Domain + Tech + Impact

My role has always been innovating / initiatives/domain knowledge-driven based use cases for more revenue opportunities / lowering the cost of operations / better customer service. Recollecting some of my milestone projects.

Reverse Logistics

Customer Service Projects

• Better delivery insights/emails to measure status at each leg
• Provided more touchpoints for better repair/refurbishment delivery

Warranty rewrite

• Rewrite warranty with traceability to new rules
• Data lineage for different warranty rules
• Tracking between repairs/exchanges

Vision for Retail Innovation

• RFID, EAS, and legacy devices for people counting, loss prevention vs vision-based solutions
• Ideate, prototype, demonstrate, patent. I wasn't there to collaborate or see how intel scaled it up but happy for the ideas that went till NRF / products

Startups collaboration

• Vision for ad effectiveness, Measuring sales impact from digital displays
• Vision for logo damage assessments - measuring logos to be replaced due to wear and tear in aircrafts
• Vision for Agriculture - Duplicate vendor detection and alert

Startups pitch and failed attempts. Vision doesn't work alone. It has to be a combination of vision + data to be a successful product.

Keep Thinking!!!

AL / ML Work - The three categories to focus on

Research perspective

• Apply new approaches + solve new problems
• Differentiation in terms of approach / performance / patenting / renewed opportunities

Practitioners

• Apply ML Use cases in current projects
• Bring in the project insights/awareness within the Project team
• Differentiation in terms of moving towards ML adoption/implementation

Business Perspective

• Contribute to AI-driven business lens focused use cases in the domain
• Build industry focused generic solutions 
• Differentiation in terms of wider community impact / collaboration with business / clients

Mastering the latest tech vs Coding vs ML ops vs Building domain knowledge, Scale as much as you can, or narrow down and pick your battles!!!

Keep Thinking!!!

The challenges to put ML models in production (Healthcare)

 Very good thread, Summarizing insights

Can you detect COVID-19 using Machine Learning? 🤔

You have an X-ray or CT scan and the task is to detect if the patient has COVID-19 or not. Sounds doable, right?

None of the 415 ML papers published on the subject in 2020 was usable. Not a single one!

Let's see why 👇 pic.twitter.com/Vrd91ZpXy3

— Vladimir Haltakov (@haltakov) June 9, 2021

Observations from papers ?

• None of the 415 ML papers published on the subject in 2020 was usable. Not a single one!
• Black small square 2212 papers, Black small square 415 after initial screening, Black small square 62 chosen for detailed analysis, Black small square 0 with potential for clinical use
• Many papers were using very small datasets often collected from a single hospital - not enough for real evaluation
• Some papers used a dataset that contained non-COVID images from children and COVID images from adults. These methods probably learned to distinguish children from adults
• Training and testing on the same data 
• Many papers failed to disclose the amount of data they were tested or important aspects of how their models work leading to poor reproducibility and biased results
• Many papers didn't even consult with radiologists.
• Rushing to publish results based on small and bad quality datasets undermines the credibility of ML
• At some point people start figuring out how to fine tune on the test set
• Dataset is not diverse enough and bias-free
• Authors find that covid-19 detectors often attend to the position of the shoulders and not the lungs. Models can easily learn shortcuts as opposed to robust features

Take everything with a pinch of salt. Real world data is not kaggle data. Kaggle does not reflect the reality or quality or the challenges we spot on data.

Keep Exploring!!!

How to get download your favorite video, audio from youtube

1. Download from youtube using python - youtube-dl package
2. Clip a portion of video - vlc advanced editor. Steps - link, C:\Users\username\Videos
3. Convert video to audio - ffmpeg

ffmpeg -i sample.avi -q:a 0 -map a sample.mp3

Steps - link

Keep Checking!!!

Retail and Supply Chain Reads

Retail and Supply Chain Reads.

Learning Tech vs Knowing current trends in each domain vs Getting insights from different reports vs Building your own insights/predictions, The cycle never ends!!!

