
Site Title
Add a subtititle here.
Embed an App
Or write your own HTML code! (HTML is Pro only)
Skip to contentIndependent consulting & contract R&D
From neural principles
to deployable AI.I help teams build efficient AI for signals, sensors and autonomous systems, connecting mathematical ideas with working algorithms and real hardware constraints.
Spatiotemporal AI / Event-based vision / Neuromorphic computing
Neural signaling · conceptual illustration
Research & engineering experience
BrainChipKelzalBrain Corp / QualcommSony CSLSalk Institute01 / Capabilities
Deep research.
Practical engineering.I work across theory, implementation and technical direction. An engagement can start with a specific bottleneck, a new architecture, or a research question that needs a working prototype.
01 Design the right
neural architectureSpatiotemporal networks, neural state-space models, spiking networks and learning rules for streaming data.
Architecture design · mathematical modeling · research prototypes
02 Make perception
work at the edgeEvent-based vision, RF sensing and spectroscopy, radar processing, sensor fusion and adaptive control under memory, latency and power constraints.
Algorithm development · benchmarking · hardware-aware optimization
03 Set a clear
R&D directionEvaluate competing approaches, define a roadmap, and connect research, engineering and product teams.
Technical assessment · proposal development · scientific leadership
Temporal structure → useful memory●Signal / events→Temporal kernels→PredictionTemporal dynamicsTemporal kernels · illustrative schematic Event-based visionDriving scene · conceptual illustration 02 / Selected projects
Selected projects.
Neural architectures, perception systems and robots that learn. A selection of my work across computational neuroscience and applied AI.
Giga-pixel event-based image generation from a moving vehicle over bumps · Kelzal demonstration with a low resolution (180 x 240) sensor 8 projects
BrainChip · Research with Yan Ru Pei
01PLEIADES & TENNs
Neural architectures for streaming data.
I invented PLEIADES and co-invented Temporal Event-Based Neural Networks (TENNs), connecting structured temporal kernels with efficient processing of streaming data.
My roleArchitecture invention, mathematical formulation and research direction
Temporal kernels built from orthogonal polynomial bases.
Figure from the PLEIADES paper by Pei & Coenen.100%DVS128 hand gesture test accuracy
192K parameters · Small output filterReported DVS128 test result after about 0.44 s of network warm-up. The small majority output filter adds 150 ms of output delay.
Efficiency in published benchmarks
≈57× fewer parameters
Road-scene object detection
PLEIADES + CenterNet used 576K parameters versus 32.8M for Gray-RetinaNet, with mAP of 0.556 versus 0.430.
About 17× fewer MACs per second, while producing predictions at 100 Hz versus 20 Hz for the reference.
PLEIADES paper, Table 3
≈46× fewer parameters
Camera-based object detection
On KITTI, PLEIADES + CenterNet used 570K parameters versus 26M for the SimCLR / ResNet-50 reference, with mAP of 0.576 versus 0.570.
About 4.6× fewer MACs per second at the same 10 frames per second.
PLEIADES paper, Table 9
≈26× fewer parameters
Event-based eye tracking
PLEIADES + CenterNet used 277K parameters versus 7.1M for CETM. Pupil predictions fell within 10 pixels of the ground-truth pupil position in 99.58% versus 99.26% of cases.
Comparison on the AIS 3ET+ eye-tracking benchmark.
PLEIADES paper, Table 2
Ratios calculated from the reported parameter and operation counts. Each comparison names its task and reference model. MACs are multiply–accumulate operations.
Neural state-space models · Spatiotemporal learning · Efficient inference
More about this project
PLEIADES represents temporal kernels with orthogonal polynomial bases. The network learns the coefficients that shape its response over time, giving it a compact way to model temporal structure.
The broader TENN work connects parallel training through temporal convolutions with recurrent computation for online inference. Applications span event-based perception, raw audio and physiological signals.
The PLEIADES paper reports results on hand gesture recognition, eye tracking and automotive object detection. Its DVS128 result reaches 100% test accuracy with a small majority output filter, with 192K network parameters.
The BrainChip TENN white paper reports an earlier KITTI detector with 165K parameters versus 26M for its SimCLR / ResNet-50 reference: about 158× fewer parameters. It reports 57.6% versus 57.5% mAP and about 6.3× fewer MACs per second at 20 inferences per second. This is a separate TENN configuration from the published PLEIADES comparisons above.
With Yan Ru Pei, I also applied TENNs to a large language model that ran in real time on an FPGA in BrainChip demonstrations.
Pixel intensity at events and event-based optical flow · Kelzal Kelzal / Qelzal
02Event-based perception
Detecting motion and recognizing objects as events arrive.
