• Alvarez presents the Juelich architecture evolution including booster, dragonfly interconnect and more. They also compare Juelich procurement methods vs. standard
• James Coomer of DDN presents storage options for AI
• Gregory Kurtzer presents rocky linux, cloud native enterprise linux, warewulf, apptainer, hpc2.0.

​

​

HPC conferences in March 2022 - 2 of 2

• HPC Forum organized by Dirk Pieper:

slides presented by VW on container technology, rescale on simplifying HPC cloud usage, GNS Systems on Digital Engineering Platform, Atos/Science and Computing on a cloud HPC application for batteries design, and Airbus on how computing architectures evolve for HPC and AI and also on Quantum Computing

​

​

software development conference in April 2022

fundamental knowledge for web applications and cloud native applications development. See here a summary blog

GTC 2021

Here are my notes and slides collection on:

​

workshop: Self-Supervision, BERT, and Beyond, keynote, conversational AI based interactions in the Metaverse, AstraZeneca applications of transformers for chemical applications,

Nvidia Inception program, Siemens Omniverse and Digital Twins,, Physics Informed Neural Networks, Meet the experts @ GTC, AeroBert: using NLP at RollsRoyce, Graph Neural Networks at PayPal, Nvidia improves data science experience

​

SC2021

Here are my notes and slides collection on the Cancer Workshop

Smart Data Innovation Lab

microprojects : collaboration with DFKI , KIT, Fraunhofer, Software AG, IBM, SAP - focus on practical applications of AI and operation data

​

Industrie 4.0 SDCS-BW EDI uses OCR for technical drawings in PDF format

​

Price intelligence product information management, classification

​

Condition monitoring : Fraunhofer IOSB: data aggregation, predictive maintenance, anomaly detection

TECO – Technology for Pervasive Computing

Scale IT : Industrie 4.0 applied to Mechatronic systems, co-design among leading sensors providers like SICK and Carl Zeiss, use cases in manufacturing and logistics

​

Smart Data Solutions: for use cases in SMEs including MES, BI

​

Digital Mountains: projects in 3D printing, AI, VR and more

​

Heliopas AI: smart data solutions for agriculture

Major ML trends: natural languages & platform approach

Natural Language Processing -

beyond recurrent neural networks

• Transformers, BERT, GPT-3, Beijing Academy of AI ‘Wu Dao 1.0’
• BERT tutorial
• NEW: my Coursera NLP training
• CTRL novel writer assistant
• IBM WATSON Debater
• Omni-Channel Relationship Management
• Voice, voice, voice

​

​

​

MLOps

highly integrated platforms, automatic tools wherever possible, CI/CD

​

• Azure : prd ML is NOT ML
• AWS SageMaker
• bigml.com
• SAS Viya & M$FT Azure

​

SuperComputing 2020

Here is a detailed slides-deck including my notes

AWS: SageMaker, Predictive Maintenance and Lambda Services

Great portfolio of cloud services for AI experts and less experts, focusing on CD/CI

​

There is also a predictive maintenance solution that I need to compare with the Azure-based solution (WIP)

​

Another Predictive Maintenance solution by H2O

​

AWS Lambda Services

AWS Summit in June 2021

my linkedin blog on the conference

​

Q&A with Suse

SAS Viya: Cloud Computing & Analytics

Agenda: Cloud, ML, analytics: making it easier for customers to implement

​

Ms. Nevala’s and Ms. Shaw’s event I recommend Bernard Marr’s keynote, the detailed webcast on SAS/Azure engineering collaboration and the SAS Viya demo.

​

Deep Integration of SAS Analytics and Microsoft Azure: Azure cloud native applications are integrated with Sas Viya, thereby enhancing analytics. Cloud customers will have easier access to the combined technologies.

SAS Global Forum 2021

power of analytics

make analytics easy to consume by business users

manage analytics in the cloud, SAS Viya and MSFT Azure

cool presentation & demo on reinforcement learning

great story at Bupa

conference proceedings

SAS Be Curious - November 2021

A good demo showing the integrated ML capabilities of SAS Viya similar to those demonstrated by Ms Shaw :

• Scenario: Loan Approval Risk Assessment
• scikit-learn-like Viya modules or open source models can be used side-by-side
• pipeline generation and automation
• registry, governance
• decision flow
• low code, user does not need to be a programmer

Accenture and AI

Several recent AI companies acquisitions including:

​

Analytics8 : they also have an interesting Predictive Maintenance model

Byte Prophecy for scaling AI; they operate in Asia Pacific region

Pragsis Bidoop from Madrid with EU-wide reach

​

Several acquisitions across the board, potentially there is synergy with AI

MLOps --- more information

• CLOUD NATIVE AI WITH KUBEFLOW

​

​

AIOps = Artificial Intelligence for IT operations

YASH Technologies and ScienceLogic

​

GNS Systems ? Christopher Woll, CEO

​

Ubercloud study, optimizing jobs placement using fuzzy logic

collaboration between GNS and ubercloud

​

cloudwise and cloudwise presentation at the hdc.cloud conference

​

The growing popularity of no-code

Azure Cognitive Services NEW see this link provided during the MSR conference in October 2021, and this link

​

AWS Lookout for Vision

​

Quote from Andrew Ng:

It used to take me months to deploy a model. With a no-code platform, I can train a RetinaNet demo, carry out error analysis, use a data-centric approach to clean up inconsistent data, retrain, and deploy to an edge device — all in 60 minutes. I get a thrill every time I go through the machine learning project lifecycle so quickly.

