Yigit Demirag

I am a PhD candidate at the Institute of Neuroinformatics of ETH Zürich. I am interested in gradient-based learning on neural network accelerators. My PhD advisor is Giacomo Indiveri.

My PhD is focused on algorithm-hardware codesign for efficient training on neural network accelerators built with in-memory computing technologies. I had pleasure to work with Melika Payvand, Emre Neftci, Charlotte Frenkel, Angeliki Pantazi (IBM Research Zürich) and Elisa Vianello (CEA-LETI) on several projects. I designed the course material and teached Spiking Neural Network on Neuromorphic Processors for undergraduate students at ETH Zürich.

As of July 2023, I am a Student Researcher at Google. My previous internships include bio-inspired quantized training of neural networks with Blake Richards at MILA in 2023, modeling memory technologies at Advanced Technology Development Group of Samsung Electronics in 2018, working on backprop implementation on crossbar arrays at EPFL in 2017 and SIMD optimization of monte carlo simulations at CERN in 2015.

Email  /  CV  /  GitHub  /  Google Scholar  /  Twitter

Publications

Mosaic: in-memory computing and routing for small-world spike-based neuromorphic systems

Network of biologically plasuible neuron models can solve motor tasks through heterogeneity

DenRAM: Neuromorphic Dendritic Architecture with RRAM for Efficient Temporal Processing with Delays

Overcoming phase-change material non-idealities by meta-learning for adaptation on the edge

Reconfigurable halide perovskite nanocrystal memristors for neuromorphic computing

Biologically-inspired training of spiking recurrent neural networks with neuromorphic hardware

PCM-Trace: Scalable synaptic eligibility traces with resistivity drift of phase-change materials

Online training of spiking recurrent neural networks with phase-change memory synapses

Analog weight updates with compliance current modulation of binary ReRAMs for on-chip learning

Modeling electrical resistance drift with ultrafast saturation of OTS selectors

Multiphysics modeling of GST-based synaptic devices for brain inspired computing

Open Source

Gradients without Backpropagation

JAX implementation of the Gradients without Backpropagation paper, which gets rid of backpropagation of errors and estimate unbiased gradient of loss function during single inference pass via perturbed Jacobian-vector product.

e-prop local learning rule for training recurrent networks

JAX implementation of e-prop algorithm. It is written to be simple, clean and fast. I replicated the pattern generation task as described in the paper.

Dynap-SE1 Neuromorphic Processor Simulator

A minimal and interpretable Brian2 based DYNAP neuromorphic processor simulator for educational purposes.

Spiking Neural Network Simulator for TPU

A simple Colab notebook for the training of recurrent spiking neural networks on TPU or multi-GPU settings. On a single A100 GPU, it can train a small size network of 512 LIF neurons via SHD dataset under <50 sec.

Physical model of HfO2/TiOx Bilayer ReRAM memory cell

Compact model simulation of ReRAM cell. The model simulates IV characteristics of the cell, incorporating gradual SET and abrupt RESET processes. It considers filament geometry as a combination of plug and disc, and employs the Schottky barrier to represent the metal/oxide interface’s high work function, which depends on the oxide’s defect concentration.

2023 🪴