User of AI on Consumer Products - NU.pdf

Use of predictive AI in consumer products. Some algorithms in and example use cases

1. Aug., 2024Gautam Borgohain
AI INAI INAI IN
CONSUMERCONSUMERCONSUMER
PRODUCTSPRODUCTSPRODUCTS
Past, Present and Future of AI in everyday prodcuts

2. TODAY'S AGENDATODAY'S AGENDATODAY'S AGENDA
111
What is AI ?
222
AI in Consumer Product 1
333
AI in Consumer Product 2
444
The future .....

3. # UFC Singapore
ABOUT MEABOUT MEABOUT ME
Hello, I am Gautam
Data Scientist and Machine learning engineer by trade
10+ years in building end to end large scale real time ML systems
in various domains Recommendations to Computer Vision to Ads
ranking systems that serve 300 million mDAU
Outside of work I love martial arts and music
But of late I spend a lot of free time playing with LLMs

4. So when we think about consumer products, I am actually talking about products that we use on a daily basis
Consumer Software products
Consumer electronics by technology
companies
Food, Beverage, Household, Personal
Care etc
CONSUMERCONSUMERCONSUMER
PRODUCTSPRODUCTSPRODUCTS
What are they ?

5. WHAT IS AI ?WHAT IS AI ?WHAT IS AI ?
AI vs Machine Learning vs LLM vs Data Science etc....
Artificial
Intelligence
Reactive Machines
Limited Memory
Theory of Mind
Self Aware
Machine
Learning
Supervised Learning
Un / Semi Supervised Learning
Deep Learning
Reinforcement Learning
Further Reading: What is AI : Google, IBM
Generative AI
Text
Code
Image
Audio
Molecules etc...

6. WHAT IS AI ?WHAT IS AI ?WHAT IS AI ?
Predictive / Discriminative models
Need labeled data (unless using
unsupervised clustering)
Evaluation metrics by type of variable
being predicted :
Precision / Recall / F1 / MAE / RMSE
etc.

7. Auto Regression Sequence Transduction
Time Series Forecasting
Further Reading: M Competition
Seq 2 Seq models
Auto Encoders
Further Reading: Autoencoders;
Image source: AE vs VAE
WHAT IS AI ?WHAT IS AI ?WHAT IS AI ?
Predictive models - Auto Regression & Sequence Transduction

8. VAE GAN Diffusion
WHAT IS AI ?WHAT IS AI ?WHAT IS AI ?
Generative models
Image source: AE vs VAE
Image source and further reading :
dev.google-GAN
Image source: Diffusion models
Further Reading: Diffusion models

9. They can be used for Text as
well but lately as you know, its
been LLMs that have been
leading text generation

10. Language Models are probabilistic model of a natural
language that is applied to predict the “next” word
WHAT IS AI ?WHAT IS AI ?WHAT IS AI ?
“Large” Language Model
Source: Evaluating Language models
Large because they are trained on more data
and the have more parameters

11. Paper; How attention works
Attention is all you need
Word2Vec
RNNs & LSTM
Deep learning based
sequence models
Conditional Random Fields and
Hiddne Markov models
Sequence Model
If <>
then
reply <>
Rule Based LM
WHAT IS AI ?WHAT IS AI ?WHAT IS AI ?
“Large” Language Model

12. WHAT IS AI ?WHAT IS AI ?WHAT IS AI ?
“Attention is all you need”
paper introduced the concept
of transformer networks.
These networks consisted of
multiple attention blocks.
These attention blocks are in
some ways “smaller models”
that remember which parts of
the sequence to “attend” to.
Transformers need more
data though, so labs
started training on larger
corpuses of data and with
even more parameters. ->
Pre-trained models
paperBERT was the first such
model, starting the trend
of “Sesame Street” models
Transformer

13. Further Reading: Transformer Explainer

14. Now, we will explore two consumer products we all use and
learn how AI has changed how these companies develop
these product and how we use them
Recap
Predictive AI vs Generative AI
Predictive AI is more narrow - specialized to use cases while
Generative AI is almost use case agnostic but dependent on
modality

