(May. 2023 – present) “Audio Source Localization using SNNs based on novel Hilbert Transform Spike Encoding”.

Description: Audio source localization using microphone arrays is of crucial importance in applications such as video conferencing, virtual reality, gaming, etc. 

This is a follow up project for the SNN localiztaion project I did before. One of the issues was that in that project, we could not reach the same precision as well-known super-resolution methods such as MUSIC albeit using a veriety of spike encoding techniques.

To solve this issue we had to go deeper and develop a novel mathematical structure based on Hilbert Transform. In the new method, we apply Hilbert Transform to extract the in-phase and quadrature components of the incoming audio which we use to extract the generalized phase of the audio. We then do spike encoding, which we call conjugate spike encoding, based on the generalized phase.

This method and this type of spike encoding is quite new in SNN litearture and seems to be the right one for localization applications. Also, using Hilbert Transform allows to reduce the implementation resources such as filterbanks, etc. and power consumption dramatically.

Outcome: Patented audio source localization method in SNNs -- Version 2.

(Apr. 2023 – June. 2023) “Automatic Gain Controller for Audio Applications”.

Description: In audio applications such as keyword spotting, baby crying detection, etc. one always needs to make sure that the input audio is amplified sufficiently so that one can extract efficient features in the presence of quantization noise at ADC. To do this, one needs to estimate the signal power over quite large windows and change the gain of programmable gain amplifier (PGA) to reverse the attenuation due to audio source distance from the microphone. However, one should be careful not to over-amplify the signal since that pushes the amplifier to the saturation region and clips the sampled audio signal.

In this project, we designed an efficient AGC for audio applications based on a Markov State Machine (MSM) with 16 states, where each state specifies how large signal amplitude after quantization is. The output of AGC module is a 4-bit digital signal that is sent as a command to PGA to adjust its gain. 

One of the challenges in AGC design is to have a feedback from the digital (MSM) to the analog (PGA) part. And one needs to design PGA to avoid any possible oscillation.

Outcome: Patented automatic gain controller (AGC) for audio applications.

(Jan. 2022 – Jan. 2023) “Exact gradient computation for spiking neural networks via forward propagation."

Description: Spiking neural networks belong to the class of nonlinear dynamical systems and have nonlinear and discontinuous dynamics in time due to their spike generation and membrane potential reset. As a result, classic back-propagation algorithm for training traditional networks has been notoriously difficult to apply to SNN due to the discontinuities at spike times. As a matter of fact, a large majority of prior work on SNN training believes that exact gradients for SNN w.r.t. their weights (trainable parameters) do not exist at all and has focused on approximation methods to produce surrogate gradients (including our recent work on EXODUS algorithm listed below). In this project, 

• by applying the implicit function theorem to SNN at the discrete spike times, we prove that, albeit being non-differentiable in time, SNNs have well-defined gradients w.r.t. their weights. More specifically, non-differentiable nature of SNNs in time does not cause any non-differentiability issue w.r.t. the trainable parameters. This is a crucial observation that had been missed by the previous literature. 

• we propose a novel training algorithm, called forward propagation (FP), that computes exact gradients for SNN. FP exploits the causality structure between the spikes and allows us to parallelize computation forward in time. It can be used with other algorithms that simulate the forward pass, and it also provides insights on why other related algorithms such as Hebbian learning and also recently-proposed surrogate gradient methods may perform well.

You can take a look at recent paper published on International Conference on Artificial Intelligence and Statistics, pp. 1812-1831, 2023.

(Jan. 2022 – present) “EXODUS: Stable and efficient training of spiking neural networks."

Description: In this project, we develop a GPU-accelerated and scalable algorithm for training spiking neural networks under surrogate gradients and back-propagation through time (BPTT) algorithm. The conventional BPTT is not scalable when the temporal dimension is large. It is also sequential in time, which makes it quite slow. We fix both issues in this project:

• we develop a vectorized version of BPTT along the time dimension by aggregating the network parameters in time. This helps us to speed up computation by running the vectorized version on GPU.

• one of the issues we encounter is this vectorizaion creates loop in the computation graph such that that the conventional chain-rule based BP algorithm is not applicable. We fix this issue by developing a new method based on Implicit Function Theorem (IFT) to unlock the loops in computation graph and let the gradient to be back-propagated across the layers.

You can take a look at our latest publication on arXiv for further information.

(Jan. 2022 – present) “Novel methods for gradient computation in ANNs and SNNs”.

