7. Natural Language Generation
• Simple, factual, data
– Ex: financial data: « At 1pm
Stock <X> is <up|down>, as is
<Y> »
• Context-aware content
– Ruleset
– A « true » agent that learns
from thousands or millions
examples: chatbots
8. Signal processing: fraud detection, automation,HF
Trade…
• Analyze a signal
• Derive model from past
signal
• Forecast/predict
See « Fraud detection via AI », Arnaud Wald,
18h30, Callisto
9. Product recommendation
• Suggest the right product at the right
time to the customer is a cardinal rule
of commerce
• Error prone & time consuming if
manual
• Retail operations generate large
volume of behavioural data
– Order history
– Session time / heat maps
– In store & online traffic patterns
10. Product recommendation
• Build an order history matrix ( buyer x product):
– Very large matrix (500k customers & 10k
products: 5B cells!)
– Extremely sparse
• From order history matrix, build a product x
product matrix
– Much smaller
– Symetrical
• Attraction between 2 products becomes a
similarity measure between two vectors: dot
product!
1 2 0 3
2 1 3 5
0 3 1 0
3 5 0 1
A
11. From neural networks to deep learning
A form of massive multi layered neural networks:
• Iterate over a dataset
• Measure accuracy between prediction &
reference
Based on:
• Large Datasets availability (open & closed
sources)
• Increase in computional power & storage
12. Deep learning examples
Inputs:
profile images at an angle
Outputs:
frontal images of the face
Training dataset:
• 700 000 pairs
• 13GB
See « Harnessing the power of GANs »
Olga Petrova, 17h, Amalthée
13. Deep learning examples
« Learning the Depths of Moving People by Watching Frozen People »
Zhengqi Li & al, April 2019
14. AI challenges & solutions
• Business process modelling
• Data availability
• Business validation - explainability
• Solutions
• Costs
15. Modelling the business case
• What’s the UC?
• Forecast
• Create value
• Reduce costs
"Computers are useless. They can only give you
answers."
Pablo Picasso
16. Data Availability
Map the data pertaining to the UC
• Internal (siloed?)
• External
• Line of business
(competitors, trade
associations,…)
• Opensource (gov bodies,
resarch…)
18. Business validation & explainibility
Stakeholders must:
• Be domain experts
• Accept results from « the machine »
AI inception incoming:
• AI explainability is still an open field
19. AI & specialized hardware
Heavily computationnaly intensive:
• Process many arithmetic operations in parallel
• Trainable parameters: ~ 8M in the frontalization
GAN example
Solution: specialized hardware
• GPUs are optimised for parallel arithmetics
• NVIDIA Tesla P100 3584 cores
20. AI Hardware costs
GP1-XS 4 vCPUs RENDER-S
(Tesla P100 GPU)
Pricing
39€/month
0.078€/hour
500€/month
1€/hour
Training 8.5 hours 18 minutes
Cost 0.66€ 0.30€
Duration & cost for « frontalization » example