Retail Trends Playbook 2020

Key Notes

• Retailers can transform data into dollars by using customer information to determine better marketing, service and product opportunities.
• Data intelligence is a key ingredient to the customer experience
• success in this new era is dependent on understanding and anticipating the needs of customers at every stage of the retail journey.

Consumer trends

• Customers are Comfortable Sharing Data In Exchange For Better Experiences
• Customers expect Personalization At Every Stage
• Customers want Knowledgeable Staff On-Hand For Service And Support
• Data-Powered Warehouses Reduce Cost Of Ownership
• Consumers Spend More With Better Service

Supply Chain Resilience Report 2021

Key Notes

• The number of organizations who report on disruptions is continuing to increase
• COVID-19 has increased the number of organizations using technology for supply chain management
• Over half of organizations admitted COVID-19 increased their use of technology for supply chain mapping
• Cross border land transport has been the primary cause of logistics disruption in 2020
• Management have become more committed to managing supply chain risk

Beyond COVID-19: Supply Chain Resilience Holds Key to Recovery

Key Notes

• COVID-19 has now unleashed a global supply chain crisis across a huge number of organizations, stemming from a lack of understanding and flexibility of the multiple layers of their global supply chains and a lack of diversification in their sourcing strategies.
• Where possible, diversified supply-chains across companies and geographies greatly
• reduce exposure and if firms are tied to single suppliers, risks from supply-chain disruptions 
• Big data analytics can assist firms in streamlining their supplier selection process, cloud-computing is increasingly being used to facilitate and manage supplier relationships and logistics and shipping processes can be greatly enhanced through automation and the internet of things. 

Supply-chain recovery in coronavirus times—plan for now and the future

Key Notes

Keep connecting the dots !!! Diversified and predicting risks / setting up a Digital connected supply chain!!!

Complicating computer vision use cases

Sometimes we design solutions because we have to use the existing environment. They may be feasible simpler / better solutions. We need to plan computer vision from hardware/software / better solutions not over overengineered solutions

Use Case #1 - Fisheye camera for dwell time computation for curbside pickup

Implementation Challenges

• Fisheye has different sizes for far/near objects
• Side view at the entrance
• Top view at the exit
• One model for entrance / one model for exit
• Occlusion problems
• Detecting vehicles types / Detecting attributes
• Compare and match and count 

Ideal way

• Edge Device near entry/exit capture license plate

Lesson learned - This needs a custom model for every type of camera. Not the best way to implement, I did call out but it was beyond my influence to change the aspects. I was partially happy with the approach although we could do it much better and simpler.

Use Case #2 - Vehicle counting in a four-lane junction

Environment challenges

• Service roads
• No clear division of vehicles
• Vehicle flow in all directions
• Limited space between vehicles
• Occlusion
• All kinds of vehicles/colors/models

Ideal way

• Number plate license number recognition 
• Having vehicle counting camera set on individual roads than at junction to have a clear line of crossing for people counting

Lesson learned - Indian roads are very different. We have to design solutions from hardware, model implementation, challenges. It's best to design considering all constraints and real-world scenarios. Vision is not going to solve everything unless you are not smart enough on how to use vision and get results in a quick time.

Object Tracking - Tracking is memory intensive, In many ways, I try to avoid tracking because it needs exhaustive computation and frame-by-frame operations. Demo projects work great with frame-by-frame tracking but we cannot do the same on terabytes of videos.

Keep Thinking!!!

What all to learn ? The endless to-do list

First, it was Docker, then Next it was Kubernetes. KFServing + Vision + Mysql for an end-to-end architecture on GCP. Next for project purpose worked on Vision + AWS. Worked on AWS Redshit for AI / ML and Stored Procedures. Then Again jumped on to Azure Synapse, Azure Cognitive Vision + Media Services. On and off Picked up on optimization/routing/pulp / seems interesting. Domain was mostly CVM, Supply Chain, Retail and picking up on trends

When you look back from time to time there is a lot of stuff to catch up on the vision, new approaches, ML is giving more MLOps, Feature stores, Vision new tracking techniques, Everything transformer, bert. Expertise in Vision vs Data vs NLP everything needs time, experimentation, and constant exploration.