Description: First-order methods for stochastic optimization has been the main building block for current deep learning and machine learning training algorithms. A crucial step in these methods is to be able to compute the gradient of the training loss w.r.t. the network parameters. Historically, the well-known method for computation of gradients is chain rule, which can be formulated as back-propagation (BP) algorithm as in implemented in many machine learning libraries such as PyTorch. One of the problems with BP is its sequential nature where the gradient computation starts from the last layer and proceeds to the first layer. With the evolution of GPUs for parallel processing, one may wonder if there are other training algorithms that (i) are not sequential (ii) can exploit time-memory resources for computation in a more optimal fashion. In this project, we investigated a generic framework for gradient computation which provided new method for gradient computation.

(Jan. 2022 – present) “Audio Source Localization using SNNs”.

Description: Audio source localization using microphone arrays is of crucial importance in applications such as video conferencing, virtual reality, gaming, etc. Localization is a well-studied classical problem in array signal processing. However, most of the exisiting methods are based on covariance-based subspace methods, which are difficult to implement on devices with low energy and computational resources. The goal of this on-going project is to develop the equivalent of classical localization for SNNs. This requires developing novel spike encodings as well as efficient data-driven algorithms/structures that work specifically with the spike (0-1) data.

Outcome: Patented audio source localization method in SNNs.

(Aug. 2021 – Dec. 2021) “Divisive Normalization”.

Description: Spiking neural networks (SNNs) are a new gerenation of neural networks in which signal processing is done by neurons communicating spikes (0-1 pulses) with one another. Due to sparse neuron activity, SNNs consume at least 1∼2 orders of magnitude less energy than conventional neural networks in similar machine learning tasks. As a result, SNNs are an interesting option for low-energy applications such as edge computation. This, however, comes at the cost of more complicated training and signal processing algorithms. The goal of this project was to design an algorithm for denoising spike signals at the input of SNN to eliminate the huge number of spikes produced because of background noise activity.

Outcome: Patented divisive normalization algorithm for eliminating background noise activity. 

(Jan. 2021 – May 2021) “Vital Signs Monitoring”.

Description: Contact-less vital signs measurement is of tremendous importance in health applications such as sleep monitoring, apnea detection, etc. Recently due to Covid-19 situation, there is yet another surge of interest to measure vital signs (respiration and heart rate) of patients/subjects in a contact-less fashion. The goal of this project within CSEM was to use the remote & high-resolution sensing capability of wideband radars to detect chest movements of patients/subjects due to respiration and heart beat. By applying robust signal processing techniques combined with data-driven machine learning algorithms, it was possible to extract respiration (with a high precision) and heart rate (with a reasonable precision) from chest movements.

Outcome: Demo of vital signs monitoring with the implemented algorithms at CSEM. 

(May 2020 – Dec. 2020) “Subject Presence Detection & Counting”.

Description: Detecting the presence of live subjects is of tremendous importance in applications such as smart offices, people counting, border control, vehicle occupancy and driver monitoring, etc. The goal of this project within CSEM was to use to design a robust system for presence detection and people counting based on wideband radars. High resolution of wideband radars enable them to detect micro-motions such as subject’s chest movement due to respiration and heart beat. Using this and also applying smart machine learning algorithms enables to detect the presence and also count the number of such live subjects in the environment with a high precision.

Outcome: Demo of people counting with the implemented algorithms at CSEM.

(Feb. 2020 – Apr. 2020) “Theory and Algorithms for Smart Sensing using Wideband Radars”.

Description: Use of wideband radars has tremendously increased recently due to their high sensing resolution, low cost, easy installation, and also their robustness against environmental effects such as mist, rain, dust, etc. The signal received at the radar reveals valuable information about moving objects in the environment, their locations, and also their movement pattern, which is of tremendous use in applications such as vital signs monitoring, subject presence detection, gesture recognition, etc. The signal received at the radar, however, goes under impairments such as fading due to multi-path propagation in the environment. The goal of this project at CSEM was to develop optimal and low-complexity signal processing algorithms based on the phase of the received signal that are robust against such impairments.

Deep Learning (TU Berlin)

(Mar. 2018 – Jan. 2020) “Information Theory of Deep Neural Networks (DNN)”.

Description: The goal of this Deutsche-Israeli project is to obtain a better theoretical understanding of the excellent performance of Deep Neural Networks (DNN) in problems related to learning, classification, data mining, etc.

(Jul. 2018 – Jan. 2019) “Scalable and low-complexity signal processing in wireless systems”.

Description: The goal of this project from “Berliner Zentrum fu ̈r Maschinelles Lernen” (BZML) is to design scalable Machine Learning algorithms for efficient signal processing, users mobility tracking/management, distributed data caching, massive device connectivity planning, etc. for high-throughput wireless systems and Internet-of-Things (IoT).

