This presentation provides a comprehensive overview on ALMA, the EU-funded project aimed at leveraging AML properties to develop a new generation of interactive, human-centric machine learning systems.
The presentation provides a deep overview of the whole project, covering from the basics on Algebraic Machine Learning (AML) technology to the specifics of the ALMA project.
1. May 19, 2022
This project has received funding from the European Union's Horizon 2020
research and innovation programme under grant agreement No 952091.
This project has received funding from the European Union's Horizon 2020
research and innovation programme under grant agreement No 952091.
HUMAN CENTRIC ALGEBRAIC MACHINE LEARNING
AML & ALMA Project overview
Presenters:
● Fernando Martin-Maroto (ALG)
● Raúl Sánchez-Mateos Lizano (EPROS)
June 1, 2023
4. | | 4
AML & ALMA Project Overview - Algebraic Machine Learning (AML)
June 1, 2023 H2020-EIC-FETPROACT-2019
AML - The new technological direction
Research in a new Machine Learning paradigm based on Algebra
ALGEBRAIC MACHINE LEARNING
Machine Learning from semantic embeddings of data and
formal knowledge into discrete algebraic structures
5. | | 5
AML & ALMA Project Overview - Algebraic Machine Learning (AML)
June 1, 2023 H2020-EIC-FETPROACT-2019
AML - What is it?
AML vs Traditional approaches to machine learning:
Symbolic AI:
⚙ Description of the world using formulas
⚙ Difficulty learning from data
⚙ High transparency
Statistical Learning (including Neural Networks)
⚙ Learning from data
⚙ Difficulty using formal descriptions
⚙ Usually opaque
Algebraic Machine Learning:
⚙ Description of the world using formulas
⚙ Learning from data
⚙ Can combine data and formal descriptions
⚙ High transparency
6. | | 6
AML & ALMA Project Overview - Algebraic Machine Learning (AML)
June 1, 2023 H2020-EIC-FETPROACT-2019
AML - What is it?
AML vs Traditional approaches to machine learning:
Symbolic AI:
⚙ Uses symbols
⚙ Symbols represent real world objects
⚙ Mostly uses discrete mathematics
⚙ Symbols are permanent
Statistical Learning (including Neural Networks)
⚙ Uses parameters
⚙ Parameters can map to the world or to intermediate
internal descriptions
⚙ Mostly uses continuous mathematics
⚙ Parameters can change
Algebraic Machine Learning:
⚙ Uses symbols
⚙ Symbols can map to the world (constants) or to
intermediate internal descriptions (atoms)
⚙ Uses discrete mathematics
⚙ Can create new symbols
⚙ Symbols can change
7. | | 7
AML & ALMA Project Overview - Algebraic Machine Learning (AML)
June 1, 2023 H2020-EIC-FETPROACT-2019
AML - What is it?
AML vs Traditional approaches to machine learning: underlying principles
Symbolic AI:
⚙ Satisfiability,
⚙ Logic, deduction, inference
Statistical Learning (including Neural Networks)
⚙ Error minimization
⚙ Fitting
⚙ Statistical inference
Algebraic Machine Learning:
⚙ Indecomposability
⚙ Maximization of algebraic freedom
⚙ Small size
⚙ Stability of indecomposable components
8. | | 8
AML & ALMA Project Overview - Algebraic Machine Learning (AML)
June 1, 2023 H2020-EIC-FETPROACT-2019
AML - What is it?
Traditional approaches to machine learning
DATA
FORMAL KNOWLEDGE
TRAINING DATA
EMBEDDING IN AN
ALGEBRAIC THEORY
AML ENGINE
= ….
Φ Φ Φ Φ Φ
symbols
defined by the
user
symbols
generated by
the engine
9. | | 9
AML & ALMA Project Overview - Algebraic Machine Learning (AML)
June 1, 2023 H2020-EIC-FETPROACT-2019
AML - Proven Cases
Proven cases of Algebraic Machine Learning
Learning a maze
Queens
completion problem
Handwritten
digit recognition
- Supervised learning (MNIST)
- Atoms: algebraic elements
resulting from learning
- Learning from formal knowledge
- Rules encoded in the algebra
- AML understands the game from
the beginning of learning process
- Learning from formal knowledge
- The path concept and geometry are
encoded in the algebra
10. | | 10
AML & ALMA Project Overview - Algebraic Machine Learning (AML)
June 1, 2023 H2020-EIC-FETPROACT-2019
AML - Proven Cases
Proven cases of Algebraic Machine Learning
Classification of human
motion
Finding hamiltonian
paths
Resolving and
creating sudokus
- Learning from formal knowledge
to solve sudoku games
- Rules are encoded in the algebra
- Inventing new sudoku games
- Finding a hamiltonian path from a
description of the tasks.
- More efficient than naive
backtracking methods
- Learning from both formal knowledge
and data (OPPORTUNITY)
- Activity recognition.
11. | | 11
AML & ALMA Project Overview - Algebraic Machine Learning (AML)
June 1, 2023 H2020-EIC-FETPROACT-2019
AML - Ongoing case studies
Ongoing Algebraic Machine Learning applications
World models
Robot control and
path planning
Interaction with
gesture keyboard
- Gesture keyboard
- Confidence feedback interface
- Robot control and landscape
navigation
- Formal description of high level
real-world concepts
- Ethical aspects
12. | | 12
AML & ALMA Project Overview - Algebraic Machine Learning (AML)
June 1, 2023 H2020-EIC-FETPROACT-2019
AML - Features
Research in a new Machine Learning paradigm based on Algebra
Less sensitive to
statistical features
of training data
No tradeoff between
memorization and
learning (no overfitting)
High mathematical
transparency
Large-scale
distributed learning
Interactive ML
13. | | 13
AML & ALMA Project Overview - Algebraic Machine Learning (AML)
June 1, 2023 H2020-EIC-FETPROACT-2019
AML - Features
Unique features of Algebraic Machine Learning
Less sensitive to
statistical features
of training data
High mathematical
transparency
Distributed ML
ecosystem
Interactive ML
No tradeoff between
memorization and
learning (no overfitting)
14. | | 14
AML & ALMA Project Overview - Algebraic Machine Learning (AML)
June 1, 2023 H2020-EIC-FETPROACT-2019
Symbolic AI capable of learning from:
⚙ Semantic embedding of data
⚙ Identify patterns in images (supervised learning)
⚙ Formal specification of human knowledge
⚙ Solve the N-Queen completion problem from a formal
description of the rules of the game
(unsupervised learning)
AML - Features
Research in a new Machine Learning paradigm based on Algebra
15. | | 15
AML & ALMA Project Overview - Algebraic Machine Learning (AML)
June 1, 2023 H2020-EIC-FETPROACT-2019
AML - Features
Unique features of Algebraic Machine Learning
Less sensitive to
statistical features
of training data
High mathematical
transparency
Distributed ML
ecosystem
Interactive ML
No tradeoff between
memorization and
learning (no overfitting)
16. | | 16
AML & ALMA Project Overview - Algebraic Machine Learning (AML)
June 1, 2023 H2020-EIC-FETPROACT-2019
AML - Learning and memorization
Unique features of Algebraic Machine Learning
17. | | 17
AML & ALMA Project Overview - Algebraic Machine Learning (AML)
June 1, 2023 H2020-EIC-FETPROACT-2019
AML - Features
Unique features of Algebraic Machine Learning
Less sensitive to
statistical features
of training data
High mathematical
transparency
Distributed ML
ecosystem
Interactive ML
No tradeoff between
memorization and
learning (no overfitting)
18. | | 18
AML & ALMA Project Overview - Algebraic Machine Learning (AML)
June 1, 2023 H2020-EIC-FETPROACT-2019
AML - Features
Unique features of Algebraic Machine Learning
Less sensitive to
statistical features
of training data
High mathematical
transparency
Distributed ML
ecosystem
Interactive ML
No tradeoff between
memorization and
learning (no overfitting)
19. | | 19
AML & ALMA Project Overview - Algebraic Machine Learning (AML)
June 1, 2023 H2020-EIC-FETPROACT-2019
AML - Features
Unique features of Algebraic Machine Learning
20. | | 20
AML & ALMA Project Overview - Algebraic Machine Learning (AML)
June 1, 2023 H2020-EIC-FETPROACT-2019
AML - Features
Unique features of Algebraic Machine Learning
Less sensitive to
statistical features
of training data
High mathematical
transparency
Distributed ML
ecosystem
Interactive ML
No tradeoff between
memorization and
learning (no overfitting)
21. | | 21
AML & ALMA Project Overview - Algebraic Machine Learning (AML)
June 1, 2023 H2020-EIC-FETPROACT-2019
AML - References
⚙ Method for large-scale distributed machine learning using formal knowledge and training data,
(2018) PCT application, US patent application US20190385087A1 F. Martin-Maroto
⚙ Algebraic Machine Learning. F. Martin-Maroto, & G. de Polavieja (2018). Algebraic Machine
Learning. arXiv:1803.05252.
⚙ Finite Atomized Semilattices. F. Martin-Maroto, F., & G. de Polavieja (2021). Finite Atomized
Semilattices. arXiv:2102.08050.
⚙ In-memory Processing of Algebraic Machine Learning, (2021) PCT application, US patent
application, F. Martin-Maroto, N. Abderrahaman-Elena, G. de Polavieja.
⚙ Semantic Embeddings in Semilattices. F. Martin-Maroto & G. de Polavieja. (2022).
Publicly available documents
23. | | 23
AML & ALMA Project Overview - General Overview
June 1, 2023 H2020-EIC-FETPROACT-2019
General Overview
AML - a new generation of interactive human-centric learning systems
Training &
Decision Making
Process
Machine decisions can be challenged,
interpreted, refined and adjusted.
Mutual exchange, introspection and active
learning of both system and user.
User introspection
Explore models beyond the state.of-the-art
offline and centralised data processing.
Pursue new avenues, such as incremental,
unsupervised, active, one-shot and ‘small data’
ML.
Machine learning
24. | | 24
AML & ALMA Project Overview - General Overview
June 1, 2023 H2020-EIC-FETPROACT-2019
General Overview
Objectives of ALMA project
Models, ethics and culture with AML
Dissemination of AML
Use cases
4
5
6
1
2
3
Foundations of AML
Methodologies to work with AML
Computing and networking tools
Principles of generalization in
AML and combination with
other ML techniques
AML Description Language to
enhance Human-Computer
interaction
Decentralised platform to
integrate AML-based nodes and
connect them with other SWs
Represent complex human
concepts with AML
Promote the adoption of AML
Verify AML ideas and
requirements of project
developments
25. | | 25
AML & ALMA Project Overview - General Overview
June 1, 2023 H2020-EIC-FETPROACT-2019
General Overview
AML - The new technological direction
Problem
⚙ Traditional ML
⚙ High sensitivity to statistical
properties of training data
⚙ Major difficulties combining
heterogeneous knowledge
⚙ Current ML algorithms models
⚙ Difficult to understand
⚙ Statistical learning “black-boxes”
⚙ Implicit biases in decision making
Approach
⚙ AML - a new viable Artificial
Intelligence paradigm
⚙ New radical approach based on
algebraic embeddings
⚙ Next AI frontier with verifiable
features of
⚙ Explainability
⚙ Trustworthiness
⚙ Transparency
27. | | 27
AML & ALMA Project Overview - ALMA Project
June 1, 2023 H2020-EIC-FETPROACT-2019
ALMA Project Overview
1. Consortium responsibilities
2. Overall architecture definition
3. Market positioning
Table of contents
28. | | 28
AML & ALMA Project Overview - ALMA Project
June 1, 2023 H2020-EIC-FETPROACT-2019
ALMA Project Overview
1. Consortium responsibilities
2. Overall architecture definition
3. Market positioning
Table of contents
29. | | 29
AML & ALMA Project Overview - ALMA Project
June 1, 2023 H2020-EIC-FETPROACT-2019
ALMA Consortium
Proyectos y Sistemas de
Mantenimiento SL (eProsima)
German Research Center for
Artificial Intelligence
Technical Research Centre of Finland
30. | | 30
AML & ALMA Project Overview - ALMA Project
June 1, 2023 H2020-EIC-FETPROACT-2019
WP1 - PM, Architecture & Tech. coordination
Project management
● Project organisation and communication
● Reporting, financial management
● Progress monitoring and risk mitigation
Overall Architecture definition
● Coordinate scientific and technical inputs
● AML-DL and AML-IP specifications
● Complete software, interfaces, dependencies, and interactions design
Innovation management plan
Consortium responsibilities regarding WP1
1
2
3
31. | | 31
AML & ALMA Project Overview - ALMA Project
June 1, 2023 H2020-EIC-FETPROACT-2019
WP2 - Fundamentals of Interactive AML
Generalization of AML
● How well AML generalizes outside training dataset
Work with traditional ML systems
● Compare AML result with statistical learning systems
● Couple other ML techniques (deep learning) with AML
Human-AML interaction
● Test the ability of AML to learn from formal knowledge
● How AML teach the human the results
(human in the training loop)
Collective learning
● Learn from many algebras running in parallel
Consortium responsibilities regarding WP2
1
2
3
4
32. | | 32
AML & ALMA Project Overview - ALMA Project
June 1, 2023 H2020-EIC-FETPROACT-2019
WP3 - AML Description language
AML-DL specification
● Write the AML-DL specification and AML-DL interpreter
Consistency checker
● Validate algebraic instruction blocks
Debugging tools
● Tools to assist AML-DL developers
AML accelerator
● Research on SW/HW AML acceleration
Consortium responsibilities regarding WP3
1
