Hoy en día es difícil no hablar de la Inteligencia Artificial y pensar en cómo se ha aplicado para resolver tareas difíciles y repetitivas para el ser humano. Pero en los últimos años, gracias a la llegada de las Redes Generativas Adversariales (GANs), la IA adoptó capacidades creativas que le permiten generar información artificial. Es la era de los Deepfakes, en la que puedes poner tu cara al actor de tu película favorita o ser felicitado por el presidente de los Estados Unidos. En esta charla, veremos gran parte de estas capacidades adquiridas por la IA, algunos ejemplos, y pondremos a prueba nuestro ojo para comprobar si estamos preparados para detectar que es real y que no.
4. How do you define 'real'? If you're talking about
what you can feel, what you can smell, what you
can taste and see, then 'real' is simply electrical
signals interpreted by your brain.
MorpheusWhat about AI and its own existence?
5.
6.
7. A computer would deserve to be called
intelligent if it could deceive a human
into believing that it was human.
Alan Turing
8. What is Artificial
Intelligence?
• More than simply code
• Solve hard human tasks
• Capacity to learn from the environment
• Intelligent Agents
Features
Any system that interacts with the environment, perceives
it, learn from it, and takes actions to achieve successfully
its goals and tasks through flexible adaptation.
13. Why using AI now?
Challenges
Reducing Costs
Escalating demand
Optimization
Usability
Interactions
Data Increase
Increase of
unstructured data
Organize data
Sensor information
Increase articles
Technology
Maturity
Previous effort
Improved
algorithms
Specialized HW
Cloud platforms
General Access
Democratization
Open source code
Frameworks
Cognitive services
Entrepreneurship
Investment
Specialization
Innovation
Business
transformation
16. Current AI Focus
• Narrow or Weak AI
• Only for solving specific problems
• Problems can be decomposed
• Complex models based on
• Model composition
• Deeper structures
• Towards a more general AI
• Models that can solve more than one task
• Making decisions
• Evolve from their response and environment
Which feature is necessary?
19. D
Discriminator
How can we create our own fashion?
G
Generator
Latent
space
Noise (s)
Real
Samples
𝑧(𝑥, 𝑠)𝑥
ො𝑢 = 𝐺(𝑧)
𝑦
𝐿 𝐺𝐴𝑁 𝑙𝑜𝑠𝑠
Is D
correct?
Fine Tune Training
𝐿1(𝑦, ො𝑢)
𝐷(𝑦, 𝑧)
Generate fake samples to fool the discriminator
Classify fake images vs real images
39. FaceSwap: Facial Modification
Training
Reconstructed A
Encoder Decoder A
Latent face A
Latent face B
Encoder Decoder B
Original A
Original B Reconstructed B
Generation
Encoder Decoder A
Latent face A
Latent face B
Encoder Decoder B
Original A
Original B Reconstructed B
From Face A
Reconstructed A
From Face B
45. What about generating
text?
• OpenAI project
• Based on Transformer approach
• Training on larger datasets:
• books, webs…
• Increases:
• Reading comprehension
• Translation
• Summarization
• QA
• Entire code not publish → Model with 1.5b words
GPT-2
46.
47. Conclusions
• AI is key to
• Automatize processes
• Create new business models
• It is here to help!!!
• Decision making support
• Towards a Generalized AI
• New features → Creativity
• Explaining what is inside
• Harder, better, faster, stronger
• Get ethics inside!!!
• The risks of fakes
• Images rights
• Biased data
• DeepFake forensics → https://github.com/ondyari/FaceForensics
49. Thanks and …
See you soon!
Thanks also to the sponsors.
Without whom this would not have been posible.
O R G A N I Z A T I O N
P L A T I N U M S P O N S O R S
C O L L A B O R A T O R S