Advanced applications of neural networks (deep learning)

This class provides students hands-on experience in training modern neural networks architectures, acting as universal feature extractors (deep learning). Specialized feed-forward (convolutional network) and recurrent (long short-term memory networks) architectures are introduced. The material is organized around specific applications concerning topics important for cognitive science, for example image recognition, language modeling, modeling action and perception, cognitive robotics. Students train their own models, and experiment with already published models from various domains. The course uses Python programming language and popular neural network libraries (PyTorch, Keras, TensorFlow).

Neural networks form a very interesting group of computational models with a rich history of applications in cognitive science. Initially they were

devised as a simplified model of biological neurons, but later it was discovered that they may be used to model arbitrary dynamicalprocesses, learn mappings between points in high-dimensional spaces and generalize that knowledge. Research in the field of neural networks was pioneered also by cognitive psychologists, who used them to model processes of attention, language acquisition, language production etc.

In the last few years multi-layer neural networks gained popularity as trainable extractors of meaningful features from unstructured data. Progress in network architectures as well as in computer hardware resulted in unprecedented successes in image and audio recognition, text processing, robotic control. Techniques of transfer learning allow generalization from one domain to another. This makes neural networks not only natural candidates for conceptual models of cognitive processes, but also practical tool for analyzing experimental data.

The course will be structured around concrete applications of neural networks relevant to cognitive science. It should provide students with

intuitions regarding strengths and limitations of these models. After this class students should be able to use existing models and adapt them to

their purposes.

Literature:

• Ian Goodfellow, Yoshua Bengio and Aaron Courville (2016). Deep Learning

https://www.deeplearningbook.org/

• Vincent François-Lavet, Peter Henderson, Riashat Islam, Marc G. Bellemare and Joelle Pineau (2018), An Introduction to Deep Reinforcement Learning

https://arxiv.org/abs/1811.12560

• Li Deng and Dong Yu (2014), Deep Learning: Methods and Applications

https://www.microsoft.com/en-us/research/publication/deep-

learning-methods-and-applications/

• Nikola Kasabov (2019). Time-Space, Spiking Neural Networks and Brain-Inspired Artificial Intelligence

https://link.springer.com/book/10.1007/978-3-662-57715-8

Student knows and understands:

- Python libraries for building deep neural networks (K_W04, K_W08)

- strength and weaknesses of neural networks, their modern applications and different roles they perform in cognitive science (K_W01, K_W02)

Student is able to:

- discuss particular applications of neural networks within the domain of cognitive science (K_U01)

- train new deep learning models or adapt existing ones to model particular phenomena (K_U02, K_U03, K_U04, K_U05)

- track recent advances in a rapidly evolving field of deep learning (K_K01, K_K02)

Projects (100%) Students work in pairs and prepare one larger project during the semester. The project should concern applications of neural networks to cognitive phenomena. Topics are discussed individually with the instructor.

Two unexcused absences are allowed in the semester.