Wireless Communication/Optimization Theory (TU Berlin)

(Jan. 2017 – Jan. 2019) “Internet-of-Things (IoT)”.

Description: Internet-of-Things (IoT) is a new approach envisaged for “Technology for Smart Cities” in the next generation of wireless network (5G) currently under development. The goal of this project is to design scalable techniques for processing very large-dim signals and managing massive device connectivity encountered in IoT setups.

(Jan. 2019 – Jan. 2020) “Obtaining Gains of massive MIMO in FDD”.

Description: FDD and TDD are two Duplexing options for massive MIMO in the next generation (5G) of wireless networks. However, implementation of FDD is too costly due to the lack of channel reciprocity between the uplink and the downlink, and the need for channel state feedback. The goal of this project with Huawei, Munich, Germany, was to obtain novel signal processing techniques to reduce feedback (and system-level) overhead of FDD by developing a novel Covariance Extrapolation and Active Channel Sparsification technique.

(Jan. 2015 – Jan. 2016) “Visible Light Communication (VLC): LiFi Networks”.

Description: The goal of this project was to investigate the possibilities and also the new opportunities for providing internet access by communicating data packets with Light Emitting Diodes (LEDs) used for lighting houses, offices, airplanes, etc. instead of using the traditional WiFi technology.

Wireless Communication/Compressed Sensing (TU Berlin)

(Jun. 2016 – Jun. 2017) “Robust Beam-Alignment for mmWave Systems”.

Description: The goal of this project with Intel, Santa Clara, CA, USA, was to design a fast and scalable algorithm for channel acquisition (generally known as Beam Alignment) in mmWave systems. The outcome of the project (duly published) was a novel compressed sensing based method whose computational and system-level complexity was almost independent of the number of users and dramatically outperformed the already-existing methods.

(Jan. 2015 – Jan. 2018) “Compressed Sensing for Information Processing (CoSIP)”.

Description: The goal of this 3-year DFG funding was to develop new compressed sensing based methods for low-complexity signal processing in massive MIMO and mmWaves, as the two prominent technologies in the next generation of wireless networks (5G).

PhD Thesis in Compressed Sensing

(Sept. 2010 – Oct. 2014) “Compressed Sensing of Memoryless Sources: A Deterministic Hadamard Construction”.

Description: Recovery of signals from their linear projections (through devices, sensors, etc.) appears in almost any engineering and signal processing problem. It is known from linear algebra that if signal has dimension (degree of freedom) `n` one needs to takes at least `n` measurement to be able to recover the signal. This worst-case bound, however, can be broken if one has extra information about the signal. Namely, one may be able to take `m << n` measurements and still be able to recover such a structured signal. This has been the starting point for Compressed Sensing theory which was a very hot research topic during 2006-2020. 

In my thesis, I study compressed sensing when signals have a probablistic and sparse structure. I mainly worked on construction of partial Hadamard matrices for measuring such signal. I also used tools from math and statistics to show that the constructed class of matrices yields close to optimal number of measurement. 

I also extended the results to study compressed sensing of distributed sparse signals where the signal can be seen from `T` terminals (e.g., temperature measured through `T` sensors at different locations). I studied what are the possible measurements `m1, m2, ..., mT` one needs to take from these `T` terminals to make sure that the signal can be recovered perfectly at all `T` terminals (distributed signal recovery).

Amateur Satellite 

(Sept. 2009 – Sept. 2010) “LEO Amateur Satellite for Terrestrial Imaging”.

Description: I was managing a group of engineers working on the design of a low-earth-orbit (LEO) amateur satellite with a focus on terrestrial imaging applications. 

Satellites rotate faster than earth at low orbits. As a result, satellite appears and disappears from field-of-view several times during a day. Depending on the orbit, the effective time for communicating with sattelite can be as low as 10-60 mins. As a result, one needs very efficient acquisition methdos to find the satellite on the horizon when it comes to the field-fo-view and robust tracking methods to have the satellite always on the peak of the radar beam. Also having very robust communication protocols strong channel codes, suitable packets and frames structure, etc. is absolutely necessary for data transfer.

Machine Learning/Radar Signal Processing

(Aug. 2007 – Aug. 2009) “Radar sea clutter cancellation using chaotic signal processing”.

Description: Detecting small targets floating on the sea level is a very challenging problem in radar signal processing. In particular, all detection algorithms based on statistical signal processing are known to perform very poorly due to the wild dynamics of the sea clutter (the signal reflection from the sea level). The goal of this project (as M.Sc. thesis) was to design a radar detector using chaotic signal model for the sea clutter. The theoretical and practical results using real data showed that this detector outperforms traditional detectors by several orders of magnitude.