15. CONSUMERCONSUMERCONSUMER
PRODUCT 1PRODUCT 1PRODUCT 1
VACUUMSVACUUMSVACUUMS
Source: LG

16. VACUUMSVACUUMSVACUUMS
Its just a broom - What could it possible have to do with AI ?

17. AI IN CONSUMER PRODUCT -AI IN CONSUMER PRODUCT -AI IN CONSUMER PRODUCT -
VACUUMSVACUUMSVACUUMS

18. AI IN AI IN AI IN IDEATION AND DESIGNIDEATION AND DESIGNIDEATION AND DESIGN
Battery Testing
Find the best method of charging
a battery that maximizes it’s
overall lifetime
Use ML to predict how batteries
would respond to different
charging approaches.
Reduce testing from 2 Years to 16
Days
Further Reading: ML speed arriaval ultra fast charging electric car; more details

19. AI IN MANUFACTURINGAI IN MANUFACTURINGAI IN MANUFACTURING
Predictive Maintainence
Equipment condition monitoring, fault diagnosis, residual life
prediction and maintainence decision making
Use digital twins of products that reflects physical characteristives
to simulate real machine behaviour.
Increase performance prediction and health management
mechanism by four times.
Further Reading: Predictive Maintenance based on digital twin

20. AI IN MARKETINGAI IN MARKETINGAI IN MARKETING
Online Ads
Using your profile data to target
relevant ads to you
Use ML for Recommendations
Systems along with second price
auction. More recently used for
generating creatives.
Further Reading: Wide and Deep for Recsys ; Second Price ; Ad creatives with Pencil;
Source: Twitter ads prediction
Source: Twitter ads bidding

21. AI IN RETENTION &AI IN RETENTION &AI IN RETENTION &
PERSONALIZATIONPERSONALIZATIONPERSONALIZATION
Connected Features
Machine sensor and usage
data is streamed to
companies to provide smart
features
Increase customer engagement
and retention

22. Recap
Predictive AI models are applied throughout a product’s lifecycle.
The focus on automation or optimization and their efficiency unlocks new
capabilities for businesses.
Now, let’s look at how AI is used in another product, but this time a software...

23. CONSUMERCONSUMERCONSUMER
PRODUCT 2PRODUCT 2PRODUCT 2
MUSICMUSICMUSIC
STREAMINGSTREAMINGSTREAMING
Source: LG

24. MUSIC STREAMINGMUSIC STREAMINGMUSIC STREAMING
Source: Slashgear, cassette
tapes are making a comeback
Source: Whathifi
Source: Redpoint
Fun video: extreme audio setup video

25. Predictive Content Caching
MUSIC STREAMING - FASTMUSIC STREAMING - FASTMUSIC STREAMING - FAST
Music tracks are stored as files in servers around the world and the app fetches the file in chunks (usually 512Kb)
Metrics : Playback latency (time from click to sound) and Stutter (number of skips/pauses during playback)
Solutions:
Reduce delays and buffering by anticipating what a user will listen to next <- ML1.
Implement BBR (a TCP congestion control algorithm) which uses an optimum ‘send rate’ based on client.2.
Further Reading: Spotify engineering

26. MUSIC STREAMING -MUSIC STREAMING -MUSIC STREAMING -
EXPERIMENTATIONEXPERIMENTATIONEXPERIMENTATION
Uplift modeling
Source: Uplift modeling at Uber; Uplift modeling explainer

27. MUSIC STREAMING -MUSIC STREAMING -MUSIC STREAMING -
EXPERIMENTATIONEXPERIMENTATIONEXPERIMENTATION
Source: Uplift modeling at Uber; Uplift modeling explainer

28. Recap
AI can be thought of as a ‘function’... a general purpose function
that train to do specific tasks for us.
Such a function can help us to identify which songs to prefetch and
can also be trained to help us understand our customers via data.

29. THE FUTURETHE FUTURETHE FUTURE

30. THE FUTURETHE FUTURETHE FUTURE

31. THAT'S ATHAT'S ATHAT'S A
WRAP!WRAP!WRAP!
Thank you, Questions ?

32. Books:
Other sources:
Engineering blogs : Meta, Google, Spotify