Cloudwise Algorithm Technical Architecture

courtesy of Huawei

ML for MBA students: a webcast on 2-17-2021 & related slack

AutoML

​

BigML event in February 2021: my questions

​

ML is becoming more ‘buy’ than ‘build’

​

AutoML examples

ML for marketing communications (slide under construction)

Umania

The AI data and team-building challenge

• Today’s AI systems need to be trained on large datasets, but entrepreneurs, whose company hasn’t yet been built, and who don’t have data, can’t create an AI product as easily ---> chicken-and-egg
• Complex projects require different skills: here differences between data engineers and data scientists are explained
• Microsoft has developed a comprehensive approach on Azure
• Production Machine Learning is not the same as Machine Learning: Algorithmia explains why
• And finally, a data strategy/ ML strategy is required. This is my summary on organizations and management perspective

AI for graph

shortest path algorithms, with implementations in Python

Angela Bonifati Query driven graph analysis

​

Why I’m lukewarm on graph neural networks

​

Amazon Neptune

​

​

data architecture and data governance

• transform workshop - July 2022
• move data to the cloud
• snowflake
• ( or databricks ? see MSFT conference in Las Vegas 2018 )

​

• Roche uses snowflake and datamesh

Data Lineage

Infogix Unveils Industry’s First and Only Automated Data Lineage Capability that Delivers Business Knowledge Around Enterprise Data Assets

​

Octopai.com: Think of data lineage like your family tree. BI users need to trace the data’s roots like you

would your ancestry, mapping out the data, its context over the course of its life-cycle - where

it came from, what its relationships are, and how they all connect.

​

Automated data discovery: With automated data discovery you can instantly locate metadata from across multiple systems, and deliver faster to the business.

MLPERF

Industry standard benchmark for Machine Learning

Mlperf HPC

​

Some preliminary work is required to become familiar with docker

Because I expect GPUs to be the premiere platform, I am using NGC and start with this use case on an Ubuntu system.

Here is a status report in March 2021

AI for photonics

my post about the conference

Narrow vs. general AI

Interview with DELL

Samsung

AI -powered memory

Huawei

HDC.Cloud and AI Pangu

ML Resources

and more on Linux, Windows

Diffusion models

Latent Diffusion Models: The Architecture behind Stable Diffusion

I have posted a question on the data input on 9-5-2022

From Siri to Photoshop to Google Search — Large AI Models Will Redefine How We Live

​

Demos

https://playground.tensorflow.org

​

a data science guide and cheat sheets for a quick reference

A data science guide, easy to read; it explains skills sets and provides sample codes

NN cheat sheet

Scikit-learn cheat sheet

Python cheat-sheet

Pandas cheat sheet

collection of videos on Machine Learning by the University of Erlangen

transformers from scratch

bigger NN is better, explanations on GPT3

useful links for data-science

• Python course
• features engineering
• statistics
• confusion matrix
• machine learning course
• deep learning
• NLP
• feature engineering how-to guide
• O’reilly book
• PCA
• thermo-dynamics

​

and more

How to in linux and windows

a collections of notes on useful tools and usage

Variational AutoEncoders

UNDER CONSTRUCTION

​

VAEs in Economics

​

My own research – added in July 2023

​

​

GPU-accelerated data science

The time to publication of Data-Driven Research can be decreased through GPU-Accelerated Data Science. Large proportion of tools used for scientific analysis and machine learning (e.g. Pandas, SciKit Learn, NetworkX or Spark) have GPU-accelerated variants. We'll discuss how, by changing just a few lines of code, research teams can accelerate their experiments by several orders of magnitude, significantly reducing hardware-associated cost or data processing/training times.

​

​

We will show a demo on how to dramatically accelerate Histopathological image pre-processing (tiling and thresholding) pipeline using RAPIDS. In addition we will show how to take advantage of GPU-Accelerated graph analytics tools for the purpose of understanding the role of certain cell-types in the tumor microenvironment.

We will discuss the common problems that hinder Data-Driven Research.

​

The Nvidia webcast link is W.I.P -- June 15, 2021

Quantum Computing

Quantum Computing resources

Atos - ISC2021

Microsoft - ISC 2021 potential and limitation, only few use cases have exp speedup / more information from the same source

Microsoft presentation: Basic infos on QC ; Deutsch Oracle: video, slides

Jülich and VW - ISC 2022 A primer on superposition and entanglement: VW use cases

Quantinuum , TKET, lambeq (Honeywell) , InQuanto, chemistry, Python API, above a given number of atoms standard computers cannot do

Coveney 2022 ISC 2022 quantum chemistry, wave function, Schrödinger

Purdue, Kais - Restricted Boltzman

more details ISC 2022

cuQuantum workshop at GTC September 2022

Quantum Computing resources - continued

​

QML course

​

Schrödinger Equation explained

Quantum Computing - projects

D-wave github repos and discussion with an expert

D-wave in AWS marketplace, including PoC - october 2022 announcement

Atos and Atos Quantum simulator

Cambridge Quantum Releases Quantum NLP Toolkit and Library :

CQ said [the toolkit] lambeq converts sentences into a quantum circuit and is designed to accelerate the development of practical, real-world QNLP applications

Bosch

Rolls Royce and Classiq : QC for CFD using the HHL algorithm. Classiq make it easier to develop Quantum circuits,which is a big deal given the complexities even in a basic workshop

error correction: Natalie Brown, Quantinuum, said that classical error correction principles fail with quantum computers because of the basic nature of quantum mechanics.

