An end-to-end argument mining (AM) pipeline takes a text as input and provides its argumentative structure as output by identifying and classifying the argument units and argument relations in the text. In this work, we approach AM using fine-tuned large language models (LLMs). We model the three main sub-tasks of the AM pipeline, as well as their joint formulation, as text generation tasks. We fine-tune eight popular quantized and non-quantized LLMs – LLaMA-3, LLaMA-3.1, Gemma-2, Mistral, Phi-3, Qwen-2 – which are among the most capable open-weight models, on the benchmark PE, AbstRCT, and CDCP datasets that represent diverse data sources. Our approach achieves state-of-the-art results across all AM sub-tasks and datasets, showing significant improvements over previous benchmarks.

Attractor dynamics underlie memory mechanisms in both neuroscience and machine learning. This study explores the attractor dynamics of Boolean recurrent neural networks governed by two plasticity mechanisms: spike-timing-dependent plasticity (STDP), driven by random background activity, and global plasticity (GP), triggered by specific patterns. In the Boolean framework, attractors can be explicitly computed and quantified for small networks. We demonstrate theoretically and experimentally that Boolean networks can exhibit a vast number of attractors, each encoding a potential memory. The attractor count is strongly modulated by synaptic strengths, emphasizing the pivotal role of plasticity. While STDP predominantly reduces attractor regimes, GP significantly enhances them. Furthermore, attractor regimes can be stabilized through carefully designed input streams, even in the presence of STDP. The processes of gaining, losing, and stabilizing attractors are interpreted as mechanisms of learning, memory loss, and memory consolidation, respectively. These findings highlight the complementary roles of local and global plasticity in shaping high-capacity attractor landscapes.

Few-shot learning and high-dimensional feature selection represent significant challenges in machine learning. In this work, we introduce LPNN-FS, a novel feature selection (FS) technique based on a Lagrange Programming Neural Network (${\mathcal{P}_k}$-LPNN) originally developed in the context of compressive sampling. LPNN-FS is a continuous-time recurrent neural network whose dynamics inherently converges to an optimal sparse solution of the LASSO, with its equilibrium point acting as a feature selector. Evaluated in combination with specific downstream classifiers, our model is tested on both synthetic and real-world benchmark datasets characterized by high dimensionality and a limited number of observations. The results demonstrate that LPNN-FS competes with or outperforms state-of-the-art methods such as LASSO, k-best, and sparse PCA in this few-shot, high-dimensional setting. Overall, this study opens new avenues for leveraging compressive sampling-based techniques in challenging feature selection problems.

We introduce BERT–MINUS, a modular, feature-enriched and transfer learning enabled model for Argument Mining. BERT–MINUS consists of: 1) a joint module which embeds the paragraph text, and 2) a dedicated module, consisting of three customized BERT models, which contextualize the argument markers, argument components and additional features given as text, respectively. BERT–MINUS implements two kinds of transfer learning – auto-transfer (transfer from a task to itself) and cross-transfer (classical transfer from one task to another) – via a novel Selective Fine-tuning mechanism. BERT–MINUS achieves state-of-the-art results on the Link Identification task and competitive results on the Argument Type Classification task. The synergy between the Features as Text and Selective Fine-tuning mechanisms significantly improves the performance of the model. Our work reveals the importance and potential of transfer learning via selective fine-tuning for modular Language Models. Moreover, this study dovetails naturally into the Prompt Engineering paradigm in NLP.

Synfire rings are fundamental neural circuits capable of conveying self-sustained activities in a robust and temporally precise manner. We propose a Turing-complete paradigm for neural computation based on synfire rings. More specifically, we provide an algorithmic procedure which, for any fixed-space Turing machine, builds a corresponding Boolean neural network composed of synfire rings capable of simulating it. As a consequence, any fixed-space Turing machine with tapes of length$N$can be simulated in linear time by some Boolean neural network composed of O(N) rings and cells. The construction can naturally be extended to general Turing machines. Therefore, any Turing machine can be simulated in linear time by some Boolean neural network composed of infinitely many synfire rings. The linear time simulation relies on the possibility to mimic the behavior of the machines. In the long term, these results might contribute to the realization of biological neural computers.

