YerevaNN

Join us on February 13 at YandexHall for a presentation on Scaling Laws. Attendance is free, but registration is required. Find all the details in the event description!

Hrant Khachatrian from YerevaNN discusses the latest advancements in AI with Amalya Gabrielyan of News.am. [Video is in Armenian]

???10/10 Our Donors & Partners With this final summary post of our activities in 2024, we would like to express our heartfelt gratitude to our friends and supporters. This year, we provided five named fellowships with generous support from our industry sponsors: Yandex Armenia, Layerswap, and Fimetech. We are happy to establish a new tradition of supporting researchers with fellowships from the industry, and we are grateful to these companies for becoming the pioneers. Our research initiatives and presence at international conferences were supported by Technotun Club, ProfHolod, Formula VC, the H. Hovnanian Family Foundation, Nuve, and Mayro. Nebius.ai helped us enhance our GPU infrastructure with their exceptional cloud service. Recently, we made it easier for everyone to contribute to our activities. We have established a campaign page on the reArmenia platform with support for monthly recurring donations (tax-deductible in the US). For other ways to help, please visit our website at yerevann.com. On behalf of YerevaNN team, we wish you a prosperous and innovative year ahead! Let 2025 bring exciting opportunities for growth, discoveries, and success in your endeavors. We are enthusiastic about strengthening our collaboration and working together to advance the frontiers of research. Here's to a year of meaningful achievements and shared success!

?? 9/10 Internship Program: Supported by Voskanyan and Cognaize Scholarships Recent developments in AI research have opened vast opportunities. We strongly believe it is possible to solve many pressing issues facing humanity by further pushing the frontier of AI in directions beyond plain text, with more focus on biology, chemistry, material science, and other fields. To achieve this, we need an increasing number of talented people advancing scalable architectures and algorithms and addressing domain-specific challenges. We are committed to bringing new students into impactful AI research activities. In 2024, we hired five interns: Anna Khosrovyan, Filya Geikyan, Armen Manukyan, Narek Nurijanyan, and Davit Gyulnazaryan. The internship program was supported by Voskanyan Scholarships (thanks to Hrant Davtyan, PhD and friends) and Cognaize Scholarships (thanks to Vahe Andonians and Al Eisaian). We are planning to launch another internship program in Summer 2025. Contact us if you would like to support our efforts in scaling these capacity-building activities.

This year was monumental in AI research globally. Over the final 10 days of 2024, we summarize our team's efforts from this year to advance modern AI, sharing one highlight per day. ??? 8/10 In-context Learning in Presence of Spurious Correlations It is well known that Transformers can perform in-context learning. They can even handle simple regression tasks in-context, which can be interpreted as *learning* a learning algorithm (instead of designing one). In this paper, we focus on classification tasks with spurious features. We demonstrate that it is possible to obtain a learner for such tasks that can even outperform baselines like regular ERM or GroupDRO. We propose a simple trick to prevent the memorization of the task in the Transformer's weights and enforce in-context learning. Additionally, we show that it is possible to transfer in-context learners to unseen tasks to a certain degree. An early version of this work was presented at the In-Context Learning Workshop at ICML ‘24. The preprint is available on arXiv: https://lnkd.in/eQ5DXNAH Hrayr Harutyunyan Samvel Karapetyan Rafayel Darbinyan Hrant Khachatrian

This year was monumental in AI research globally. Over the final 10 days of 2024, we summarize our team's efforts from this year to advance modern AI, sharing one highlight per day. ??? 7/10 Generalization of Remote Sensing Models Over the past two years there was an explosion of visual foundation models tailored to remote sensing applications. They can be fine-tuned for various kinds of downstream tasks, from scene classification to change detection. Are these fine-tuned models capable of generalizing to lower resolution imagery, or to imagery with a different set of bands? We designed a benchmark that measures these in detail. We see that none of the tested approaches generalize to bands unseen during fine-tuning (but known from pretraining). Hakob Tamazyan presented this work at the OOD-CV workshop of ECCV 2024 conference. Ani Vanyan, Tigran Galstyan, Alvard Barseghyan, Anna Khosrovyan, Vahan Huroyan, Hrant Khachatrian. Ani’s work is supported by the Fimetech Fellowship. Alvard’s research is supported by the Layerswap Fellowship. Compute resources for the largest experiments were generously provided by Nebius.ai (Anastasia Zemskova). Conference participation was sponsored by H. Hovnanian Family Foundation Travel Grant.

