Sign off 2020

2020 was a very different year
It was challenging both personally and professionally
Professionally it was a long stretch to learn/work on time-bound projects
Teaching was ok-ok
Personally not so well, overall.
A big thank you to everyone with whom I shared my 2020. Apologies for anyone whom I may have hurt. At the end of 2020 in terms of memories are some vague travel moments, work from home always.

Keep Going!!!

Mental Health & AI - Research paper reads

Paper #1 - Forecasting the onset and course of mental illness with Twitter data

Key Notes

• Dataset - 204 individuals (105 depressed, 99 healthy).
• Models - Classify depressed and healthy content, supervised learning algorithms, time series model
• Depression­ related terms such as diagnosis, antidepressants, psychotherapy, and hospitalization

Paper #2 - Predicting User Emotional Tone in Mental Disorder Online Communities

Paper #3 - Mental Health and Sensing

Features

1. Facial expression
2. Heart rate variability
3. Eye movement
4. Electrodermal activity
5. Mobility and location
6. Speech patterns
7. Technology use
8. Activity
9. Social interaction
10. Communication patterns
11. Social media

More Reads

Big Data Analytics and AI in Mental Healthcare

What are We Depressed about When We Talk about COVID19: Mental Health Analysis on Tweets Using Natural Language Processing

Temporal Mental Health Dynamics on Social Media

Personal Mental Health Navigator: Harnessing the Power of Data, Personal Models, and Health Cybernetics to Promote Psychological Well-being

Keep Thinking!!!

Five Things for better User Privacy in Social Media, Ads, Google, Youtube

• Apply policy of Individuals have the right to request the restriction or suppression of their personal data in all the countries
• Explicit consent requires a very clear and specific statement of consent with the purpose
• Investigate on Statistical purposes of Data usage at Individual Level
• What level of data is aggregated at the user level. Apply the policy of restriction, archival
• Callout every Ads based on identifiable data that was used (Name, IP address, Location data)
• Warn them of Obsessive behavior but do not alert it or use it for advertising
• Advise them if you spot depression symptoms but do not remember the historical data more than a few months
• Have a more social approach toward better sex education, awareness, restrict on content but provide more relevant content, not addiction

After months of stalling, Google finally revealed how much personal data they collect in Chrome and the Google app. No wonder they wanted to hide it.
⁰
Spying on users has nothing to do with building a great web browser or search engine. We would know (our app is both in one). pic.twitter.com/lJBbLTjMuu

— DuckDuckGo (@DuckDuckGo) March 15, 2021

Keep Thinking!!!

Review GDPR

Key Points

Paper - Link

Key notes

• General Data Protection Regulation (GDPR)
• Controller determines the purposes
• Processor is responsible for processing personal data on behalf of a controller

Personal Data

• Information about identified or identifiable individual
• Name, IP address or a cookie
• Content of the information, the purpose
• Identification number;
• Location data; and
• Online identifier

Impact on Individual

• The content of the data – is it directly about the individual or their activities?;
• The purpose you will process the data for; and
• Results of or effects on the individual from processing the data
• Lawfulness, Fairness, Transparency
• The GDPR does not dictate how long you should keep personal data. 

Consent: the individual has given clear consent for you to process their personal data for a specific purpose.

• Are they vulnerable?

Purpose

• Scientific or historical research purposes; or
• Statistical purposes

Key Points

• Avoid making consent to processing a precondition of a service
• Explicit consent requires a very clear and specific statement of consent
• Are you processing children’s data?
• Is any of the data particularly sensitive or private?
• Sensitive Data - race;ethnic origin;politics;religion;trade union membership;genetics;biometrics (where used for ID purposes);health;sex life; or sexual orientation.

Rights for individuals:

• The right to be informed
• The right of access
• The right to rectification
• The right to erasure
• The right to restrict processing
• The right to data portability
• The right to object
• Rights in relation to automated decision making and profiling.

This part of the guide explains these rights. Individuals have the right to request the restriction or suppression of their personal data.

The right to data portability allows individuals to obtain and reuse their personal data for their own purposes across different services.

I am not sure how the data in FANG (Facebook, Amazon, Netflix, Google), Microsoft, Cookies, Browser info how much they are used to what extent :(

We hire people to handle laws, circumvent the clauses. There is always a catchup game being bending rules vs line of privacy. Debatable from both sides.

