Institutional Repository [SANDBOX]
Technical University of Crete
EN  |  EL

Search

Browse

My Space

Bioinspired DNN architectures with dendritic structure

Pantzekos Lampros

Full record


URI: http://purl.tuc.gr/dl/dias/E3466A49-13C0-41E6-8AC9-484A0C47CFC4
Year 2023
Type of Item Diploma Work
License
Details
Bibliographic Citation Lampros Pantzekos, "Bioinspired DNN architectures with dendritic structure", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2023 https://doi.org/10.26233/heallink.tuc.97389
Appears in Collections

Summary

Artificial Neural Networks (ANNs) implemented in Deep Learning architectures have been successfully used to solve a wide range of challenging machine learning tasks. However, in order to achieve top performance, they typically require a substantial amount of energy. In contrast, the brain operates at a very low energy level. Drawing inspiration from biological dendrites and the aforementioned limitations of ANNs, the Poirazi lab of IMBB-FORTH introduced a bio-inspired ANN architecture with a dendritic structure and receptive field. Regarding the learning rule, backpropagation is fully applied. Training parameters are updated using the Adam optimization algorithm instead of the classical gradient descent algorithm. Based on their initial high-level Keras implementation, a lower-level Numpy implementation was developed in this thesis to analyze and understand in depth this model and its training process. Following this, an FPGA-based architecture for the training process of this bio-inspired ANN was designed, implemented, and downloaded onto the Xilinx ZCU 102 board in this thesis. In this developed FPGA implementation, training has been accelerated and power/energy consumption has been greatly reduced as a result of leveraging the high parallelism and power efficiency of the FPGA. In particular, our proposed FPGA implementation executes an epoch of training (for the MNIST dataset) in only 2.3797 seconds rather than 37 seconds on the CPU (Keras) and 17 seconds on the GPU (Keras). Furthermore, it achieves 106.15 times greater energy efficiency than the CPU implementation (Keras) and 56.5 times greater energy efficiency than the GPU implementation (Keras).

Available Files

Services

Statistics