"No one is harder on a talented person than the person themselves" - Linda Wilkinson ; "Trust your guts and don't follow the herd" ; "Validate direction not destination" ;

March 31, 2021

Applications of Machine Learning in the Supply Chain

Key Notes

  • AI is the new electricity
  • ML is at hype cycle


Realtime AI

  • Dynamic routing - dynamically learn best route
  • Worker / Picker assessment
  • Optimal pricing through real time market feedback
  • Forecasting - Prophet


Dynamic Routing


  • The traditional approach - Regression
  • Online learning-based approach
  • Exploration vs Exploitation trade-off
  • Routing under uncertainty
  • Minimize cost on average
  • Optimization problem
  • Dynamic routing
  • Identify new routes based on the current status








AI, Machine Learning & Supply Chain // Manuel Davy, CEO of Vekia

Key Notes
  • Global, Agile
  • Collect information of Stores, Products, Location
  • Demand forecasting - at different timescales










Using Graph + Machine Learning to Optimize Logistics in Supply Chain












Keep Thinking!!!









March 26, 2021

Weekend Reads - Roadmap for Transportation Savings

Keynotes
  • Manage holistically, Transportation across Enterprise holistically
  • Inbound, Outbound, Facility
  • Mapping it to career capacity with advanced optimization technologies
  • Getting the best transportation solution, Optimize preferred integrated DC network
  • Map capacity to Freight flow, Run few empty miles
  • Turn data into information, Timely fashion
  • Build baseline as is cost, Restructure, reoptimize what does cost look like
  • Overall complete management portfolio
  • Getting visibility is vital

Happy Learning!!!

March 25, 2021

Why I may never work for a startup

I was thinking about what motivates me

  1. Technology does not motivate me, The same CRUD(Create / Read / Update / Delete) can be done in million ways and dozen tools. All I care about is some tech that would fit my tech need.
  2. My mind keeps wandering on multiple ideas. I don't pick one idea and sit on it forever. Sometimes you read papers, you try some blog code, examples. It is a mix, collect lot of ideas and let your creativity and interest pick what you like
  3. It is a race to compete for technology and domain knowledge. I would rather focus on solving business problems and learn required tech on a need basis than trying to master tech and go for domain
  4. In 20 years so much has changed in tech. To observe the new business process, how technology shapes the 2.0, 3.0 supply chain is great learning. Applying a mix of both domain and tech to be on par with the evolving landscape is my interest
  5. Health is going down these days. I would rather be comfortable and creative than pushing my limits. I wish to be a learner, coder. You cannot win all the battles. Pick and chose
  6. At some point, I need to write books, more blogs on my lessons both personal, professional. In a way, we are in a mentally pressured space. Too much competition and clarity and purpose and being clear needs constant revival and focused mindset
  7. I have lost interest in money. I have connected with dozens of startups. I personally know the landscape, the tech they are solving. In a way, I am happy with the tech and business problems. Able to see and see how things evolve is more important than titles and company.
  8. I believe teaching is equally important, mastering fundamentals. Adding value does not come from being online or replying to every email. It comes with deliverables, passion to do personally. We may not excel in everything but few things close to our heart we can be at our best
  9. Now looking back on my 24 hours deployment, those critical moments in my work I do cherish those awards and being able to stand in times of crisis. Always domain knowledge helped me shape up to solve things and push for them. 
  10. Learn more, Live the best you can, regret less, add less guilt. In the end, only memories matter, Halfway done. Keep the last half a bit more contented and satisfied.

Keep going!!!

March 22, 2021

Learning Notes - Azure Synapse Analytics

Everything gets evolved into the next level. Earlier it was SQL DW, Now it has evolved into MPP with ML components.

Azure Synapse Analytics – Azure Synapse Analytics is a new offering available on Microsoft Azure. It’s a combination of SQL Data warehouse (MPP offering), Apache Spark, pipelines, and a workspace to manage this entire ecosystem

What is dedicated SQL pool (formerly SQL DW) in Azure Synapse Analytics?

  • Dedicated SQL pool (formerly SQL DW) stores data in relational tables with columnar storage
  • PolyBase uses standard T-SQL queries to bring the data into dedicated SQL pool
  • Dedicated SQL pool uses PolyBase to query the big data stores.

Reference - Link

Architecture - Link

Dedicated SQL pool (formerly SQL DW) uses a node-based architecture.

Applications connect and issue T-SQL commands to a Control node. The Control node hosts the distributed query engine, which optimizes queries for parallel processing, and then passes operations to Compute nodes to do their work in parallel.

Similar to Map Reduce here you see distributed parallel processing. Hope to experiment few more examples in the next posts.

Building real-time enterprise analytics solutions with Azure Synapse Analytics 

Key Notes

  • Dedicated SQL pools
  • Serverless consumption pools
  • Azure Synapse Analytics


Workspace Features

  • SQL Pools
  • Spark Pools
  • Pipelines - Integration and Orchestration
  • All resources governed by common security model
  • Connected service to expand synapse
  • Linked services for Data Integration


Demo 1

  • Synapse Analytics workspace

Demo 2

  • Azure Synapse and Azure ML
  • Synapse Notebook
  • Hummingbird generates in onnx format
  • Connect to AzureML Workspace



Demo 3

  • Use the model in Synapse workspace



Happy Learning!!!

