Przemysław Uznański

Adrian Kosowski, Przemysław Uznański, Jan Chorowski et al.

The relationship between computing systems and the brain has served as motivation for pioneering theoreticians since John von Neumann and Alan Turing. Uniform, scale-free biological networks, such as the brain, have powerful properties, including generalizing over time, which is the main barrier for Machine Learning on the path to Universal Reasoning Models. We introduce `Dragon Hatchling' (BDH), a new Large Language Model architecture based on a scale-free biologically inspired network of \$n\$ locally-interacting neuron particles. BDH couples strong theoretical foundations and inherent interpretability without sacrificing Transformer-like performance. BDH is a practical, performant state-of-the-art attention-based state space sequence learning architecture. In addition to being a graph model, BDH admits a GPU-friendly formulation. It exhibits Transformer-like scaling laws: empirically BDH rivals GPT2 performance on language and translation tasks, at the same number of parameters (10M to 1B), for the same training data. BDH can be represented as a brain model. The working memory of BDH during inference entirely relies on synaptic plasticity with Hebbian learning using spiking neurons. We confirm empirically that specific, individual synapses strengthen connection whenever BDH hears or reasons about a specific concept while processing language inputs. The neuron interaction network of BDH is a graph of high modularity with heavy-tailed degree distribution. The BDH model is biologically plausible, explaining one possible mechanism which human neurons could use to achieve speech. BDH is designed for interpretability. Activation vectors of BDH are sparse and positive. We demonstrate monosemanticity in BDH on language tasks. Interpretability of state, which goes beyond interpretability of neurons and model parameters, is an inherent feature of the BDH architecture.

Dariusz Dereniowski, Stefan Tiegel, Przemysław Uznański et al.

We consider the problem of searching for an unknown target vertex $t$ in a (possibly edge-weighted) graph. Each \emph{vertex-query} points to a vertex $v$ and the response either admits $v$ is the target or provides any neighbor $s\not=v$ that lies on a shortest path from $v$ to $t$. This model has been introduced for trees by Onak and Parys [FOCS 2006] and for general graphs by Emamjomeh-Zadeh et al. [STOC 2016]. In the latter, the authors provide algorithms for the error-less case and for the independent noise model (where each query independently receives an erroneous answer with known probability $p<1/2$ and a correct one with probability $1-p$). We study this problem in both adversarial errors and independent noise models. First, we show an algorithm that needs $\frac{\log_2 n}{1 - H(r)}$ queries against \emph{adversarial} errors, where adversary is bounded with its rate of errors by a known constant $r<1/2$. Our algorithm is in fact a simplification of previous work, and our refinement lies in invoking amortization argument. We then show that our algorithm coupled with Chernoff bound argument leads to an algorithm for independent noise that is simpler and with a query complexity that is both simpler and asymptotically better to one of Emamjomeh-Zadeh et al. [STOC 2016]. Our approach has a wide range of applications. First, it improves and simplifies Robust Interactive Learning framework proposed by Emamjomeh-Zadeh et al. [NIPS 2017]. Secondly, performing analogous analysis for \emph{edge-queries} (where query to edge $e$ returns its endpoint that is closer to target) we actually recover (as a special case) noisy binary search algorithm that is asymptotically optimal, matching the complexity of Feige et al. [SIAM J. Comput. 1994]. Thirdly, we improve and simplify upon existing algorithm for searching of \emph{unbounded} domains due to Aslam and Dhagat [STOC 1991].