I founded and led Kelzal, developing algorithms for event-based vision sensors. The work covered object detection and tracking, short-latency recognition, stereo depth and optic flow.
- Car / non-car recognition
- <3 msDemonstrated using less than 3 ms of event-based sensor data.
- Less vision data
- 400–1,000×Compared with conventional frame cameras in the demonstrated scenes.
My roleFounder, CEO and CTO · Algorithm and software development
Event cameras · Spiking networks · Stereo vision · Sensor processing
More about this project
Developed event-based methods for vehicle and pedestrian detection, tracking and action recognition, including spiking neural networks trained for short-latency responses. Demonstrated car/non-car recognition using less than 3 ms of sensor data, and standing/walking classification with under 1 ms of processing after a few milliseconds of sensor input.
Built stereo vision and motion-estimation algorithms, together with denoising, event-timing calibration, image registration and reconstruction. These components supported perception under camera motion and vibration. I also developed event-based resolution upscaling, including image generation from a low-resolution sensor carried over uneven terrain.
Stereo depth estimation Event-image registration Aircraft tracking in front of the sun: Conventional camera (left), event-based vision camera (right) Kelzal / Qelzal
03Aircraft & drone tracking
Finding small targets while the sensor itself is moving.
Developed aircraft and drone detection and tracking, including image stabilization driven by inertial measurements and online estimation of flashing-light frequencies.
- Aircraft flashing-light detection
- 11 kmDemonstrated frequency-based discrimination of aircraft lights from street lighting.
- Drone detection and tracking
- ≈1 kmUsing event-based resolution upscaling for small, distant targets; drone tracked up to 1/10 the size of a single pixel.
My rolePerception algorithms · Sensor integration · Prototype development
Target tracking · Sensor fusion · Frequency estimation · Autonomous systems
More about this project
Combined event-based tracking with IMU-driven stabilization to handle camera motion.
Developed an efficient event-based online Fourier transform and frequency estimation to distinguish flashing light sources. Applications included aircraft detection and tracking, alongside NSF and DARPA work on sense-and-avoid systems and optical guidance.
The online Fourier method extracted light-frequency information as events arrived. A later experiment extended aircraft-light detection to 11 km, and the resolution-upscaling work included drone detection and tracking at approximately 1 km.
Drone tracking with image stabilization BrainChip
04RF & radar signal intelligence
Temporal models for radio-frequency signals.
Invented SCORPIUS for RF and radar signal modeling and automatic modulation classification. Co-authored the technical proposal and served as a principal investigator on an Air Force program for low-power neuromorphic radar classification.
- Automatic modulation classification
- State of the artSCORPIUS redefine the state-of-the-art performance on RadioML 2018 RF data at BrainChip.
- Air Force R&D program
- $1.8MPrincipal investigator and co-author of the technical proposal.
My roleArchitecture invention · Principal investigator · Technical proposal development
RF and radar · Automatic modulation classification · Neuromorphic computing
More about this project
The work brought temporal neural modeling to RF and radar classification, with an emphasis on efficient processing under hardware and power constraints.
SCORPIUS redefine the state-of-the-art performance on RadioML 2018 RF data at BrainChip. The project combined architecture invention with the scientific and proposal work needed to establish a funded development program.
Kelzal retail interface concept. Kelzal / Qelzal
05Vision for autonomous retail
Reading products, text and barcodes on embedded systems.
Developed camera-based barcode decoding, OCR and text localization, alongside logo recognition and fruit-and-vegetable classification. Built recognition pipelines and optimized their execution on embedded NVIDIA GPUs.
- Recognition through optimized inference
- Embedded GPUsTensorRT / CUDA engine generation, XLA compilation, Tensor Cores and mixed precision.
My roleModel development · Inference optimization · Product R&D
Transformers · OCR · Computer vision · TensorRT · Embedded GPUs
More about this project
Adapted neural machine-translation models to barcode decoding and OCR, using CNNs, LSTMs with attention and Transformer variants. Investigated one- and two-dimensional positional encodings for these tasks.
Converted trained networks into TensorRT models and CUDA inference engines. Optimized embedded GPU execution using XLA compilation, Tensor Cores and automatic mixed precision.
Brain Corporation / Qualcomm
06Spiking networks for adaptive control
Learning to control movement through neural dynamics.
Developed spiking models of motor control, reinforcement learning and predictive control. Contributed to DARPA research on an artificial nervous system for autonomous aircraft.
- Neural modeling and learning rules
- Multiple patentsThe research also contributed to Brain Corporation’s Series B funding with Qualcomm Ventures.
My roleSenior scientist · Neural modeling · Learning-rule development
Spiking neural networks · Reinforcement learning · Predictive control
More about this project
Worked across neural representations, learning rules and stochastic control to connect biologically inspired models with adaptive robotics.
The work produced multiple patents and directly contributed to closing Brain Corporation’s Series B funding with Qualcomm Ventures, connecting neuroscience research with the company’s technology development.
Sony Computer Science Laboratories, Paris
07SENSOPAC & SPIKEFORCE
Research programs connecting neuroscience with robot control.
Created and coordinated two European research consortia that brought neuroscience, machine learning, tactile sensing and real-time spiking networks together to study perception, manipulation and robot control.
- Combined research funding
- €8.9M€2.4M for SPIKEFORCE and €6.5M for SENSOPAC.
- SENSOPAC collaboration
- 9 countriesTwelve partner institutions and around 60 researchers.
My roleProject creator · Scientific coordinator · Research-team leadership
Computational neuroscience · Sensorimotor learning · Scientific leadership
More about this project
SPIKEFORCE brought together five partner institutions, and SENSOPAC twelve from 9 different countries. The programs involved 14 and around 60 researchers, respectively. I defined research directions, helped select and recruit researchers, and coordinated scientific work across the consortia.
My own research included cerebellar models, real-time spiking networks, tactile and visual coding, and active perception. Responsibilities also covered budgets, consortium agreements, project proposals and research reviews.
Intentiva
08Visual gesture interfaces
Controlling devices through hand movements.
Co-founded and led Intentiva, developing visual hand-gesture recognition to control televisions and other interactive devices from a distance.
My roleCo-founder and CEO · Research and product development
Gesture recognition · Human–machine interaction · Startup development
More about this project
Led the visual-recognition and product-development work alongside fundraising, customer and business development. As co-founder and CEO, I also handled company operations, including legal and financial responsibilities.
RF sensors / current development
AI for RF sensing
& spectroscopySpectral analysis · sensor modeling · efficient inference
I develop machine-learning models for RF sensors, including SansEC sensing technology and SansEC Spectroscopy. This combines spectral signal processing, physical-parameter estimation and compact neural networks, with validation across operating conditions.
NASA’s SansEC technology uses wireless, open-circuit resonant sensors. SansEC Spectroscopy characterizes material properties such as permittivity, permeability and conductivity.
Current focus
Intelligence closer
to the sensor.Governed execution layer & stable API
A stable API above model formats and vendor runtimes, with reproducible build records, verified artifacts and qualification evidence.
Alongside ongoing research in efficient spatiotemporal and neuromorphic systems.
03 / Research interests
Questions that
shape the work.My interests sit where neuroscience, mathematical physics and machine intelligence meet.
↗ Time, memory & causality
Neural dynamics, state-space models and representations that preserve the timing of events.
↗ Learning from the brain
Neural coding, synaptic plasticity, independent representations and spiking computation.
↗ Embodied intelligence
Active perception, sensorimotor learning and adaptive control for robots and autonomous systems.
↗ Efficient intelligence
Sparse and event-driven processing, cognitive RF and AI that works within device constraints.
04 / Background & toolkit
Physics, neuroscience
& AI engineering.I’m Olivier Coenen, an AI scientist and research leader working from mathematical foundations through algorithms to embedded systems.
My background combines a PhD in Physics (Biophysics) at UC San Diego, research with Terrence J. Sejnowski at the Salk Institute, founding Sony CSL Paris’s Neuroscience Group, and scientific and entrepreneurial roles across AI and robotics.
I’ve led research teams, funded R&D programs and startups, with work spanning publications, patents, prototypes and product-facing systems.
Professional profile on LinkedIn ↗PhDPhysics / Biophysics
UC San DiegoBScHonours Physics
McGill UniversitySkills across the full R&D cycle
Modeling & learning
Mathematical physics; dynamical systems; optimal control; information theory; sparsity; independent coding; computational neuroscience; deep learning; reinforcement learning; neural state-space models; spiking networks.
Software & implementation
Python, PyTorch, TensorFlow, NumPy/SciPy, C/C++, CUDA/TensorRT, XLA/AMP, ROS and Linux. Reproducible training, benchmarking, ablations and model optimization.
Signals, sensors & control
Event-based processing, computer vision, image and signal processing, RF/radar, resonant sensing and spectroscopy, tracking, stereo/depth, optic flow, Kalman filtering, sensor fusion, IMU stabilization and adaptive robotics.
Scientific & business leadership
Research direction, technical strategy, competitive analysis, consortium coordination, proposals and grants, project and budget management, hiring, mentoring, entrepreneurship, fundraising, partnerships, product strategy and intellectual property.
Let’s work together
What are you
trying to build?Tell me about the problem, your target platform and the constraints that matter. We can explore a focused consulting engagement or hands-on contract R&D.