Tuesday, January 9, 2024

// // Leave a Comment

Skoltech - Skolkovo Institute of Science and Technology - Neural Optimal Transport, the convergence of different AI paths and comparison of research institutes missions

An interesting research institute in Russia which I discovered while reviewing a paper of theirs about optimal transport used to create maps between images, similar to Wasserstein GANs.

Published as a conference paper at ICLR 2023
NEURAL OPTIMAL TRANSPORT, Alexander Korotin, Daniil Selikhanovych, Evgeny Burnaev https://arxiv.org/pdf/2201.12220.pdf   
We present a novel neural-networks-based algorithm to compute optimal transport maps and plans for strong and weak transport costs. To justify the usage of neural networks, we prove that they are universal approximators of transport plans between probability distributions. We evaluate the performance of our optimal transport algorithm on toy examples and on the unpaired image-to-image translation

As of the optimal transport - as discussed in the darker ages of AGI, IMO at a high level many  or  all working approaches are actually analogical, homeomorphic, isomorphic and convergent. The same problems are solved with different terminology and formulation. Some minimize the "earth mover's distance (Wasserstein distance), other "the energy" or "the free energy", or the "cost", or find the maximum "reward" (which is functionally the same), shortest path. An underground AGI-er which I know is also framing the problem as a logisitcs one, transporting "items". Numenta's/Hawkin's "frames of reference", Levin's/Field's "navigating different spaces" (which is also a direct consequence of just that the Universe is a Computer, and the basic cognitive primitives of time, space, causality; see also :"Embodiment is just coordinate spaces, interactivity and modalities - not a mystery"   https://artificial-mind.blogspot.com/2011/12/embodiment-is-just-coordinate-spaces.html 

Deep learning in general is doing that, minimizing "the loss function", the difference, maximizing match, and the sequences of activations could be viewed as "paths". Another interpretation, which is similar, is Clustering (CogAlg, also Theory of Universe and Mind, which encompasses many views). It is all "variational computation", "optimization" and mapping (matching) etc., within the hierarchical prediction-causation, and the actual problem is the definition of adequate configurations, core representations of the problem spaces, the space of the development, the possible actions and measurements etc. The rest are the technical details of the "optimization", the traversal of these spaces, which is search and match. Etc. The huge datasets are probably the bigger part of these spaces and that's one reason why the DL critics are blaming them for being "just big DBs", "hash tables" etc. See "Unvierse and Mind 6" when it is published with notes about why they are not "just...": https://github.com/Twenkid/Theory-of-Universe-and-Mind/blob/main/Universe-And-Mind-6.md 

Now the "embeddings" or "vectors" are the preferred terms of what was formats, representations, records, general "types" of the data.  

That reminds me of a short rant of B.K. where he complains about the underground AGI developers, "hackers", who claimed they knew the "secret of AGI", but they kept it proprietary etc., on Twitter. They could have only tricks, but they couldn't explain it etc.

Well, IMO there's no secret at the broad theoretical level. AGI was conceptually explained and clear as of what is required to be performed, to achieve computationally and what to "optimize" in the early 2000s (prediction-compression, hierarchy, incremental precision and range, multimodality, intermodality, "creativity is imitation at the level of algorithms" etc.), e.g. in the works/ideas which I try to get credit for being rediscovered and now praised by academic researchers - The Theory of Universe and Mind. The secret could be in efficiency, as the "cranks" and the ones with less resources are required to be more clever, but with more or less resources, if a system is demonstrating AGI and is able to communicate, to produce comparable patterns, to solve corresponding problems etc.as another AGI/human/cognitive system, that implies that they have some isomorphic structures and representations at some level or some way of reviewing/observing/measuring them. Etc.


Back to Skoltech:

The institute is interdisciplinary and has diverse research directions, not only AI, and it's founded in 2011 in a collaboration with MIT.

