Tipus de document
Treball de fi de màsterData de publicació
Llicència de publicació
Si us plau utilitzeu sempre aquest identificador per citar o enllaçar aquest document: https://hdl.handle.net/2445/230679
Handwriting Generation Using Spiking Neural Networks
Títol de la revista
Autors
Director/Tutor
ISSN de la revista
Títol del volum
Recurs relacionat
Resum
Spiking Neural Networks (SNNs) provide a biologically inspired framework for processing temporal information through discrete spike events. Their event-driven nature makes them especially suitable for sequence-generation tasks, where the timing of the signal is central to the problem. This Master’s Final Project studies the generation of handwritten characters using recurrent SNNs trained with eligibility propagation (e-prop) in the NEST simulator.
The project is based on the idea that handwriting should not only be treated as a static image, but as a temporal movement trajectory. A handwritten character is produced by a sequence of pen displacements, stroke transitions, pressure variations, and pen lifts. Therefore, the model developed in this thesis generates handwriting through three temporal readout signals: horizontal displacement, vertical displacement, and pen pressure. The spatial readouts are represented as derivatives, $\Delta x(t)$ and $\Delta y(t)$, and the visible trajectory is reconstructed only after simulation by cumulative summation.
The work develops a controlled pipeline for training, executing, and decoding recurrent SNN handwriting models. The system receives frozen spike-train inputs associated with characters, processes them through a recurrent spiking population, and generates the corresponding trajectory through a small readout layer. The pressure readout allows the model to represent multi-stroke symbols without drawing artificial connections between strokes, while also providing a more natural interpretation of stroke visibility and intensity.
The thesis focuses on improving control over the generated output. This includes the use of derivative targets to obtain a clearer movement representation, pressure-based decoding for multi-stroke generation, input spike perturbations to produce natural variation. The results show that e-prop trained SNNs can reproduce legible handwritten symbols from very limited data, while also showing that decoding consistency, input timing, and training duration strongly affect the quality of the generated trajectory.
Overall, this work presents a compact and data-efficient approach to handwriting generation with recurrent SNNs. Its main contribution is not only to generate visible characters, but to define a more interpretable and controlled framework for online handwriting generation, where movement, pressure, variability, and stability are treated as separate components of the system.
Descripció
Treballs finals del Màster en Matemàtica Avançada, Facultat de Matemàtiques, Universitat de Barcelona: Any: 2026. Director: Albert Ruiz
Matèries (anglès)
Citació
Col·leccions
Citació
POMAR PALLARES, Marc. Handwriting Generation Using Spiking Neural Networks. [consulted: 25 of July of 2026]. Available at: https://hdl.handle.net/2445/230679