Fantastic Models
and Where to
Find Them
Rishit Dagli
High School
TEDx, TED-Ed Speaker
rishit_dagli
Rishit-dagli
Slide 2
Slide 2 text
Fantastic Models and
Where to Find Them
Rishit Dagli
Slide 3
Slide 3 text
TensorFlow Hub
rishit_dagli
A repository of trained
machine learning models ready
for fine-tuning and deployable
anywhere
Slide 4
Slide 4 text
TensorFlow Hub
rishit_dagli
Slide 5
Slide 5 text
● High School Student
● TEDx and 2xTed-Ed Speaker
● I contribute to TF Hub
● I ❤ ML Research
● My coordinates - www.rishit.tech
$whoami
rishit_dagli Rishit-dagli
Slide 6
Slide 6 text
● Devs who have worked on Deep Learning
Models (Keras)
Ideal Audience
rishit_dagli
Slide 7
Slide 7 text
rishit_dagli
● Off the shelf inference with SoTA
machine learning models
Could be super helpful
rishit_dagli
Slide 8
Slide 8 text
rishit_dagli
●
● You do need to ask yourself:
○ Is there enough information in P to represent Q?
○ how much new data?
○ can unlabelled data do it?
○ what’s the right mix of P and Q data?
○ How far is from P from Q?
Transfer Learning
rishit_dagli
Slide 9
Slide 9 text
rishit_dagli
●
● You do need to ask yourself:
○ Is there enough information in P to represent Q?
○ how much new data?
○ can unlabelled data do it?
○ what’s the right mix of P and Q data?
○ How far is from P from Q?
● Samory Kpotufe: Some Recent Insights
on Transfer-Learning
Transfer Learning
rishit_dagli
Slide 10
Slide 10 text
rishit_dagli
●
● You do need to ask yourself:
○ Is there enough information in P to represent Q?
○ how much new data?
○ can unlabelled data do it?
○ what’s the right mix of P and Q data?
○ How far is from P from Q?
● You can do this easily with TF Hub
too!
Transfer Learning
rishit_dagli
Slide 11
Slide 11 text
rishit_dagli
● Less data
Could be super helpful
rishit_dagli
Slide 12
Slide 12 text
rishit_dagli
● Less data
● Less domain expertise
Could be super helpful
rishit_dagli
Slide 13
Slide 13 text
rishit_dagli
● Less data
● Less domain expertise
● for different domains including
image, video, audio etc.
Could be super helpful
rishit_dagli
Slide 14
Slide 14 text
rishit_dagli
● Less data
● Less domain expertise
● for different domains including
image, video, audio etc.
● Take advantage of SoTA machine
learning models
Could be super helpful
rishit_dagli
Slide 15
Slide 15 text
rishit_dagli
Ready to Use
rishit_dagli
Credits: @TheSandeepGupta