RESUMO
This article investigates the application of spiking neural networks (SNNs) to the problem of topic modeling (TM): the identification of significant groups of words that represent human-understandable topics in large sets of documents. Our research is based on the hypothesis that an SNN that implements the Hebbian learning paradigm is capable of becoming specialized in the detection of statistically significant word patterns in the presence of adequately tailored sequential input. To support this hypothesis, we propose a novel spiking topic model (STM) that transforms text into a sequence of spikes and uses that sequence to train single-layer SNNs. In STM, each SNN neuron represents one topic, and each of the neuron's weights corresponds to one word. STM synaptic connections are modified according to spike-timing-dependent plasticity; after training, the neurons' strongest weights are interpreted as the words that represent topics. We compare the performance of STM with four other TM methods Latent Dirichlet Allocation (LDA), Biterm Topic Model (BTM), Embedding Topic Model (ETM) and BERTopic on three datasets: 20Newsgroups, BBC news, and AG news. The results demonstrate that STM can discover high-quality topics and successfully compete with comparative classical methods. This sheds new light on the possibility of the adaptation of SNN models in unsupervised natural language processing.
Assuntos
Potenciais de Ação , Modelos Neurológicos , Redes Neurais de Computação , Humanos , Potenciais de Ação/fisiologia , Neurônios/fisiologia , Plasticidade Neuronal/fisiologia , Processamento de Linguagem NaturalRESUMO
For a long time, the ability to solve abstract reasoning tasks was considered one of the hallmarks of human intelligence. Recent advances in the application of deep learning (DL) methods led to surpassing human abstract reasoning performance, specifically in the most popular type of such problems-Raven's progressive matrices (RPMs). While the efficacy of DL systems is indeed impressive, the way they approach the RPMs is very different from that of humans. State-of-the-art systems solving RPMs rely on massive pattern-based training and sometimes on exploiting biases in the dataset, whereas humans concentrate on the identification of the rules/concepts underlying the RPM to be solved. Motivated by this cognitive difference, this work aims at combining DL with the human way of solving RPMs. Specifically, we cast the problem of solving RPMs into a multilabel classification framework where each RPM is viewed as a multilabel data point, with labels determined by the set of abstract rules underlying the RPM. For efficient training of the system, we present a generalization of the noise contrastive estimation algorithm to the case of multilabel samples and a new sparse rule encoding scheme for RPMs. The proposed approach is evaluated on the two most popular benchmark datasets I-RAVEN and procedurally generated matrices (PGM) and on both of them demonstrate an advantage over the state-of-the-art results.
RESUMO
Spike-timing-dependent plasticity (STDP) is one of the most popular and deeply biologically motivated forms of unsupervised Hebbian-type learning. In this article, we propose a variant of STDP extended by an additional activation-dependent scale factor. The consequent learning rule is an efficient algorithm, which is simple to implement and applicable to spiking neural networks (SNNs). It is demonstrated that the proposed plasticity mechanism combined with competitive learning can serve as an effective mechanism for the unsupervised development of receptive fields (RFs). Furthermore, the relationship between synaptic scaling and lateral inhibition is explored in the context of the successful development of RFs. Specifically, we demonstrate that maintaining a high level of synaptic scaling followed by its rapid increase is crucial for the development of neuronal mechanisms of selectivity. The strength of the proposed solution is assessed in classification tasks performed on the Modified National Institute of Standards and Technology (MNIST) data set with an accuracy level of 94.65% (a single network) and 95.17% (a network committee)-comparable to the state-of-the-art results of single-layer SNN architectures trained in an unsupervised manner. Furthermore, the training process leads to sparse data representation and the developed RFs have the potential to serve as local feature detectors in multilayered spiking networks. We also prove theoretically that when applied to linear Poisson neurons, our rule conserves total synaptic strength, guaranteeing the convergence of the learning process.
Assuntos
Modelos Neurológicos , Redes Neurais de Computação , Aprendizagem/fisiologia , Plasticidade Neuronal/fisiologia , Neurônios/fisiologiaRESUMO
The goal of General Game Playing (GGP) has been to develop computer programs that can perform well across various game types. It is natural for human game players to transfer knowledge from games they already know how to play to other similar games. GGP research attempts to design systems that work well across different game types, including unknown new games. In this review, we present a survey of recent advances (2011 to 2014) in GGP for both traditional games and video games. It is notable that research on GGP has been expanding into modern video games. Monte-Carlo Tree Search and its enhancements have been the most influential techniques in GGP for both research domains. Additionally, international competitions have become important events that promote and increase GGP research. Recently, a video GGP competition was launched. In this survey, we review recent progress in the most challenging research areas of Artificial Intelligence (AI) related to universal game playing.
Assuntos
Inteligência Artificial , Jogos de Vídeo/psicologia , Teoria dos Jogos , Humanos , Método de Monte Carlo , Inquéritos e Questionários , Jogos de Vídeo/provisão & distribuiçãoRESUMO
Artificial neural networks, trained only on sample deals, without presentation of any human knowledge or even rules of the game, are used to estimate the number of tricks to be taken by one pair of bridge players in the so-called double dummy bridge problem (DDBP). Four representations of a deal in the input layer were tested leading to significant differences in achieved results. In order to test networks' abilities to extract knowledge from sample deals, experiments with additional inputs representing estimators of hand's strength used by humans were also performed. The superior network trained solely on sample deals outperformed all other architectures, including those using explicit human knowledge of the game of bridge. Considering the suit contracts, this network, in a sample of 100,000 testing deals, output a perfect answer in 53.11% of the cases and only in 3.52% of them was mistaken by more than one trick. The respective figures for notrump contracts were equal to 37.80% and 16.36%. The above results were compared with the ones obtained by 24 professional human bridge players--members of The Polish Bridge Union--on test sets of sizes between 27 and 864 deals per player (depending on player's time availability). In case of suit contracts, the perfect answer was obtained in 53.06% of the testing deals for ten upper-classified players and in 48.66% of them, for the remaining 14 participants of the experiment. For the notrump contracts, the respective figures were equal to 73.68% and 60.78%. Except for checking the ability of neural networks in solving the DDBP, the other goal of this research was to analyze connection weights in trained networks in a quest for weights' patterns that are explainable by experienced human bridge players. Quite surprisingly, several such patterns were discovered (e.g., preference for groups of honors, drawing special attention to Aces, favoring cards from a trump suit, gradual importance of cards in one suit--from two to the Ace, etc.). Both the numerical figures and weight patterns are stable and repeatable in a sample of neural architectures (differing only by randomly chosen initial weights). In summary, the piece of research described in this paper provides a detailed comparison between various data representations of the DDBP solved by neural networks. On a more general note, this approach can be extended to a certain class of binary classification problems.