MISANU

PROGRAM

Plan rada Seminara iz veštačke inteligencije za APRIL 2023.

Registraciona forma za učesće, i link na predavanje ako ste već registrovani:
https://miteam.mi.sanu.ac.rs/asset/CW5nJWDSEZDj7p32p
Ukoliko želite samo da gledate predavanje bez mogućnosti aktivnog učešća, prenos će biti dostupan na:
https://miteam.mi.sanu.ac.rs/asset/4LNW8WtML7rLKojoz
Na ovom linku se mogu pronaci kratka uputstva na srpskom i engleskom:
https://miteam.mi.sanu.ac.rs/asset/Kc7qJtEvoMFx9MFnz

SREDA, 05.04.2023. u 19:00, Online
Max Talanov, Institute for Artificial Intelligence and Kazan Federal University
MEMRISTIVE IMPLEMENTATION OF THE SPIKING NEURAL NETWORK OF THE SPINAL CENTRAL PATTERN GENERATOR
Currently more than a billion people around the world have neurological conditions. In other simple words they have problems with their neurons. We know that neural networks do their good job in a lot of fields like natural language processing and computer vision. Can we use their neurons to help those people? The strait forward answer is NO, for at least one reason the Rosenblatt model of neuron is not bio-plausible. The good news are we know how the bio-plausible model should work and more than that there are more than thousand bio-plausible models of neurons in the model db - in Yale university. The project i want to present is the memristive (hardware) implementation of the spinal central pattern generator neuronal circuitry. Only four neurons with the self-learning option (STDP) indicated the power of the self-organization after the learning process to create the neuronal walking pattern replicating the biological results.

SREDA, 12.04.2023. u 19:00, Online
Miloš Košprdić, Nikola Milošević, Bayer, Istraživačko-razvojni institut za veštačku inteligenciju Srbije
PREPOZNAVANjE BIOMEDICINSKIH IMENOVANIH ENTITETA BEZ I SA MALO PRIMERA
Kod prepoznavanja imenovanih entiteta u specifičnim domenima kao što je biomedicinski često je potrebno dodati novi tip imenovanog entiteta. Međutim, kreiranje podataka sa primerima za mašinsko učenje je često skup, dugotrajan, i dosadan posao. U ovoj prezentaciji ćemo predstaviti algoritam mašinskog učenja zasnovan na transformer modelima koji je u stanju da uspešno prepozna nove imenovane entitete bez zadatih primera tog entiteta u toku učenja (zero-shot learning). Takođe, razvili smo i metodu koja omogućava fino podešavanje modela na osnovu samo nekoliko primera. Fino podešavanje uz pomoć samo jednog ili 10 primera daje rezultate koji su gotovo uporedivi sa treniranjem modela na celom skupu od nekoliko hiljada primera. Ovaj metod je plod saradnje farmaceutske kompanije Bayer i Istraživačko-razvojnog instituta za veštačku inteligenciju Srbije.

SREDA, 19.04.2023. u 19:00, Online
Joao Paulo Carvalho, INESC-ID / Instituto Superior Técnico, Universidade de Lisboa, Portugal
FUZZY FINGERPRINTS: FROM USER IDENTIFICATION TO LARGE LANGUAGE MODELS
Fuzzy Fingerprints (FFP) are a frequency-based compact classification technique inspired by human fingerprints. They usually perform and compare well against other machine learning techniques when the number of classes is very large. FFPs have been used and adapted for tasks such as mobile phone user identification, web user identification, text authorship attribution, Tweet Topic Detection, prediction of Intensive Care Unit readmissions from medical text notes, Memory-based Collaborative filtering solutions in the Recommendation Systems domain, or cyberbullying detection in social networks.
FFPs have been successfully used as an interpretable text classification technique, but, like most other techniques, have lately been largely surpassed in performance by Large Pre-trained Language Models (LLM), such as BERT or RoBERTa. However, these LLM suffer from the lack of interpretability and explainability. Recently we were able to combine the interpretability and compact characteristics of the FFP framework with the robustness of the large pre-trained models and shown that, even with a small FFP size, this new architecture can generalize and compete with the results from fine-tuned LLM models.

SREDA, 26.04.2022. u 19:00, Online
Laszlo T. Koczy, Szechenyi Istvan University and Budapest University of Technology and Economics, Hungary
DISCRETE BACTERIAL MEMETIC EVOLUTIONARY ALGORITHMS FOR SOLVING HIGH COMPLEXITY PROBLEMS
Evolutionary algorithms attempt to copy the solutions nature offers for solving (in the quasi-optimal sense) intractable problems, whose exact mathematical solution is impossible. The prototype of such algorithms is the Genetic Algorithm, which is, however rather slow and often does not find a sufficient solution. Nawa and Furuhashi proposed a more efficient modified one, under the name of Bacterial Evolutionary Algorithm (BEA). Moscato proposed the combination of evolutionary global search with nested local search based on traditional optimization techniques, and called the new approach memetic algorithm (MA).
Our group started to combine BEA with Levenberg-Marquardt local search and we obtained very good results on a series of benchmarks. The next step was to apply the new type of MA for NP-hard discrete optimization, starting with the classic and well known Traveling Salesman Problem (TSP), applying discrete local search, and thus proposing the novel Discrete Bacterial Memetic Evolutionary Algorithm (DBMEA). Then, we continued with a series of related, but mathematically different graph search problems, applying the same approach. Although we could not improve the tailor made Helsgaun-Lin-Kernighan (HLK) heuristics for the basic TSP, we got comparably good results, and in some other problem cases, we obtained new, so far the best accuracy and running time combinations. The Traveling Repairman Problem is an eminent example, where DBMEA delivers the best solutions.
The advantages of the new approach are as follows:

General applicability. With minimal adaptation to the concrete problem type the same method could be successfully applied, there was no need to construct new tailor made algorithms for every new problem
Predictability. Knowing the problem size, it was easy to give a good estimation of the running time, assuming a certain accuracy. This is not true for any of the other approaches, including the HLK, and especially not true for other methods, finding approximate solutions (often with large error)

In the talk, several examples will be presented with standard benchmarks going up to large numbers of graph nodes, and the DBMEA results will be compared with the best practices from the literature. The predictability feature will also be illustrated by size-running time graphs.
Reference will be made to the importance of determining the initial population in achieving fast and accurate results. A new approach, the Bounded Radius Heuristics will be presented.
In the last part of the talk, a series of fuzzy extensions of the Time Dependent TSP (TD TSP) will be introduced, an extension of the TSP with real life aspects where the natural fluctuation of the traffic in certain areas causes non-deterministic features causing additional difficulties in the quasi-optimization. The novel extensions will be also tackled with the DBMEA approach successfully.
As a conclusion, one more example will be mentioned where the discrete NP-hard problem is of a rther different nature, and it will be shown that by changing the local search technique appropriately, DBMEA can still deliver superior results.

Ovaj onlajn seminar nastao je kao nastavak sastanka “Serbian AI Meeting” i zamišljen je da na njemu istraživači iz Srbije i iz dijaspore, kao i istraživači sa univerzteta, naučnih instituta i iz prakse predstavljaju naučne teme i rezultate iz oblasti veštačke inteligencije.
Link za svako pojedinačno predavanje biće dostavljen dan pre održavanja predavanja.

Andreja Tepavčević
Rukovodilac seminara