Skolkovo Institute of Science and Technology (Skoltech) in Moscow is a new model university in Russia, established with the vision of being a world-leading institute of science and technology. Skoltech mission is to impact economy and society development based on academic and technology excellence and entrepreneurial spirit. Integrating entrepreneurship and innovation, Skoltech delivers graduate educational programs to shape next generations of leaders in science, technology and business. Skoltech is recognized among top-100 world young universities in Nature Index ranking, taking # 65 place. https://360.skoltech.ru
A summary of the AI department: https://crei.skoltech.ru/ai 

 "Find out more about Skoltech AI research groups:
Computational Intelligence, Prof. Ivan Oseledets
Mobile Robotics, Prof. Gonzalo Ferrer
Natural Language Processing, Prof. Alexander Panchenko
Intelligent Signal and Image Processing, Prof. Anh Huy Phan
Multiscale Neurodynamics for Intelligent Systems, Prof. Jun Wang
Mathematical Foundations of AI, Prof. Dmitry Yarotsky
AI & Supercomputing, Prof. Sergey Rykovanov
Quantum algorithms for machine learning and optimisation, Prof. Vladimir Palyulin
Computational Imaging, Prof. Dmitry Dylov
AI for Materials Design, Prof. Alexander Shapeev
AI-driven Modeling, Prof. Ekaterina Muravleva
Parallel algorithms for AI, Prof. Alexander Mikhalev
Tensor Networks & Deep Learning, Prof. Andrzej Cichocki"
The institute INSAIT in Sofia, Bulgaria, which opened in 2022, has a similar philosophy. It is currently  more narrower (Computer Science, Math, AI, Institute for AI and Technology) and was also declared as created in a tight collaboration with established institutes: EPFL and ETH, and working as a department of Sofia University, was presented and is still advertised as the first of its kind in Eastern Europe, etc. https://insait.ai/what-is-insait/ 

INSAIT’s mission is to establish a first-of-its-kind research institute for computer science and artificial intelligence in Eastern Europe with sole focus on scientific excellence. INSAIT’s faculty and staff will conduct world-class research, attract outstanding international scientists, and training the next-generation of graduate and undergraduate students.

INSAIT is expected to have transformational effects on society and economy at large: attracting high-quality diverse talent to the region, preventing brain drain, creating new state-of-the-art educational programs, pushing towards a more product-driven economy by inventing high-valued intellectual property (IP), enabling deep research-guided technological companies, attracting big technology companies, and many more.

Compare to the Sacred Computer's Mission and Strategy from 2003, 21 years ago:


(Originally in Bulgarian, Translated by GT)

How would I invest a million with the greatest benefit for the development of the country? - Interdisciplinary research and creative institute 

"(...)The science of Artificial Intelligence (AI) is "middle aged" and some of its "fathers" are also "fathers" of the computer age. As soon as they "gave birth" to programmable computing machines, the "fathers" began to dream of the moment when the calculator would become a thinker... (...)

I believe that it is only a matter of years until we find the right "parts" and build a "machine" from them to overcome the "Wall" and meet the Machine.


According to my strategy, a scientific-research Institute would be founded, which would unite computer scientists, engineers, art critics, linguists, philosophers, psychologists, neuroscientists; multilingual translators; creators in various arts - writers and poets, composers and musicians; artists, photographers and film directors. The members of the Institute will be, with advantage, having knowledge and skills in several fields, both scientists and creators, because the goal of the searches will be to discover the commonality between all manifestations of reason, between the sciences and the arts. The form of thought is different in different manifestations of thinking, but its essence, the underlying mechanisms, are the same, and only the data with which it works - word, sound, images, sequences of images, abstract concepts and etc.

The institute will also play the role of a "wing" that finds, "protects and wings" gifted people to support their development and, if they wish, to enjoy their talent in research.

The Institute will have a program house, in which "incidentally" "smart" application software will be produced, using the developments of the Institute on the way to IR: programs for automated design, multimedia, word processing, translators, games, etc. application programs.

The goal of the Institute will be the programmatic creation of a MM, possessing universal capabilities for exchanging information with other computing machines, in particular robotic modules. The robots created by the robotics department will be, in addition to a way to use IR for physical activities, a means of attracting public attention and advertising for the Institute.

Once a Thinking Machine is realized, it will be able to be used in all creative spheres of human activity and in the work of the Institute itself.

I suppose that after the Discovery and the creation of the MM, running on standard computers, the Institute will be "armed" and will be able to set aside a design department to develop new complete computing systems specially adapted for the operation of the Machine.(...)"

0 коментара: