Povilas Daniušis

Artūras Nakvosas, Povilas Daniušis, Vytas Mulevičius

In this paper, we propose and describe the first open Llama2 large language models (LLMs) for the Lithuanian language, including an accompanying question/answer (Q/A) dataset and translations of popular LLM benchmarks. We provide a brief review of open regional LLMs and detailed information on the proposed LLMs and their training process. We also conduct an empirical evaluation, comparing the perplexities of the proposed LLMs with those of other modern open LLMs. In addition, benchmarking the proposed LLMs against language understanding tasks reveals that high-quality pretraining datasets may be essential for achieving models that perform efficiently on these benchmarks. The full realisations of the described LLMs are available in the accompanying open repository~\url{https://huggingface.co/neurotechnology}.

Povilas Daniušis, Shubham Juneja, Lukas Kuzma et al.

In this paper, we focus on the problem of statistical dependence estimation using characteristic functions. We propose a statistical dependence measure, based on the maximum-norm of the difference between joint and product-marginal characteristic functions. The proposed measure can detect arbitrary statistical dependence between two random vectors of possibly different dimensions, is differentiable, and easily integrable into modern machine learning and deep learning pipelines. We also conduct experiments both with simulated and real data. Our simulations show, that the proposed method can measure statistical dependencies in high-dimensional, non-linear data, and is less affected by the curse of dimensionality, compared to the previous work in this line of research. The experiments with real data demonstrate the potential applicability of our statistical measure for two different empirical inference scenarios, showing statistically significant improvement in the performance characteristics when applied for supervised feature extraction and deep neural network regularization. In addition, we provide a link to the accompanying open-source repository https://bit.ly/3d4ch5I.

Shubham Juneja, Povilas Daniušis, Virginijus Marcinkevičius

In this article, we focus on the pre-training of visual autonomous driving agents in the context of imitation learning. Current methods often rely on a classification-based pre-training, which we hypothesise to be holding back from extending capabilities of implicit image understanding. We propose pre-training the visual encoder of a driving agent using the self-distillation with no labels (DINO) method, which relies on a self-supervised learning paradigm.% and is trained on an unrelated task. Our experiments in CARLA environment in accordance with the Leaderboard benchmark reveal that the proposed pre-training is more efficient than classification-based pre-training, and is on par with the recently proposed pre-training based on visual place recognition (VPRPre).

Vytenis Šliogeris, Povilas Daniušis, Artūras Nakvosas

In this technical report, we empirically investigate the relationship between linguistic fluency and domain knowledge in the context of continual learning with large language models (LLMs). Specifically, we enhance the linguistic fluency of the Gemma2 LLM for the Lithuanian language by autoregressively pretraining its full parameter set on the first 10\% of the Lithuanian language component of the CulturaX dataset. To prevent catastrophic forgetting of the model's existing domain knowledge, we apply Elastic Weight Consolidation (EWC), leveraging Fisher information estimated using data from the Massive Multitask Language Understanding (MMLU) benchmark. In the post-training evaluations, we assess linguistic fluency through perplexity and evaluate domain knowledge using accuracy on a suite of language understanding benchmarks, including ARC-Easy, Belebele, GSM8K, HellaSwag, MMLU, TruthfulQA, and Winogrande, in both English and Lithuanian. The empirical results demonstrate that EWC not only mitigates catastrophic forgetting by preserving the model's performance in terms of both linguistic fluency and domain knowledge but also improves or maintains these capabilities for the newly added Lithuanian language. These findings highlight the potential for more efficient adaptation of general-purpose LLMs to under-represented languages without requiring access to the original training data. The accompanying codebase is openly accessible at https://github.com/Neurotechnology/LLM_EWC.

Povilas Daniušis

We propose a simple multivariate normality test based on Kac-Bernstein's characterization, which can be conducted by utilising existing statistical independence tests for sums and differences of data samples. We also perform its empirical investigation, which reveals that for high-dimensional data, the proposed approach may be more efficient than the alternative ones. The accompanying code repository is provided at \url{https://shorturl.at/rtuy5}.