March 07, 2021

Datasets Discussions

OpenEDS: Open Eye Dataset

Keynotes

  • Open Eye Dataset, of eye-images captured using a virtual-reality (VR) head mounted display mounted with two synchronized eyefacing cameras
  • 12,759 images with pixel-level annotations for key eye-regions: iris, pupil and sclera
  • 252,690 unlabelled eye-images

Metadata data per participant:

OVERHEAD MNIST: A BENCHMARK SATELLITE DATASET

More Reads

HandSeg: An Automatically Labeled Dataset for Hand Segmentation from Depth Images

  • Curated the overhead imagery from multiple public sources
UC Merced Land Use Dataset

There are 100 images for each of the following classes:

  • agricultural, airplane, baseballdiamond, beach, buildings, chaparral, denseresidential, forest, freeway, golfcourse, harbor, intersection, mediumresidential, mobilehomepark, overpass, parkinglot, river, runway, sparseresidential, storagetanks, tenniscourt

We need more mnist for different domains.

Keep Thinking!!!

More open source Datasets

To boost more powerful ML models, we need quality datasets, crowdsourced free datasets. More than code/logic data is key. We need to democratize data to build better models.

  • mnist for plants
  • mnist for vehicles
  • mnist for fishes
  • mnist for animals
  • mnist for ships
  • mnist for toops
  • mnist for equipements
  • mnist for highways
  • mnist for indian vegetables 
  • mnist of indian pets
  • mnist of railways
  • mnist for Trucks
  • mnist for Gender and Age bucketization
  • mnist for footwear types
  • mnist for highway symbols
  • mnist for airport safety signals
  • mnist for skin issues
  • Deepfake based marketing
  • Deepfake based face swapping
  • mnist for Indian Cusines
  • mnist for beauty Items
  • mnist for tamil

This needs more collective work for better well-equipped models.

Keep Thinking!!!

Preparing for certification

A bit of confusion, learning for ideas implementation vs learning for certification. Hoping to make an honest attempt to cover. I am poor in MCQ always so slow and study revise, post, blog and see if it works :)

The handbook is key - Link

  • Building and training neural network models using TensorFlow 2.x
  • Image classification - CNN, ImageDataGenerator
  • Natural language processing (NLP) - binary categorization, multi-class categorization, LSTM
  • Time series, sequences and predictions

Youtube playlist - Link1 

Good Reference of materials - Link

Example codes - Link

For every topic

  • Refer TensorFlow documentation
  • Git Examples
  • Example codes
  • Summary of concepts, learning
  • Pycharm familiarity

Books

  • Data Science on the Google Cloud Platform - Valliappa Lakshmanan
  • Machine Learning Design Patterns - Valliappa Lakshmanan, Sara Robinson, Michael Munn

Start Slowly and Keep Going!!!

March 01, 2021

Back to Basics - Fundamentals - RNN - Transformers

 It needs a bit more careful *attention* to understand the crux of transformers. This lecture was useful

Slides - Link

Session - 

Transfer Learning

  • Use Neural Network on imagenet and finetune on custom data
  • Better performance than anything else
Convert words to vectors

  • One hot encoding
  • Scales poorly with vocabulary size
  • Sparse and high dimensional
  • Map one hot to dense vectors (Embedding matrix)
  • Finding Embedding matrix - Learn as part of tasks
  • Learn the Language model
  • Training on large corpus of text - wikipedia
  • N-Grams, Sliding Window forming rows
  • Binary classification - 0 / 1 - Neighbouring word or not
NLP Imagenet moment - Elmo / ULMfit

  • ELMO - bidirectional stack LSTM
  • ULMfit

Good Paper Read - SQuAD: 100,000+ Questions for Machine Comprehension of Text






Attention

  • Only attention no LSTM
  • Self-attention, positional encoding, Layer normalization
  • Attention and Fully Connected Layers

Self Attention

  • Input sequence of vectors
  • Output weighted sum of input sequence

Learn weights

  • Compute attention weight for its own output
  • Compute every other vector to compute attention weights for its own output y_i (query)
  • Compare to every other vector to compute attention weight w_ij for output y_j (key)
  • Summed with other vectors to form the results of the attention weighted sum (value)

Multihead attention

  • Weight matrices - query, key, value weights
  • Multiple heads of attention just mean learning different sets of query, key and value matrices simultaneously

Transformer

  • Self attention layer - layer normalization - dense layer



Layer Normalization

  • Data scaling, weight initialization
  • Rest things between uniform mean and standard deviation

Position Embedding

  • Word embedding depends on word
  • Position embedding depends on position 
  • Combine both and run through transformers
  • Both position and content reasoned

Attention is all you need

  • Translation
  • Encoder - Decoder architecture

GPT - Generative pretrained transformer

  • Generating text
  • ELMo, ULMFIT
  • Preceeding words
  • GPT2 1.5 Billion parameters

BERT

  • Bidirectional encoder representations from transformers

T5 - Text to Text Transfer Transformer

  • Input and output as text streams
  • 11 billion parameters

Keep Thinking!!!