• LARGE NUMBER OF QUBITS
• Impact on HPC - Scientific Computing 2021 webcast
• Universal Quantum - Prof. Winfried Hensinger
• Pistoia Alliance : promoting knowledge sharing
• Quantinuum
• Quantinuum Closes in on Break-even Point in Quantum Error Correction
• it marks the first time that logical qubits have been shown to outperform physical qubits — a critical step towards fault-tolerant quantum computers.

​

• A review of QC providers in the Scientific Computing magazine, 2022:

​

A review of QC providers in the Scientific Computing magazine, 2022

• Universal Quantum: error rates, coherence time, number of qubits
• Quantum Brilliance systems can be smaller and more easily integrated with existing computing
• Quantinuum is the combination of Cambridge Quantum with Honeywell:
• Quantum Solutions,” : integrated solution.
• “A lot of the algorithmic work is in imagining this future where you don’t
• have to worry about qubits and how they interact, because all of that has been ‘taken care of’ by universal fault tolerance.”

​

​

quantinuum staff

quantinuum software

Exascale

Exascale applications

Porting a Particle-in-Cell Code to Exascale Architectures

​

Industrie 4.0

NEW : Hannover Messe 2024

1. linkedin summary
2. check the detailed posts in the comments to the post at #1
3. The next slide is a sheet with the details per topic

Hannover Messe 2023

Digital Twin , Asset Automation Shell, OPC UA and Digital Product Passport

​

AI and Generative AI

​

5g, Metaverse/Omniverse for Manufacturing, and other topics

​

​

​

Hannover Messe 2023 - Industrie40, Digital Twin , Asset Administration Shell, OPC UA and Digital Product Passport (DPP)

• my first day at the Messe
• AAS and Digital Twin Tour [slides, blog]
• Bosch and participation in IDTA
• Digital Data Chain and Asset Administration Shell in context of the DPP , blog
• Digital Twin for Industrial Sites
• DPP panel
• OPC UA - Herden , Hoppe
• Software BOM

Hannover Messe 2023 - AI, Industrial AI, and Generative AI

• AI in Industry including Siemens, Monolith and more [slides, blog]
• How does AI get into Manufacturing - Tour:
• Beckhoff --- Fabian Bause
• AWS, Pepperl & Fuchs, Syntax : IIOT to Cloud
• PSI - Rudolf Felix : causality and explainability of AI ; sequencing and scheduling in the automotive industry.
• Weidmüller: mechanical and plant engineers can drive forward the development of analysis models by themselves without having to be trained as data scientists
• AI-powered Industrial Transformation, panel: automated and integrated quality control in prd process and assembly [ref]
• Generative AI and LLM: Aleph Alpha for Maintenance Applications with explainability

Hannover Messe 2023 - Metaverse/Omniverse for Manufacturing, 5g, and other topics

• Accelerating Time-to-Market: Harnessing the Industrial Metaverse with Robotics Simulation and XR - Nvidia Omniverse [slides]
• 5g / Fraunnhofer
• Hydrogen

Hannover Messe 2022

here is a preview and summary of past editions

Hannover Messe 2021

day 1

day 2

day 3

day 4

day 5

​

My notes

DFKI and Project SPAICER

• Production Anomaly Detection using Autoencoders and Reinforcement Learning

Digital Twin

• 2022 - siemens and Nvidia in Omniverse
• Siemens, Nvidia and NetAllied Digital Twin in Nvidia Omniverse:
• Digital Twins at the factory level and the machine level
• KPI : cost, safety, energy
• a DT for shop floor interacts with ERP for: supplies availability, machine maintenance required, ect
• from macro-scale to machine level, control on each machine, PLC

​

• BMW factory digital twin
• JP’s blog on Digital Twins

​

​

Digital twin NAFEMS - December 2021

Un Jumeau Numérique Pour Mieux Connecter La Simulation Multiphysiques Et Le Monde Réel

​

Hexagon

​

Cosmo

​

AI

Siemens

collaboration with Google cloud for AI applications in manufacturing

Bosch and AI

my post following the 2021 CEO’s keynote

​

competence center for AI

Predictive Maintenance

• GE Predix: One of the first applications, combining assets, data, analytics, and applications to transform industrial operations
• Hitachi: a pipeline of NLP applications based on Spark NLP to process field reports on maintenance actions, extract problem statement and classify into incident types, such as bearing replacement, etc.
• The heavy lifting is in pre-trained (BERT) language models fine tuned for Hitachi. The classification uses a deep NN
• Microsoft, JP, UberCloud: domain-specific features to predict time windows failures of a machine consisting of 4 components, and for which IoT data and maintenance records are available

Initial research on ML for manufacturing in 2018

slides on GE Predix, 3d printing, Digital Twins for FEM and a McKinsey cross industry report on ML applications

Hannover Messe 2019

Here is my summary

SPS Parma 2020

Here are my notes

Guidepoint HPC Cloud Q&A

Opportunity uncovered by Olivier Schreiber on July 16, 2021

team inputs, deadline, participation in the interview

• Pass 0
• ​
• Pass 1 - see next slide
• ​
• Deadline?
• ​
• Can we show up as a team?

Pass 1 (1 of 3)

Q1

Given public clouds have historically struggled in HPC, how are they closing the gaps? Are there still any major tech or strategy gaps as barriers for large scale HPC migration towards cloud?

response to Q1

The gaps are to some extent dealt with. Specifically:

* virtualization inhibits less performance and scalability for many use cases. In addition, cloud providers support containers and bare metal options.