We consider echo state networks (ESNs) for text classification. More specifically, we investigate the learning capabilities of ESNs with pre-trained word embedding as input features, trained on the IMDb and TREC sentiment and question classification datasets, respectively. First, we introduce a customized training paradigm for the processing of multiple input time series (the inputs texts) associated with categorical targets (their corresponding classes). For sentiment tasks, we use an additional frozen attention mechanism which is based on an external lexicon, and hence requires only negligible computational cost. Within this paradigm, ESNs can be trained in tens of seconds on a GPU. We show that ESNs significantly outperform their Ridge regression baselines provided with the same embedded features. ESNs also compete with classical Bi-LSTM networks while keeping a training time of up to 23 times faster. These results show that ESNs can be considered as robust, efficient and fast candidates for text classification tasks. Overall, this study falls within the context of light and fast-to-train models for NLP.

We consider a simplified Boolean model of the basal ganglia-thalamocortical network, and study the effect of a spike-timing-dependent plasticity (STDP) rule on the stabilization of its attractor dynamics. More precisely, we introduce an adaptive STDP rule which constantly updates its learning rate based on the attractors that the network encounters during a window of past time steps. This so-called network memory is assumed to be dynamic: its duration is step-wise increased every time a trigger input pattern is detected, and is decreased otherwise. In this context, we show that well-adjusted trigger inputs can fine tune the network memory and its associated STDP rule in such a way to drive the network into stable and rich attractor dynamics. We discuss how this feature might be related to reward learning processes in the neurobiological context.

Recent results have shown that finite state automata can be simulated by recurrent neural networks composed of synfire rings. The simulation process was shown to work correctly in the cases of Boolean neural networks and of Izhikevich spiking neural networks. In this paper, we generalize these results to the very biological context of the Hodgkin-Huxley neural model. We prove that any finite state automaton can be simulated by a Hodgkin-Huxley based recurrent neural network composed of synfire rings. In this framework, the inhibitory system ensuring the transition between the successive rings can be significantly simplified. These results show that a neuro-inspired paradigm of abstract computation based on sustained activities of neural assemblies is indeed possible, and potentially harnessable. They also constitute a first step towards the implementation of biological neural computers.

We propose a novel paradigm of neural computation based on synfire rings, i.e., synfire chains that loop back in on themselves. We show that any finite state automaton can be simulated by a Boolean recurrent neural network made up of synfire rings. More precisely, if the given automaton and its corresponding network are run in parallel on a same input stream, then the successive computational states of the automaton are perfectly reflected by the consecutive sustained activities of the network's synfire rings. Our construction turns out to be robust with respect to the removal of a number of connections. These considerations support the idea that a robust paradigm of neural computation based on sustained activities of cell assemblies is indeed possible.

We reconsider Aumann' s seminal impossibility theorem that agents cannot agree to disagree in a topologically extended epistemic model. In such a framework, a possibility result on agreeing to disagree actually ensues. More precisely, agents with a common prior belief satisfying limit knowledge instead of common knowledge of their posterior beliefs may have distinct posterior beliefs. Since limit knowledge is defined as the limit of iterated mutual knowledge, agents can thus be said to limit-agree to disagree. Besides, an example is provided in which limit knowledge coincides with Rubinstein's (1989) notion of almost common knowledge, and the agents have almost common knowledge of posteriors yet distinct posterior beliefs. More generally, an epistemic-topological foundation for almost common knowledge is thus provided.

The attractor-based complexity of a Boolean neural network is a measure which refers to the ability of the network to perform more or less complicated classification tasks of its inputs via the manifestation of meaningful or spurious attractor dynamics. Here, we study the attractor-based complexity of a Boolean model of the basal ganglia-thalamocortical network. We show that the regulation of the interactive feedback is significantly involved in the maintenance of an optimal level of complexity. We also show that the complexity of the network depends sensitively on the values of its synaptic connections. These considerations support the general rationale that the synaptic plasticity and the interactive architecture play a crucial role in the computational and dynamical capabilities of biological neural networks.