Domain Learning, Cloud Exposure, Data Learning, ML Techniques, Vision updates :) The list is never-ending!!!

Bookmark your learning's time to time. Looking back it feels like explored everything a bit but still so many unknowns. As always sometimes all these dots will help to paint a big picture. After an experiment you know a bit of implementation, The curiosity goes once you get the crux of it. The same TSQL, SQL Basics, performance tuning translated from RDBMS in 2000 to CAP theorem, the advent of NoSQL, large scale parallel processing with Hadoop, Columnar indexes, JSON models, Making everything read-only with Spark RDD. What is a limitation today that will be a motivation for something tomorrow?

What we know is a spark, curiosity and the ability to build the perspective connecting old and new things, learning the required pieces building up a good business use case is the satisfaction of learning!!!

Keep Going!!!

OpenCV Experiments

Experiment with

• Logging
• Threads
• Error Handling
• Writing text on Frame with grey background

Keep Exploring!!!

ffmpeg

What is ffmpeg 

• libraries and programs for handling video, audio, and other multimedia files and streams

Why it is used

• FFmpeg program itself, designed for command-line-based processing of video and audio files

What purpose it serves best

• ffmpeg project tools and programs form the basis of a ton of video and audio software. VLC, Google Chrome, and MPC all use ffmpeg libraries to decode video.

Download and Install from Link

In System variables locate and select Path row, click Edit and add value c:\ffmpeg\bin.

Live stream local video

ffmpeg -i sample.mp4 -v 0 -vcodec mpeg4 -f mpegts udp://127.0.0.1:23000

ffplay udp://127.0.0.1:23000

Keep Thinking!!!

Tabu Search

Ref1 - Link 

Notes

• Iterative procedure 
• Initial feasible solution to next solution in the current neighbourhood

Ref2 - Link

Notes

• The basic form of Tabu Search (TS) is founded on ideas proposed by Fred Glover (1977, 1986).
• Ordinary local or neighborhood search, proceeding iteratively from one point (solution) to another until a chosen termination criterion is satisfied.
• The most common attributive memory approaches are recency based memory and frequency-based memory
• Recency-based memory is the most common memory structure used in TS implementations.
• A key element of the adaptive memory framework of tabu search is to create a balance between search intensification and diversification.

Ref3 - Link

• A Comparative Study of Tabu Search and Simulated Annealing for Traveling Salesman Problem 
• Tabu List: To prevent the process from cycling in a small set of solutions, some attribute of recently visited solutions is stored in a Tabu List, which prevents their occurrence for a limited period.
• Diversification: The use of frequency information is used to penalize nonimproving moves by assigning a larger penalty
• Neighborhood: A neighborhood to a given solution is defined as any other solution that is obtained by a pair wise exchange of any two nodes in the solution

Book - Algorithms For Dummies

Notes

AVOIDING REPEATS USING TABU SEARCH

Tabu Search does the following

• Allows use of a pejorative solution for a few times to see whether moving away from the local solution can help the search find a better path to the best solution
• Remembers the solutions that the search tries and forbids it from using them anymore
• Creates a long-term or short-term memory of Tabu solutions by modifying the length of the queue used to store past solutions. 

Book - Metaheuristics and Evolutionary Algorithms

The TS is an enhancement of local search (LS) methods.

• TS solved the problem of convergence to local optima experienced with LS methods by allowing movements to non‐improving solutions when there is no better solution near the current solution. 
• Unlike the LS, in the TS the best neighboring point replaces the searching point even if it is worse than the current searching point
• The previous searching point is memorized as tabu
• If the new searching point is better than the best point, the new point replaces the best point; otherwise the best point remains unchanged.
• Neighboring points near the new searching point are generated. 

More Reads

• tabuSearch - R
• Ref4 - Link
• Code example - Link
• A Tabu Search-Based Algorithm for Airport Gate Assignment: A Case Study in Kunming, China
• Optimization Techniques — Tabu Search
• Code link 

Keep Thinking!!!