2
3
4
33. | | 33
AML & ALMA Project Overview - ALMA Project
June 1, 2023 H2020-EIC-FETPROACT-2019
WP4 - Human AML Interaction
Cognitive foundations for Human-AML interaction
● How human learn from and control interaction with AML
Interaction paradigm methodology
● Enable AI researchers to create more effective human-computer partnerships
● Requirements for AML based interactive machine learning
Working prototype
● Demonstrate the design methodology and interaction paradigm
Evaluation methods
● Efficiency of the interaction from human perspective
Consortium responsibilities regarding WP4
1
2
3
4
34. | | 34
AML & ALMA Project Overview - ALMA Project
June 1, 2023 H2020-EIC-FETPROACT-2019
WP5 - Models, ethics and culture with AML
AML based world models
● Embed complex models into AML
Represent complex human concepts with AML
● Human centric AI
Human-AML co-creation of complex models
● AML ability to recognize complex real world situations
AML ethical and cultural concepts model
● Refinement of Human-AML co-creation of complex domain models
Consortium responsibilities regarding WP5
1
2
3
4
35. | | 35
AML & ALMA Project Overview - ALMA Project
June 1, 2023 H2020-EIC-FETPROACT-2019
WP6 - System tools
Consortium responsibilities regarding WP6
AML Integrating Platform (AML-IP)
● Interconnect AML components
● Cloud and edge computing environments
Robotics and Constrained Devices
● Extend AML-IP for ROS 2 compatibility
Open source tools for AML
experimentation
● Libraries with reusable AML algorithms for AI/ML
problems
1
2
3
36. | | 36
AML & ALMA Project Overview - ALMA Project
June 1, 2023 H2020-EIC-FETPROACT-2019
WP7 - Use cases
Image Classification using interactive AML
● Study image classification improvements with AML
Intelligent Tools for supporting creative professionals
● Provide support for cultural, gender and related issues
Higher-level cognition for domestic assistance robots
● Encode domestic tasks using AML-DL
Consortium responsibilities regarding WP7
1
2
3
37. | | 37
AML & ALMA Project Overview - ALMA Project
June 1, 2023 H2020-EIC-FETPROACT-2019
WP8 - Dissemination, Exploitation and
Collaboration
Communication and dissemination strategy
● Publish results
● Participation in events and workshop organization
Collaboration with other projects and initiatives
● Collaborate with ROS and FIWARE to build open source tools
● Collaborate with AI/ML european communities
Exploitation and Outlook Plan
● Short and long term plans for AML dissemination and exploitation
● Alignment with the European Research Agenda for AI
Consortium responsibilities regarding WP8
1
2
3
38. | | 38
AML & ALMA Project Overview - ALMA Project
June 1, 2023 H2020-EIC-FETPROACT-2019
ALMA Project Overview
1. Consortium responsibilities
2. Overall architecture definition
3. Market positioning
Table of contents
39. | | 39
AML & ALMA Project Overview - ALMA Project
June 1, 2023 H2020-EIC-FETPROACT-2019
Overall architecture definition
ALMA Architecture and WP dependencies
40. | | 40
AML & ALMA Project Overview - ALMA Project
June 1, 2023 H2020-EIC-FETPROACT-2019
ALMA Project Overview
1. Consortium responsibilities
2. Overall architecture definition
3. Market positioning
Table of contents
41. | | 41
AML & ALMA Project Overview - ALMA Project
June 1, 2023 H2020-EIC-FETPROACT-2019
ALMA Mission
To provide a new ML paradigm, known as
AML
⚙ Easily understandable by (no black-box)
⚙ Ease of interaction (Human-AML interaction)
⚙ Seamless integration with AML-IP
⚙ Ensure long-term maintenance of AML environments
Spain
Portugal
Germany
France Finland
42. | | 42
AML & ALMA Project Overview - ALMA Project
June 1, 2023 H2020-EIC-FETPROACT-2019
USP - Value Proposition
Core message and brand promise of ALMA project
Controllable transparent, distributed and non-centralized machine
learning. An algebraic approach to ML that can complement statistical
methods.
Brand promise
AML: The novel Machine Learning paradigm leveraging abstract algebra
for better control and more transparent AI.
Core message
43. | | 43
AML & ALMA Project Overview - ALMA Project
June 1, 2023 H2020-EIC-FETPROACT-2019
Market positioning
Community impact and engagement
● alma-ai.eu
● eprosima.com/products-all/r-d-projects/eu-project-alma
● github.com/eProsima/AML-IP
44. | |
alma-ai.eu
44
AML & ALMA Project Overview
June 1, 2023
alma@eprosima.com alma-ai.eu
H2020-EIC-FETPROACT-2019