We consider a model of so-called hybrid recurrent neural networks composed with Boolean input and output cells as well as sigmoid internal cells. When subjected to some infinite binary input stream, the Boolean output cells necessarily exhibit some attractor dynamics, which is assumed to be of two possible kinds, namely either meaningful or spurious, and which underlies the arising of spatiotemporal patterns of output discharges. In this context, we show that rational-weighted neural networks are computationally equivalent to deterministic Muller Turing machines, whereas all other models of real-weighted or evolving neural networks are equivalent to each other, and strictly more powerful than deterministic Muller Turing machines. In this precise sense, the analog and evolving neural networks are super-Turing. We further provide some precise mathematical characterization of the expressive powers of all these neural models. These results constitute a generalization to the current computational context of those obtained in the cases of classical as well as interactive computations. They support the idea that recurrent neural networks represent a natural model of computation beyond the Turing limits.

We study the computational capabilities of a biologically inspired neural model where the synaptic weights, the connectivity pattern, and the number of neurons can evolve over time rather than stay static. Our study focuses on the mere concept of plasticity of the model so that the nature of the updates is assumed to be not constrained. In this context, we show that the so-called plastic recurrent neural networks (RNNs) are capable of the precise super-Turing computational power--as the static analog neural networks--irrespective of whether their synaptic weights are modeled by rational or real numbers, and moreover, irrespective of whether their patterns of plasticity are restricted to bi-valued updates or expressed by any other more general form of updating. Consequently, the incorporation of only bi-valued plastic capabilities in a basic model of RNNs suffices to break the Turing barrier and achieve the super-Turing level of computation. The consideration of more general mechanisms of architectural plasticity or of real synaptic weights does not further increase the capabilities of the networks. These results support the claim that the general mechanism of plasticity is crucially involved in the computational and dynamical capabilities of biological neural networks. They further show that the super-Turing level of computation reflects in a suitable way the capabilities of brain-like models of computation.

In classical computation, rational- and real-weighted recurrent neural networks were shown to be respectively equivalent to and strictly more powerful than the standard Turing machine model. Here, we study the computational power of recurrent neural networks in a more biologically oriented computational framework, capturing the aspects of sequential interactivity and persistence of memory. In this context, we prove that so-called interactive rational- and real-weighted neural networks show the same computational powers as interactive Turing machines and interactive Turing machines with advice, respectively. A mathematical characterization of each of these computational powers is also provided. It follows from these results that interactive real-weighted neural networks can perform uncountably many more translations of information than interactive Turing machines, making them capable of super-Turing capabilities.

The computational power of recurrent neural networks is intimately related to the nature of their synaptic weights. In particular, neural networks with static rational weights are known to be Turing equivalent, and recurrent networks with static real weights were proved to be super-Turing. Here, we study the computational power of a more biologically-oriented model where the synaptic weights can evolve rather than stay static. We prove that such evolving networks gain a super-Turing computational power, equivalent to that of static real-weighted networks, regardless of whether their synaptic weights are rational or real. These results suggest that evolution might play a crucial role in the computational capabilities of neural networks.

Recently, Miko{\l}aj Boja{\'n}czyk introduced a class of max-regular languages, an extension of regular languages of infinite words preserving manyof its usual properties. This new class can be seen as a different way of generalising the notion of regularity from finite to infinite words. This paper compares regular and max-regular languages in terms of topological complexity.It is proved that up to Wadge equivalence the classes coincide. Moreover, when restricted to $\mathbf{\Delta}^0_2$-languages, the classes contain virtually the same languages. On the other hand, separating examples of arbitrary complexity exceeding $\mathbf{\Delta}^0_2$ are constructed.

Epistemic game theory scrutinizes the relationship between knowledge, belief and choice of rational players. Here, the relationship between common knowledge and the limit of higher-order mutual knowledge is studied from a topological point of view. More precisely, the new epistemic operator limit knowledge defined as the topological limit of higher-order mutual knowledge is introduced. We then show that limit knowledge of the specific event rationality can be used for epistemic-topological characterizations of solution concepts in games. As a first step towards this scheme, we construct a game where limit knowledge of rationality appears to be a cogent strict refinement of common knowledge of rationality in terms of solution concepts. More generally, it is shown that for any given game and epistemic model of it satisfying some specific condition, every possible epistemic hypothesis as well as as every solution concept can be characterized by limit knowledge of rationality for some appropriate topology.