Do we have clarity from Reliance, Airtel, Flipkart other Indian providers for their GDPR similar data usage, retention, user consent. Time to check on this!!!

Good Read - Link2

Good Read - Ethical AI

Keep Thinking!!!

Data Governance - Research paper Reads

Data Governance for Platform Ecosystems: Critical Factors and the State of Practice

Key Notes

• Data Challenges - data abuse, privacy violation and proper distribution 
• Data Governance - (availability, usability, security and privacy)

Data Governance Factors

• Data Ownership access - presents who owns and uses the data in platform ecosystems.
• Data ownership definition criteria
• Monitoring - Invisible supply chain is a longstanding challenge
• Conformance - Audit for compliance based on strict processes and rules
• Data Use Case - Use the data in platform ecosystems
• Data provenance - Data transparently for all participating groups
• Contribution Estimation - User contribution against value creation by providing data

AI GOVERNANCE FOR BUSINESSES

Key Notes

• AI exhibits forms of intelligent behavior allowing for a large range of cost-efficient, wellperforming applications
• AI produces results that are partly outside the control of an organization or at least unexpected. It exhibits non-predictable, “ethics”-unaware, data-induced behavior yielding novel security, safety and fairness issues
• To mitigate AI challenges and to raise AI potentials in organisations, governance mechanisms play an important role
• Testing of ML models, ensuring fairness, explaining “black boxes”, data valuation
• Data-driven lens of AI, based on the observation that most existing AI techniques
• Prominent regulations include the European GDPR that touches upon data as well as models. Compliance monitoring, Audit

Data

• Data is the representation of facts using text, numbers, images, sound or video
• An essential characteristic of data is in addition also the primary source of data: Is it personal or non-personal?
• GDPR 2018 grants the right to explanation to individuals for automated decisions based on their data
• Governance model based on fairness, transparency, trustworthiness, accountability.

Model Explainability

Transparent models are intrinsically human understandable, whereas complex black-box models such as deep learning require external methods that provide explanations that might or might not suffice to understand the model

Data Valuation - valuation of data gains in relevance, if the acquisition of data comes with costs, e.g. data has to be labeled by humans as part of the construction of a dataset for an AI system or data requires costly processing, such as manual cleansing to raise data quality

• Data quality denotes the ability of data to meet its usage requirements in a given context
• An important data quality aspect with respect to fairness of ML systems is bias
• Data is biased if it is not representative of the population or phenomenon of study.
• Concept drift implies that the data used to train ML model does not capture the relationship that the model should capture
• Robustness defines to what extent a ML model can function correctly in the presence of invalid inputs 
• Protected characteristics such as gender, religion, familial status, age and race must not be used
• ML should allow to track provenance/lineage, ensure reproducability, enable audits and compliance checks of models, foster reusability, handle scale and heterogeneity, allow for flexible metadata usage

Unionized Data Governance in Virtual Power Plants

Collective bargaining The asset-owners should be able to bargain collectively about the conditions and purposes of the data flows. This includes which supplementary data flows to include and how to utilize them

Representation The asset-owners should be represented in a central organizational governing body, which is in charge of defining and overseeing the data principles.

Accountability Transparency measures should be put in place to ensure the asset-owners ability to audit the data usage performed by the aggregator, in order to detect misuse and assign accountability

Social and Governance Implications of Improved Data Efficiency

Data Readiness Report

Alternative Personal data governance models

Design Choices for Data Governance in Platform Ecosystems – A Contingency Model

Data Governance Strategies from Experience

Happy Learning!!!

Fashion Paper Reads - Part II

Paper #1 - Detailed Garment Recovery from a Single-View Image

Key Notes

• Global shape and geometry of the clothing
• Extract occluded wrinkles and folds
• Parameter estimation, semantic parsing, shape recovery, and physics-based cloth simulation

The current implementation of our approach depends on two

• databases: a database of commonly available garment templates
• database of human-body models.

Implementation Notes

• 14 joint positions on the image and provides a rough sketch outlining the human body silhouette
• Semantic parse of the garments in the image to identify and localize depicted clothing items

Human body - We follow the PCA encoding of the human body shape presented in [Hasler et al. 2009]. The semantic parameters include gender, height, weight, muscle percentage, breast girth, waist girth, hip girth, thigh girth, calf girth, shoulder height, and leg length

Garment Parsing

• we extract the clothing regions Ωb,h,g by performing a two-stage image segmentation guided by user sketch
• Initial garment registration results. We fit garments to human bodies with different body shapes and poses.

Implementation - We have implemented our algorithm in C++ and demonstrated the effectiveness of our approach throughout the paper

Paper #2 - M2E-Try On Net: Fashion from Model to Everyone

• Pose alignment network (PAN) - pose alignment network (PAN) to align the model and clothes pose to the target pose
• Texture refinement network (TRN) - enrich the textures and logo patterns to the desired clothes
• Fitting network (FTN) - merge the transferred garments to the target person images.
• Unsupervised learning and self supervised learning to accomplish this task.
• Generative adversarial network (GAN) [9] has been used for image-based generation
• GAN has been used for person image generation [18] to generate the human image from pose representation
• For fashion image generation, a more intuitive way is to generate images from a person image and the desired clothes image

Dataset - Deep Fashion [17] Women Tops dataset, MVC [16] Women Tops dataset and MVC [16]

PAN 

• PAN as a conditional generative module
• To train PAN, ideally we need to have a training triplet with paired images: model image M, person image P, and pose aligned model image
• self-supervised training method that uses images of the same person in two different poses to supervise Pose Alignment Network (PAN)

Texture Refinement Network (TRN)

• Combine the information from network generated images and texture preserved images produced by geometric transformation
• Loss - Reconstruction loss, perceptual loss and style loss are used only for paired training

Paper #3 - VITON-GAN: Virtual Try-on Image Generator

Trained with Adversarial Loss

Code - Link

Paper #4 - GarmentGAN: Photo-realistic Adversarial Fashion Transfer

This method divides the image generation task into two sub-tasks: segmentation map synthesis and transference of the clothing characteristics onto the previously generated map

The system comprises two separate GANs: a shape transfer network and an appearance transfer network

More Reads

• GarmentGAN: Photo-realistic Adversarial Fashion Transfer
• Mining Fashion Outfit Composition Using An End-to-End Deep Learning Approach on Set Data
• Algorithmic clothing: hybrid recommendation, from street-style-to-shop
• Aesthetic-based Clothing Recommendation

Keep Thinking!!!

Weekend Reading - Fashion Research Paper Reads # Part 1

Paper #1 - Outfit Recommender System

Perceptions

• Christian funeral, wearing black conservative clothes is customary
• Hindu funeral, wearing white conservative clothes is the norm

A recommender system is used to suggest products to customers by using information about the customer. Feature Descriptors - Histogram of Oriented Gradients (HOG), Speeded Up Robust Feature (SURF), Scale Invariant Feature Transform (SIFT)

• Event Analysis using Detected Objects
• Fifty-three categories of clothes

1. Multilabel classification and cloth detection
2. Approach Events - Clothes Mapping
3. Event to Outfit Mapping

Crop the images to reduce the background information

Paper #2 - Artificial Intelligence and the Fashion Industry

• Datafication of fashion - data generation through the digitization of content,and monitoring of activities, including real world activities and phenomena,through sensors
• Image Search. Personalized shopping is also achieved using AI applications based on computer vision and augmented and virtual reality.
• Reverse image search instead is the process by which an image is used to find another image.
• Visual search, a subset of reverse image search, refers to the possibility of finding items within an image and searching for those.

Paper #3 - Recommendation based on multiproduct utility maximization

Multi-product utility maximization (MPUM) - integrates the economic theory of consumer choice theory with a personalized recommendation

Two products could be substitutes - buy A instead of B or complements - buy A together B. Identifying and making use of such relationships are useful for recommendation systems.

We assume user make choices to maximize multi-product utility and use the multinomial consumer choice model for that, then the multi-product utility model can be learned to maximize the likelihood of observed user data

Collaborative filtering is based on the assumption that users with similar tastes for previous items will have similar preferences for new items

Content-based filtering is based on the assumption that the features (meta data, words in the description, price, tags, visual features, etc.) 

Hybrid recommendation algorithms combine collaborative filtering with content-based filtering

Code - Link1, Link2

Keep Learning!!!

Simple Utility - File Renumbering

 

Script to renumber files in all subdirectories. Bookmarking it for future use.

Keep Learning!!!

Weekend Reading - Resilient Distributed Datasets

 Paper Link 

Key notes

• Resilient Distributed Datasets (RDDs) - a distributed memory abstraction that lets programmers perform in-memory computations on large cluster
• Data reuse is common in many iterative machine learning and graph algorithms, including PageRank, K-means clustering, and logistic regression. 
• RDDs are a good fit for many parallel applications because these applications naturally apply the same operation to multiple data items. 
• RDDs are fault-tolerant, parallel data structures that let users explicitly persist intermediate results in memory, control their partitioning to optimize data placement
• If a partition of an RDD is lost, the RDD has enough information about how it was derived from other RDDs to recompute
• Formally, an RDD is a read-only, partitioned collection of records. RDDs can only be created through deterministic operations on either (1) data in stable storage or (2) other RDDs
• RDD has enough information about how it was derived from other datasets (its lineage) to compute its partitions from data in stable storage
• The main difference between RDDs and DSM is that RDDs can only be created (“written”) through coarse-grained transformations, while DSM allows reads and writes to each memory location

Keep Thinking!!!

Interesting projects and their tech choices

Project #1 - Tech for SaaS built by One Person

• Language - Python, Typescript
• Frameworks - Django, React, NextJS, Celery, Bootstrap 4
• Databases - Clickhouse, PostgresSQL, Redis
• Deployment - Terraform, Docker, Kubernetes, CircleCI

Related Good Read - Link

Project #2 - Raspberry Pi Webcam

Project #3 - Preventing Against Dos Attacks, Good Read - Someone attacked our company

Keep Thinking!!!

Kubeflow Pipelines - Learning Notes #1

To appreciate something we need to why, how, what about the Tool.

Key Notes

Why / Necessity

1. Monitoring of Model
2. Training /Serving - Differences in transformation, handling missing data
3. Frequency to refresh the model

Production System Components

Kubeflow Platform

Develop, Deploy, Manage

Pipelines, Data Management, Serving (Rest End Point) 

Pipeline Component

Commands

Setup cluster, permissions in yaml file

Demo with screenshots

Pipelines

• Domain-Specific Language
• Instantiate Components
• Define Dependency between components
• Compile and Deploy Pipeline

Custom Components

Somehow the gap between ML code vs kubeflow code there is a lot of learning. How much time it takes to port to this infra? I need to experiment to comment. A lot of features are there but we shouldn't end up rewriting ML code to pipeline code. 

Notes #2

Codify ML Workflows

Adopt pipeline mindset

Experiment, Reproduce, Share pipeline

Define Pipeline

• The description on ML Workflow
• Runs on Container
• Execution vs Runtime decoupled
• Components - one step of workflow
• Component - Packaged as Docker image
• Pod for Each Step
• Pipeline SDK

More Reads

Example

pipeline sdk key notes - Link1, Link2, Link3

SDK Summary pointers

KALE (Kubeflow Automated pipeLines Engine) is a project that aims at simplifying the Data Science experience of deploying Kubeflow Pipelines workflows.

An Argo workflow executor is a process that conforms to a specific interface that allows Argo to perform certain actions like monitoring pod logs, collecting artifacts, managing container lifecycles, etc.

Katib is a Kubernetes-native project for automated machine learning (AutoML). Katib supports hyperparameter tuning, early stopping and neural architecture search (NAS). Learn more about AutoML at fast.ai, Google Cloud, Microsoft Azure or Amazon SageMaker

Katib is the project which is agnostic to machine learning (ML) frameworks.

Ray Train, an easy-to-use library for distributed deep learning.

Dask is a flexible library for parallel computing in Python.

ML metadata (MLMD) library by Google. MLMD is an integral part of TensorFlow Extended (TFX) and a stand-alone application

The most important entities created and stored by MLMD are:

• Artifacts that are generated by the pipeline steps (e.g., the trained model).
• Metadata about the executions (e.g., the step itself).
• Metadata about the context (e.g., the whole pipeline).

Keep Thinking!!!