* bring-your-own-license is a good model for some ISVs, but some users may have complexity issues for their hybrid cloud models.

* high performance interconnects are supported at least in Azure and AWS

* data ingest and egress are less prohibitive

* there are real HPC applications that scale, for example Schlumberger on Azure Cloud

* security and governance are enforced, but some users may have concerns

Some important caveats and nuances about above points and costs will be

discussed during interview as well as the main advantages for using the cloud

for hpc i.e.: improved agility/flexibility for burst capability, notably.

​

​

​

Pass 1 - 2 of 3

Q2

Please briefly describe the strategy intent of smaller hybrid cloud / cloud-burst players like Rescale, Nimbix, Ubercloud. What are they helping solve for the customers?

response to Q2

Rescale covers the workflow for hybrid HPC cloud usage.

Ubercloud focuses on containers and Kubernetes. With this approach they help customers (mainly Fortune500) in simplifying the usage of (hybrid) cloud.

More details and some players in Germany/Europe, will be mentioned in interview

such as Simscale.

​

Q3

Across the technology stack for HPC from infra to application, who do the public clouds and smaller HPC-aaS players partner with? What do you think is the intent of these partnerships? Ignoring obvious partnerships with chip vendors.

response to Q3

Cloud data governance and certification are important factors.

There are a lot of different partnerships of public clouds and smaller HPS-aaS players, e.g. UberCloud and Microsoft Azure.

Other examples are the partnerships between GNS and AWS, MS and UberCloud.

The main intent from the CSP‘s point of view is to better serve their customer‘s

​

Pass 1 - 3 of 3

Q4

Do customers consider a colo in their HPC strategy? If so, in which situation are they most relevant?

response to Q4

Yes, e.g. BMW has used colocation (https://drive.google.com/file/d/1ml3c37_PFp_7n6FKFg2KQwsafdx1_B4J/view?usp=sharing) in Iceland since 2013, Daimler is at least considering colocation for their HPC strategy, they are currently using it for „general IT“. The benefits are energy savings and offloading the primary data center, while determining acceptable latencies depending on the hpc application.

April 2022 Guidepoint request: HPC market dynamics (#952364)

Q1 can you speak to the major players in HPC

Q2 can you speak to differentiators and value drivers and trends in HPC

Q3 Can you speak to HPC across major use cases for HPC, mission critical, analytics for enterprises, AI

Q4 What sources do you use for market insights

Q5 at a high level, what trends do you observe in HPC

Here is our response

​

Backup slides

old school ways to advertise HPCsquAIre

• presence at conferences
• ISC
• NAFEMS
• …
• partnering
• NI-SP
• UberCloud
• Sicos?
• …
• publications (print)
• Scientific Computing World
• Digital Engineering (German)
• ...

other websites

• Digital Engineering - https://www.digital-engineering-magazin.de
• Digital Engineering 24/7 - https://digitalengineering247.com
• Revolution in Simulation - https://revolutioninsimulation.org
• CAE Forum - https://cae-forum.com
• CFD FEA Service - https://cfdfeaservice.it
• Scientific Computing World - https://www.scientific-computing.com
• Science Node - https://sciencenode.org
• ...

Java sabbatical, iSCSI PoC and Parallel Computing tutorials in the years gone by

• J2EE PoC on linux notebooks connected with Ethernet (client/server) - 2009
• iSCSI setup on clustered PCs running RHEL -2015
• In 2003, I developed parallel programming tutorials for university customers:
• MPI
• OpenMP
• case study on MPI
• case study on OpenMP
• Parallel Program debugging using TotalView - courtesy of TotalView
• Parallel Program profiling using Vampire - courtesy of Pallas

​

​

Teknowledge: my first touch with AI in the 80’s

linkedin blog

​

​

introduction

​

​

advanced course

​

the HPC view by a professional HP benchmarker in the past: Logan Sankaran

Here are the slides, mainly based on HP-UX but they may still be useful to elaborate on the benchmarking methodology

meetups & communities

ACTIVE participation, the organizers solicit inputs, e.g.

AI meetup Italia

AI meetup Germany

​

AI community in Berlin

CAE forum shown in Linkedin

collaboration tools

• Key point: google docs tools are EFFECTIVE when working on shared projects, for example this document can be edited/commented. In addition:

​

• slack, for instant messaging supporting multiple channels is perfect for a complex project with multiple work packages
• I am a Slack user on a Machine Learning thread

​

• airtable, to summarize and publish contents

​

• devpost, the home for hackathons.

Kubernetes architect

• daniel.gruber@theubercloud.com

​

I-O system tuning on GPU clusters, Gordon Bell 2018 paper co-author for ML work on Summit (> 1 exaflop)

Archived slides on website development

Projects and HCSQUAIRE Website & business development

ML for Automatic Speech Recognition (ASR) and Machine Translation

1. The end user platform is TV - side Raspberry PI implementing a model to do english - spanish translation on the fly. The critical part is model training on a large VCTK corpus.
2. I started with ‘Listen_Attend_Spell’ and this PyTorch model on an Ubuntu system equipped with an Nvidia GPU.
3. The project at point #2 has been placed on hold in 2020 due to technical issues and lack of support.
4. However, in July 2021 I found a possible new approach using Nvidia NeMo. This report on ASR explains the details and preliminary results, and I also published an early report on linkedin: The model is relatively small which bodes well for simple edge devices deployment. Please refer to the repo

CFD model of several buildings for energy optimization

Mural brain-storming

website and linkedin group

1. hpcsquaire website
2. linkedin group
3. current developments to increase traction, June 2021 time frame. The linked document contains team members’ profiles
4. content base for a video on Machine Learning
5. announcement of Software Velocity as a partner of hpcsquaire

​

(the initial development details, meetings reports, bug fixing, etc are still available here. In addition, a back-up of old website pages is available here. )

classic HPC

Even though the bulk of this presentation is about Machine Learning, HPC classic is still the main goal of the hpcsquaire team.