This year was monumental in AI research globally. Over the final 10 days of 2024, we summarize our team's efforts from this year to advance modern AI, sharing one highlight per day. ??6/10 Indoor Radio Map Prediction When you place a WiFi router somewhere in an apartment, can you predict the signal strength in other rooms? The ICASSP 2025 conference announced a challenge on indoor pathloss radio map prediction. It included three test sets designed to measure increasingly challenging generalization scenarios: generalization to new building layouts, new radio frequencies, and new antenna types. Our team participated in the challenge, focusing on heavy data augmentation methods to teach the model to generalize across various axes. Our solution demonstrated competitive generalization with respect to frequencies and antenna types but struggled with new building layouts, placing us in the 8th position on the leaderboard. We are actively working on improving that aspect of generalization as well. The method we have developed so far is described in a short paper available on arXiv: https://lnkd.in/eSyMuthY Contributors: Rafayel Mkrtchyan, Edvard Ghukasyan, Hrant Khachatrian, and Theofanis Raptis. Edvard’s work was supported by Yandex Armenia Fellowship.

This year was monumental in AI research globally. Over the final 10 days of 2024, we summarize our team's efforts from this year to advance modern AI, sharing one highlight per day. ??5/10: Environment Reconstruction Using RF Signals Imagine multiple mobile devices in an urban environment receiving signals from various cellular antennas. What information do these signals contain? What if we tried to reconstruct the environment using RF signals? Using the large-scale simulated dataset WAIR-D, we demonstrated that it is possible to recover the map with 41% IoU. Rafayel Mkrtchyan presented this work at the IEEE IWCMC 2024 conference. Read the paper on arxiv:?https://lnkd.in/ekdFf2YV However, this is a fairly artificial scenario. It is unlikely that we would have access to the locations and RF signals of multiple devices and antennas without also having a map of the region. On the other hand, it is quite plausible that we would have access to a map, but the map might be outdated or contain errors. Can we use RF signals to improve the map? The short answer is yes, but this is still a work in progress. Follow us to see the results in the coming months. We show the predicted map of buildings in the visualization below: white pixels indicate true positives, red pixels false positives, and grey pixels false negatives. Orange crosses represent the antennas, while blue crosses denote the mobile devices.

This year was monumental in AI research globally. Over the final 10 days of 2024, we will summarize our team's efforts from this year to advance modern AI, sharing one highlight per day. ??4/10: Device Localization Using RF Signals What problems can modern deep learning solve in wireless and radio communications? We began exploring this question two years ago. Rafayel Darbinyan presented the initial results on device localization in mid-2023, and this year, we published our first major paper in the Ad Hoc Networks journal. We investigated radio signal representations, adapted Vision Transformers to work with radio maps, trained models for device localization, conducted a deep analysis of prediction errors, measured robustness with respect to noise in radio signal measurements, and examined the inherent ambiguity of device localization problems. We strongly believe that wireless and radio data are going to become another important modality for the next generation of AI models, and we can't wait to see major advancements in the field in 2025. This is the first journal paper within the scope of the DISTAL project at Yerevan State University (Project RL22-054), funded by the RA Science Committee. Rafayel Mkrtchyan, Theofanis Raptis, and Hrant Khachatrian. Read the paper: https://lnkd.in/eQMUa6AN

This year was monumental in AI research globally. Over the final 10 days of 2024, we will summarize our team's efforts from this year to advance modern AI, sharing one highlight per day. ??3/10: Small Molecule Optimization with Large Language Models [Chemlactica / Chemma] Language models are remarkable. But what happens when they are combined with a proper search algorithm? What if there is an external oracle function providing feedback on the search? This project demonstrates the powerful synergy of all these components. We trained the Galactica and Gemma models on a massive corpus of small molecules (40 billion tokens!). The corpus was constructed to enable the resulting language models to understand complex prompts, such as similarity to given molecules or basic molecular properties. We integrated these models into a genetic algorithm that receives supervision signals from an external oracle function. As a cherry on top, we periodically fine-tuned the language model using the scores provided by the oracle to guide the model along the optimization trajectory in molecular space. The results are impressive: state-of-the-art performance on drug-likeness (QED) optimization (popularized by NVIDIA’s RetMol), the Practical Molecular Optimization benchmark, and several benchmarks involving protein docking simulations. Various aspects of this work were presented at the ICML ML for Life and Material Sciences Workshop and recently at the NeurIPS workshop on Foundation Models for Science (although without our physical presence, as the Canadian embassy is still processing the visa application). The preprint is available on arXiv: https://lnkd.in/eREAM5af The pretraining code and optimization algorithm are available on GitHub: https://lnkd.in/egspkaT3 The models have been downloaded more than 17,500 times on HuggingFace: https://lnkd.in/edTVFzXV Model development and pretraining of the smaller models were conducted on A100 GPUs at Yerevan State University, while the larger models were trained on H100 Cloud GPUs generously provided by Nebius AI. Philipp’s work was supported by a Yandex Armenia fellowship. Philipp Guevorguian, Menua Bedrosian, Tigran Fahradyan, Gayane Chilingaryan, Hrant Khachatrian, Armen Aghajanyan