Understanding Machine Learning Model from industrys expert

VIDEO DESCRIPTION:

Play, download and edit the free video Understanding Machine Learning Model from industrys expert.

Understanding Machine Learning Model from industry’s expert

Ankita Chauhan - NASSCOM Product Council
In conversation with Dr. Om Deshmukh – Senior Director, Data Sciences; Envestnet | Yodlee

Topic - Building Machine Learning Model from Scratch Theory and Tutorial

The most asked question is “At what point, can we get 100% accuracy?”
The definition of 100% accuracy is very illusive.
Example: If you search for Jaguar, Google shows results of Jaguar car. But there can be other right results also like a Leopard or even a bathroom fitting.
This is the grey area where a lot of times ML solutions have to play an important role.

Decision making has changed with the advent of data. The steps involved in decision making are
• Formulating Hypothesis
• Evaluating Hypothesis
• How to get datasets
• Formulating data-driven insights (with caveats called out)
• Driving action by providing expert interpretation of the data outcomes

Steps for creating a product which is AI/ML/DS driven
• Know which problem to solve
• Where is the data? Data limitations?
• Scope/Scrub the data
• Formulate into a AI/ML framework
• Data “Swimming”
• Data representation
• Train for generalization; Optimize deployment
• Synergy with UI/UX for a consistent message
• Feedback/ regular checks and upgrades

Know which problem to solve / Discovery Phase
• Work with Product Team/Domain Experts
• Have Dreaming session with clients/partners
• Business problem  Functional Problem  Data Driven Problem

Where is the data? Data limitations?
• What is the “ideal” dataset?
• Settle for practical data as some systemic gaps exist in ideal dataset
• Leads to “holes” & “biases” in data

Scope/Scrub the data
• Impute missing values
• Sample the data
• Data bias often underestimated

Formulate into an AI/ML framework
• Most difficult phase and needs intensive training
• Solution has to be scalable
• “Obvious in hindsight”

Foundational work has irreversible impact on when and where the solution reaches “point of diminishing returns”

Objectivity and Subjectivity in AI/ML Models
Objectivity
• The function to optimize is well-defined, fixed and doesn’t change across data samples
• Different cost functions will lead to different optimum solutions
Subjectivity
• Human labels about right vs. wrong are used to train the system
• Only as accurate as the humans are accurate and consistent
• Repeatability and reproducibility metrics vary based on the complexity of the task

Human learning is more flexible, and we can “Self-teach” as we get more and more data. Example, as a kid, we only knew cycle as the only vehicle, but over the years as we got more data, we got to know that there are more vehicles like Bike, Car, Bus etc.

UI/UX can camouflage ML shortcomings
Where your system is not confident, you want to have a way in which you let the user know that the system is not confident.
Example: spellchecking ‘Teh’ to ‘The’ or ‘if you spellcheck visitpalace’ it gives you an option of what the likely correct answer can be but if you type some gibberish such as ‘xxhhdjdjdhhj’ the spellcheck isn’t confident hence doesn’t give you any spelling suggestions.
(Add the last screenshot)

Download, play and edit free videos and free audios from Understanding Machine Learning Model from industrys expert using RedcoolMedia.net web apps