Hence ideas and developments could be linked to this document or added as HPC slides such as:

• High performance storage, scalable NAS systems
• xyratex, Seagate, now belonging to Cray

hpcsquaire partners

NI SP for visualization

Software Velocity to support digital transformation

project proposal on July 6 2021 to develop a MVP

• HS to approach CloudFluid for potential collaboration with hpcsquaire, ubercloud. Expanding CFD code with AI
• a similar case was shown at ISC, PyTorch based LBM
• JP to work out a project proposal
• funding from EU ‘FF4EuroHPC’ EUR100k available

collaboration with MSFT

This is my presentation to Microsoft on July 2, 2021

KHP to join an hpcsquaire meeting to discuss:

• Daimler project
• other forms of collaboration with us
• JP to send a reminder

gonarrative

champion of business story-telling

can help us in this: compile inputs from each of us to deliver the hpcsquaire story

friday July 9 2021 call with gonarrative: we agreed to start working as a team in August 2021

​

the compass chart on hold

revolut (for $ transactions)

use HS’ account

details to NM

they have IBAN

Tracking HPC**2 agenda and business development

Website development plan

​

Agenda for the 2-9-2021 team meeting:

• SICOS - NI SP - HPC**2 synergies >>> Henry, Karsten 10m
• personal websites dev. status report >>> All 20m
• eye-catching HPC content: brainstorming >>> All 30m
• plan the agenda for the 2-12-2021 meeting 5m

2-19-2021 Create one page for HPC and one for AI with proof points TBD

Logan Sankaran’s (jpg) slides: these are mainly PA RISC /HP UX based, and they illustrate HPC capabilities for porting, tuning, etc.

I used the google slides API to programmatically insert all of those jpg files into THIS PRESENTATION using this Python program

(earlier attempts are described here/ and ended up with this successful insertion test ; however this programmatic approach mandates that all urls MUST be publically available, so that I could not just process the images uploaded to the google drive)

​

meeting on 3-12-2021

1. delete ‘how we store the data ‘ in privacy policy
1. work together with ubercloud and a video expert (TBD) in order to create a demo using ML for CFD, possibly showing the Renumics use case. JP to send mail to Wolfgang.
2. do the website as simple as possible, using block editor, no fancy stuff
3. determine target date to go live: April?
4. ask linkedin help desk how to create a group page rather than a company page, e.g. supercomputing group (JP)
• when done we need to invite 1st levels
• new case id

​

meeting on 3-16-2021

1. review open questions and website changes - here starting from 'JPs review ...'.
1. mission statement TBD; all team members should think about: HPC and AI
2. consolidate into 1 page who we are / what you should know and provide a summary sentence for the bullet point
3. 2 pages on offering ---> merge the 2 into 1
4. skills set: update as advised
5. our menu: eda, molecular dyn, life sciences, FSI, quantum out: get ME and kent koeninger on board (action HS); an extended team can warrant a longer list of deliverables
2. Wolfgang's proposal for a joint content hpc**2 - UberCloud -- email on 3-15-2021
3. Karsten‘s feedback on SICOS contact

about us /who we are paragraph

headcount and hpc specialist ---> keep

training on parallel processing/programming → keep

BD ---> keep

linkedin group

Please check the following link to know how to create a Group on LinkedIn: Create a LinkedIn Group . Unfortunately, we don't have a functionality of inviting your direct connections join your group automatically. You will need to invite your connections join your group: Invite People to Join a Group (Group Management) .

Q: how to automatically invite connections, lot of work if done manually

We don't have that specific functionality available, but I've sent your suggestion to our product team for consideration. When many of our members ask for the same improvement, we try our best to get it done. However, due to the large number of suggestions we receive, we can't provide a timeline.

invitation to linkedin group

Good Morning,

​

I would like to invite you to join the HPCSQUAIRE linkedin group (insert link here): an initiative to provide consulting services for AI and HPC, benchmarking, architecture design, etc.

​

You can view more information about our business & team at www.hpcsquaire.org.

​

[to be signed by the person who sends the invitation]

​

Thank you  

​

SICOS - NI SP - HPC**2 synergies [ part 1]

business development OPPORTUNITY

​

​

text for website to be provided by Karsten and Henry to describe the business opportunity of Sicos and HLRS

​

sicos looking for commercial customers that would like to use HLRS resources

SICOS - NI SP - HPC**2 synergies [ part 2]

GENERIC OPPORTUNITIES

​

• target SME
• sicos acts as a broker of HPC services & training to users
• what channels ? To be verified w/ sicos sales
• Are we pulled in by them or we collaborate in the project acquisition and qualification?
• who is doing the hands-on work?
• NI SP has customers in the SME and large enterprise space; joint marketing actions
• sicos acts as a mediator of HLRS and KIT compute resources for simulations, and a mediator for training services
• SICOS BW : SDSC-BW: smart data solution center (analytics, AI)

youtube video for the website

• Powerful tool to share ideas, story-telling, webcasts, etc.
• Focus on CONTENT e.g for a 10 or 5 minutes video
• Intro: benefits for customers 1 min
• HPC by examples
• AI: a couple of use cases
• work with https://www.vivalexis.com/en/ They should tell us what content they need from us and how much it costs
• some help from Sicos?

create a blog in wordpress; action JP

• add content incrementally
• short description besides the link to the blogs
• one blog for discussion forum
• one blog for news, e.g. introduce NI SP, UberCloud, etc
• Permanent updates, better SEO
• case #3837527 with WordPress support to clarify blog creation and how to save older pages

wordpress case #3837527

I need to make a blog out of this static page (https://hpcsquaire.org/what-we-are-offering/) How can I do it?

