A.I.: The next frontierAmparo Alonso Betanzos
CITIC-UDCGrupo LIDIA
The Primeval Soup: The perfect storm
Batch
StreamingAman Naimat. “The new Artificial Intelligence Market. The Big data Market”. O´Reilly, 2016
During 2017 the tendency of data generation has showed sustained growth.
The appetite of corporates, industry and public sector for data driven initiatives has not decreased.
There is a change of landscape that by 2017 has started to become apparent.
Data Industry Landscape
Infrastructure Challenges
Data storage
High performance
in interchange and sharing
Data format and protocols
Advancing hardware
Regulation and Ethics
Safety
Data rich vs Data
poor
Confidentiality and
scientific transparency
Reproducibility Free data
https://www.linkedin.com/pulse/national-artificial-intelligence-research-development-nco-nitrd/
High dimensionality data
Sparse data
Heterogeneous data
Missing data
Noisy data
Adversarial data
Untrustworthy data
Data Science
• Machine Learning is as valuable as how exploitable its results are.
• Lagging behind in some areas:• Visualization of clusters• Data drift• Results Assurance• Biased data2017 Big Data Coruña. Statistical inference for big-but-biased datahttps://www.youtube.com/watch?v=luTJbX3aVKA More work
is needed on:
• Feature engineering• Regression• Anomaly detection • Practical non convex optimization• Effective parameter selection• Scalable transfer learning • Data integration• Data visualization
Reliable Machine Learning
Feature Engineering
Distributed FS algorithms
Missing Data
Heterogeneous data
Unbalanced data
Norm
aliz
ed D
isco
unte
d C
um
ula
tive G
ain
(N
DC
G)
• MNIST, 256 relevant features(576pixels)• 20% missing (MAR)• Imputation using median and SVD (Singular Value Decomposition)
B. Seijo-Pardo, A. Alonso-Betanzos, K. Bennett, V. Bolón-Canedo, I. Guyon, M. Saeed. Analysis of imputation bias for feature selection with missing data. ESANN 2018
FS Original
FS Median Imputation
FS, SVD imputation
Size matters
• The study of methodologies that increase the scalability of ML principles and algorithms.
• Scalability should be seen as an abstract concept that not only includes the case of dealing with huge amounts of data points.
• Just measuring the challenge in storage units will be a narrow minded view that will be oblivious to the challenge that current times is putting on the shoulders of ML
Networks of AI systems
Scalability
• Models that can learn under privacy and anonimity constraints
• Share parameter values, not data
• Using aggregated data• Adequate accuracy?• Private data reconstruction?
Privacy-preserving ML
D. Fernández-Francos, O. Fontenla-Romero, A. Alonso-Betanzos. One-class convex hull-based algorithm for classification in distributed environments. IEEE Transactions on Systems, Man and Cybernetics: Systems (in press)
Learning to Learn
http://bair.berkeley.edu/blog/2017/07/18/learning-to-learn/
https://spectrum.ieee.org/static/ai-vs-doctors
Narr
ow
nic
he v
s G
enera
l A
I
“Armed with machine learning, a manager becomes a supermanager, a scientist a superscientist, an engineer a superengineer. The future belongs to those who understand at a very deep level how to combine their unique expertise with what algorithms do best.” Pedro Domingos, The Master Algorithm: How the Quest for the Ultimate Learning Machine Will Remake Our World
https://www.itnonline.com/content/ new-report-highlights-five-reasons-why-radiology-needs-artificial-intelligence
Human-in-the-loop
• Deep Learning is not the AI future, https://www.kdnuggets.com/2017/08/deep-learning-not-ai-future.html
• The National AI R&D Strategic plan (USA)
https://www.linkedin.com/pulse/national-artificial-intelligence-research-development-nco-nitrd/
• General Data Protection Regulation, UE
http://ec.europa.eu/justice/data-protection/reform/files/regulation_oj_en.pdf
Explainability
Transportation
service robots
Public safety, securityAI Applications
Education
Low-resource communities
AI Applications
Entertainment
Social risk of diminishing interpersonal interactions
AI applications: Employment and workplace
The 6 Laws proposed by EUAll intelligent machine should have an emergency switch
An intelligent machine could not damage a human being
It is forbidden to establish emotional links with a machine or electronic person
The biggest machines should have an obligatory insurance
Electronic persons will have rights and obligations.
Electronic persons and machines should pay taxeshttp://www.europarl.europa.eu/news/es/news-
room/20170109STO57505/delvaux-propone-normas-europeas-para-la-rob%C3%B3tica-y-un-seguro-obligatorio
http://computerhoy.com/noticias/life/estas-son-seis-leyes-robotica-que-propone-ue-56972
6,3% (16% in Software Industry)
A.I.: The next frontierAmparo Alonso Betanzos
CITIC-UDCGrupo LIDIA