Učení diskrétních modelů gradientními metodami

• Název práce v češtině: Učení diskrétních modelů gradientními metodami
• Název v anglickém jazyce: Learning discrete models by gradient descent methods
• Klíčová slova: strojové učení, neuronové sítě, gradientní metody
• Klíčová slova anglicky: machine learning, neural networks, gradient descent methods
• Akademický rok vypsání: 2018/2019
• Typ práce: disertační práce
• Ústav: Katedra softwaru a výuky informatiky (32-KSVI)
• Vedoucí / školitel: RNDr. František Mráz, CSc.
• Řešitel: skrytý - zadáno a potvrzeno stud. odd.
• Datum přihlášení: 31.10.2018
• Datum zadání: 31.10.2018
• Datum potvrzení stud. oddělením: 31.10.2018
• Konzultanti: doc. RNDr. Iveta Mrázová, CSc.

Zásady pro vypracování

Recently, neural networks trained by gradient descent methods have achieved remarkable results in various fields like visual object recognition and natural language processing. While recurrent neural networks have been shown capable of approximating given formal languages based on presented labeled sample words, their behavior is difficult to explain. Furthermore, the internal state representations of recurrent neural networks can exhibit unstable behavior during computation over long sequences, which can lead to poor generalization capabilities in such cases. On the other hand, the working of discrete models like finite automata or grammars can be explained exactly and they are naturally stable on sequences of arbitrary lengths, but their learning is a computationally hard problem.

The goal of the thesis is to propose new methods for inference of discrete models of languages from sets of labeled examples using gradient descent optimization. As the error functions of discrete models are usually not differentiable, it is necessary to use their continuous relaxations or to employ approximate gradient descent methods.

The results of the thesis will be both theoretical and experimental. Properties of the proposed methods will be analyzed with respect to the class of accepted languages. The newly proposed methods and selected relevant methods will be implemented and they will be compared empirically on various instances of the inference problem.

Seznam odborné literatury

OMLIN, Christian W.; GILES, C. Lee. Constructing Deterministic Finite-State Automata in Recurrent Neural Networks. J. ACM. 1996, vol. 43, no. 6, pp. 937–972.

WEISS, Gail; GOLDBERG, Yoav; YAHAV, Eran. Extracting Automata from Recurrent Neural Networks Using Queries and Counterexamples. In: Proceedings of the 35th International Conference on Machine Learning, ICML 2018, Stockholmsmässan, Stockholm, Sweden, 2018. JMLR.org, 2018, vol. 80, pp. 5244–5253. JMLR Workshop and Conference Proceedings.

HUBARA, Itay; COURBARIAUX, Matthieu; SOUDRY, Daniel; EL-YANIV, Ran; BENGIO, Yoshua. Binarized Neural Networks. In: Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, 2016, pp. 4107–4115.

OORD, Aäron van den; VINYALS, Oriol; KAVUKCUOGLU, Koray. Neural Discrete Representation Learning. In: Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, 2017, pp. 6309–6318.

MADDISON, Chris J.; MNIH, Andriy; TEH, Yee Whye. The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables. CoRR. 2016, vol. abs/1611.00712.

ZENG, Zheng; GOODMAN, Rodney M.; SMYTH, Padhraic. Discrete recurrent neural networks for grammatical inference. IEEE Trans. Neural Networks. 1994, vol. 5, no. 2, pp. 320–330.