Here's a guide: https://wordpress.com/support/five-step-blog-setup/#step-2-write-your-first-post

I edited a page https://hpcsquaire.org/who-we-are and then tried to save it as https://hpcsquaire.org/who-we-are-new/ -- see attachment- but now the original page is gone.

What you edited on this page is called the permalink. The old page remains, just a new URL. To keep a copy of a page elsewhere, we recommend using the copy feature: https://wordpress.com/support/copy-a-post-or-page/

Alternatively, WordPress comes with revisions every time you save/update a page wherein you can rollback changes you made, you can read more on accessing this here: https://wordpress.com/support/editors/page-post-revisions/

​

blog post

https://wordpress.com/post/hpcsquaire.org/683

​

not visible from main page ?

meeting on 3-30-2021

1 move the description of hpc**2 as a footnote

2 remove the team gif

3 icons for linkedin to be shown on top of page next to ‘about us’

4 Die Adresse wurde nicht gefunden

Ihre Nachricht wurde nicht an join@hpcsquare.net zugestellt, weil die Adresse nicht gefunden wurde oder keine E-Mails empfangen kann.

fwd to team members ‘ mail boxes

5 slide #10 action KG, items 1-4 action JP

6 jp to make a video proposal

4-6-2021 meeting

add visualization to skills

​

explain the intent of working with cadfem on predictive maintenance

4-16-2021 meeting

get rid of ‘private’ for ML and HPC pages >>> JP

rearrange in alphabet order >>> JP

email addresses >>> HS

olivier to provide a blog >>> OS

some issues when working with GitHub

this slide is about WIP, it will be removed once the issues are resolved

​

Ticket ID: 1017757

• failure to load a Jupyter Notebook
• something goes wrong when creating a new repo

​

​

(2020) invitations have been sent in linkedin [part 1]

anglano

bichlmeier

bruening

chiel

diogo

enda

graening

justyna

​

karthik

loffet

koenig

santini

schlachter

supalov

suwelack

thuerey

zoiner

(2020) invitations have been sent in linkedin [part 2]

• Renata Olejnik, IP Manager /Commercialisation specialist, University of Warsaw
• Valeriu Codreanu, Head of High-Performance Computing and Visualization at SURF – Netherlands
• Martina R. Dinkelbach, CEO Digital Transformation @ MAKOWA
• Rachael Hagan, PhD Student at Queen’s University Belfast
• Nicolò Magnanini, Pigro CEO (focusing on chatbots)
• Karlis Skuja, Community, Entrepreneurship and Innovation – Brussels
• Marjolein Deryck, PhD Researcher bij EAVISE – KU Leuven -Belgium

​

• Constantin Diez, CAE Machine Learning and Software Manager at LASSO Ingenieurgesellschaft mbH
• Wim Van Acker, O365 Platform Owner & Solution Architect (as external dlw consultant) at Ypto -Belgium
• Aleksandra Kubica, CEO, Founder at diCELLa – Krakow, Poland

​

Old slide - Biz opportunities (1/2)

• Linkedin premium
• Messages to anyone
• Training courses (tech stuff, business, marketing,…)
• Sample recording 1
• Sample recording 2
• Business insights, access to staff members profiles
• Ask me if you need anything useful for the team requiring premium subscription
• Smarttribe: a web tool to augment your network of professional peers
• https://open.smarttribe.io/welcome/aims
• Fremium model
• Chose whom you want to get in touch with
• GULP: info@gulp.de for job posting
• Chose your topics, you get notifications on available jobs, many are for contractors, mostly cloud-related

old slide - Biz opportunities (2/2)

• Expand the team, here is a sample [linkedin] invitation text:

​

Good Morning, Perhaps you remember we met [virtually] [f2f] at event ABC to discuss XYZ.

​

I would like to inform you about HPCSQUARE: an initiative to provide consulting services for AI and HPC, benchmarking, architecture design, etc.

​

You can view some preliminary information about our business & team at www.hpcsquaire.org.

Please keep in mind what you see there is work-in-progress and subject to change.�

We are experts in basic technology, and we are looking to expand our deliverables,  so that when doing project work, we could reach out to a virtual team including applications and industry experts like you.

As an initial step, I could invite you to give a short presentation in one of our weekly meetings; that could also help promote your business in an international context. �

Thank you  

OLD SLIDE on youtube video

Wolfgang's proposal for a joint content hpc**2 - UberCloud -- email on 3-15-2021

​

The HPC/ML video that I was proposing is not doable in the short term.

​

This action should not delay the completion of the website; at this time the priority is to go live

​

Natural Language Processing with Attention Models & Transformers

joepareti54@gmail.com, August 2021, updated in march 2022 january 2023

This paragraph was designed to cover a Coursera NLP workshop, then it evolved in multiple directions

NEW: April 2024

1. Using GPT4 to learn a basic implementation of a programmatic access to LLMs and resolve issues
1. Linkedin Article and troubleshooting report
2. Code : program jp-llama-langchain-2.ipynb and jp-llama-langchain-3.ipynb
2. Amplifying the exercise through Langchain, and Knowledge Base
• Linkedin Article and detailed report
• Code : program jp_llama_langchain_rag_0.ipynb
3. Amplifying through Vector Embeddings and FAISS
• Linkedin Post and detailed report
• Code : program jp_llama_langchain_rag_faiss_1.ipynb
4. Project - Q&A on user defined topics - restricted view

​

​

​

​

NEW: LLMs : open vs. closed source & optimizations - November 2023 - Updates in January & February 2024

• slides
• my document
• script

​

New slides deck, updated in January 2024 with focus on quantization and open source; modified script in english and in german

​

Deep Dive in quantization, some statements are not accurate, facts checking is required

​

​

​

​

Applying LLMs to a Linux troubleshooting case - overview

• Phase 1 Ubuntu System does not boot in GUI mode
• Time frame September-October 2023 using GPT3.5
• Scope: UEFI updates, ldconfig from another system with same OS, boot repair tools
• Results: system boots in CLI recovery mode, or using an external Linux, but the main issue remains
• Phase 2
• Time frame September-October 2023 using GPT3.5
• Scope: packages removal and reinstallation to fix conflicts
• Phase 3
• Time frame October-December 2023 using GPT3.5
• Scope: python scripts, dependency issues, focus on graphic system
• Results: packages seem OK, display manager or the Xorg system are the likely culprit
• Phase 4 (2024)
• use gpt4, focus on nvidia driver and kernel compatibility issues, better scripts
• Results: I2C Timeout Errors, script to run before gdm starts (this is further developed in the next phase)
• Phase 5 (2024)
• usage of gemini, claude, mistral and more, script to run before gdm starts, collect data on different ubuntu systems using the same script, leverage large context lengths and use chain-of-thoughts

NEW: LLMs : open vs. closed source & optimizations - Abstract November 2023 - Updates in January 2024

As advancements in Neural Networks continue, there's growing interest in local execution of Large Language Models (LLMs) instead of using public services hosted by remote companies. Running LLMs locally on one's own hardware brings a number of benefits. Firstly, local execution guarantees data privacy, ensuring sensitive information isn't inadvertently exposed to third-party API or cloud providers. This approach also offers users more control over their computational environment, allowing for specific customizations and optimizations that may not be feasible on shared cloud infrastructure.

​

In addition to these advantages, running LLMs on local hardware can offer consistent performance, free from performance fluctuations that could affect cloud-based operations. The direct hardware access can reduce latency, enhancing the real-time processing capabilities of LLMs. Moreover, without the recurring costs associated with cloud services, local deployment might prove cost-effective in the long run for enterprises with substantial computational resources.

​

Local deployments of LLMs including interesting features like quantization of models reducing the memory consumption significantly allow to run a huge number of custom LLMs dedicated to specific tasks like programming, summarizing, sentiment analysis, image analysis, ... including benefitting from LLMs with special knowledge in domains like law, financial markets, medicine, ... LLMs can run on standard hardware with or without GPUs and offer great performance in response times and quality of output.

​

The talk will introduce quantized LLMs and show different examples of running LLMs for use cases such as programming, question answering, summarizing and image analysis.

Generative AI and LLMs presentation - July 2023

• my document (playbook, updated in May, June, July 2023)
• slides
• script
• advanced use cases

​

​

chatgpt presentation - January 2023

• my document (playbook, updated in May, June 2023)
• slides & script
• initial attempt

​

​

​

NLP 2021 training - First hurdle: learn Google Trax

Learn how to use the available code ; unfortunately I spent quite some time to understand the details, see here

​

good background information here

​

linkedin blog

Next hurdle: make sense of the overall architecture

Backtrack a little

I was overwhelmed, and found that not even the Coursera lesson moved me forward. So I decided to do some basic research, and here are useful sites that help me connecting the dots:

• tokenizer: part 1, part 2, part 3: keras code snippets that explain basic things: you can also try them out on colab: exercise 1 2 3
• word embedding: reduces dimensionality while retaining the semantic word similarity in vectors, see here .
• word2vec : semantic similarity of words are reflected in the vectors
• As usual, Nvidia is the place to go
• and my first encounter with NLP back in 2018

… and backtrack a little more

1. seq2seq: this is an outstanding tutorial on a basic model
2. seq2seq: in this tutorial, they improve the performance using LSTM
3. seq2seq: encoder/decoder model to further improve items 1 and 2
4. teacher forcing: good intuition using some math
5. LSTM: a Google video that has played a fundamental role for my ML efforts
6. RNNs: great tutorial by LeCun out of an awesome series

This exercise is more complicated than it looked like at the beginning, so I am collecting the details in this document. The slides are not appropriate for the purpose.

NEW - more on LSTM 1 of 2

LSTM is used for the Predictive Maintenance workshop at GTC2022: it uses Keras, which is also used for the exercise presented before.

​

this video provides the reference explanations for the Keras input/output model and it complements other LSTM tutorials

NEW - more on LSTM 2 of 2

I have some questions on how to use the Keras LSTM code, and while seeking answers I found these useful resources:

• a report, and associated video on time series prediction that also contains interesting pandas functionalities for features engineering, and for transformations, such as dataframe to numpy matrices.
• some unclarity in the function create_dataset, that calculates the tensors input to LSTM are explained here

NLP Summit 2021

• Awesome conference on NLP technology and use cases in healthcare and financial services.
• It opened up a whole new perspective for me: on NLP through John Snow Lab and the wide availability of models one can use / adapt
• Spark cluster , Spark OCR, Spark NLP built on Apache Spark
• ( How to build a Spark NLP pipeline )
• Named Entity Recognition
• Annotations Lab
• Knowledge Graphs and Spark NLP, here is a good overview
• Use Cases: I am calling out the medical billing application which reinforces some beliefs on an opportunity in the US.

​

Spark NLP at the API World , and here are my notes, also covering the NLP Summit

• NLU for Python; my t5 code does not use it
• 4000 off-the-shelf models; it can also be used to train your model
• Learn using the demos
• Multilingual support
• Slack channel for support
• All functions supported in Spark NLP without having to export data ---> better scalability
• Open Source
• There are production projects based on Spark NLP
• Air gap
• One line of code to install
• Python, Java, Scala Support

​

Some more experiments

• My first attempt to text summarization which is also reported in a linkedin post
• the results using T5 seem strange, so I raised my concerns in the Spark NLP Slack Channel:
• i would like to ask a few questions on a text summarization experiment https://github.com/joepareti54/NLP_experiments When I use T5, the input text is summarized even further compared to another algorithm, but lots of the article insights are missed. I am neither familiar with NLP nor with T5, and this is really a starting point. I picked an article from Wall Street Journal for which I have done some homework in order to evaluate how NLP tools can accurately summarize the content. Any insight would be greatly appreciated
• ​
• why pre-processing to remove non essential characters up front

LDA and use case

• Some intuition on how LDA works: a latent layer, the topic, that helps reducing the number of threads connecting words and documents. Also, here are Serrano’s slides: part 1 and part 2
• An important LDA use case by Booz Allen and Hamilton: NER on EHR to detect adverse events: a target group and a control group are defined, then a vocabulary is filtered out from EHRs using ensemble supervised ML, then LDA is applied using that vocabulary. The result is detection of events that were not evident in the raw data
• Slides and video

linkedin blog

NLP & Transformers workshop at GTC 2021

• repeat almost the same DLI training as at GTC 2020
• details are here:
• text classification and NER using a pre-trained BERT model
• encoder-decoder BERT architecture, embeddings, LDA Part-1 and LDA Part-2
• Q&A
• certificate

Bi-directional Encoder Representations from Transformers, or BERT

• a great BERT introduction ,about encoding/decoding and a transformers overview
• more details in a Google presentation, (slides for quick reference) which explains embeddings like word2vec, and RNNs as a means for contextual embeddings, followed by an easy-to-follow introduction to the attention layer, and finally transfer learning to create industry specific BERT implementations.
• BERT embeddings A tour of basic functionalities that is used to build more complex projects
• BERT classifications using the Huggingface software
• A sentiment analysis use case based on BERT, code is provided

Christopher Manning - Stanford University

word embeddings

C. Manning: Word Embeddings and Co-occurrence Matrix (January 2022)

question on stackoverflow on how to build the matrix of column indexes

​

linkedin blog

​

github repo

​

​

Co-occurrence Matrix - test case for window=1

Co-occurrence Matrix - test case for window=2

Hugging Face

Hugging Face in 15 Minutes | Transformers, Pipeline, Tokenizer, Models

​

​

NLP Directions

Low code, no code

Active Learning

Generating speech from raw audio (Facebook Research)

BART

Projects

2026

• Implementing Retrieval-Augmented Generation on AWS: a PoC
• slides
• github
• video
• Agentic AI demo showing agent ability to make autonomous decisions
• slides
• github
• video

​

​

​

​

2025

• Implementing Retrieval-Augmented Generation with Llama 2 on Resource-Constrained Hardware: a PoC
• slides
• video
• Protein Structure Prediction for Drug Development: Focus on Alpha Fold 3
• slides
• video

​

​

​

​

2023

• Deep Learning for CAE

​

• July 2023 : VAE and pre-study on a naval customer which may be a conversational AI case
• Brainstorming AI for health science :
• Genomics
• Omics

​

​

PyTorch benchmarks on several hardware platforms

Requestor: Dr Karsten Gaier - April 2023

​

The purpose is to provide guidance to AWS users of DL training and inference applications.

​

Testing VMs on prem and in the AWS with different hardware configurations

​

This report has the detailed results and links to implementation notes, sources, scripts and system details. A milestone report is here.

using a no-code approach for NLP

A key takeaway from GTC September 2022 is no-code to fast track deployment of AI applications

​

This case is about using a pretrained bioBERT model and fine tune it for NER

Predictive Maintenance

Here is the latest report :

​

​

In this report, I am describing a project I worked on in 2019 together with Microsoft and UberCloud, and I am reviewing some AI-based Predictive Maintenance solutions available in the market. My project was a learning exercise, yet it turns out to be a rather comprehensive model that takes into account all important variables that influence the Residual Useful Life of an asset.

ASR using Nvidia NeMo

Even though this project was not pursued, it provided an interesting entry point in the technology, and it fits nicely with the Nvidia Conversational AI workshop at GTC 2022

Identify Customer Segments

​

where is the newest report?

​

This is a project to identify customer segments to increase sales. It is based on a demographic file containing ~900K records, and a ~200K customers’ records file. This is an application of unsupervised learning techniques such as k-means clustering and of Principal Components Analysis.