AGENIUM SPACE

OUR COMPANY

International reference in Deep Learning (DL) and Deep Neural Networks (DNN) simplification for DL at the edge

Created in 2018, team of 20, PhD and engineers in image processing and artificial intelligence.

National support earned:

CNES SME Label for Onboard artificial intelligence
(Deep Learning at the edge solutions)
National Priority of CNES
(GSTP-Make submission)
DeepTech Label
(BPI)

AGENIUM is selected by French Government’s Space Recovery Plan for the final demonstration of its DL at the edge service. .

Proof of cutting edge competence in AI :.

SpotGeo 1st place! https://kelvins.esa.int/spot-the-geo-satellites/home/
AIRQuality 2nd place! https://ai4eo.eu/ai4eo-challenge1
Space Debris, the origin 4th place https://kelvins.esa.int/space-debris-the-origin/home/
Cloud Cover Detection Less than 1% gap with the winner https://www.drivendata.org/competitions/83/cloud-cover/leaderboard/

OUR SOLUTION

Adaptative AI solution for resource constrained satellite missions

training

simplification

deployment

run

model training

For QI problems and information extraction

DNN sourced from AGENIUM
or a 3rd party

DISTILLATION QUANTIZATION

DNN Preparation for boarding
Adaptation
to platform / sensor

HARDWARE DEPLOYMENT

Target platform testing

Upload support

MAINTENANCE

Features and capabilities updates

FOCUS

Any neural network distillation service enabling AI on-board

reduction

Reduction rate of DNN up to 98%
Performance is almost the same of the original network (attrition rate less than 5%)

First distillation in few days, optimized distillation in few weeks
Tailored to optimize for customer application needs

WIDE APPLICABILITY

Use case independent
Application groups: target detection, segmentation, anomaly alarm triggering

Internally tested applications: detecting clouds, boats, forests, deforestation, oil spills, destroyed buildings

compatibility

Multi-platform: compatible with a large range of on-board devices
Multi-sensor: available for optical, radar and hyperspectral sensors
In flight service update: available for on going missions, not only at launch

VALUE PROPOSITION

Aim to provide smart payloads

data
reduction

DATA
SELECTION

COLLABORATION BETWEEN SATELLITES

COST
REDUCTION

INCREASED
RESPONSIVENESS

Extraction of relevant information, reduction of the data volume to be stored or downlinked

Selection of interesting data (object presence or absence, image quality analysis)

Object detection by one sensor will specify acquisition by the following sensor

Reduces development and maintenance costs of ground segments for data reception and processing

Enabling analytics based action or alarm triggering in orbit

ACCESS TO DNN REDUCTION TOOLS

Safety of customer’s data are our utmost priority

AGENIUM SPACE PROVIDES DNN REDUCTION IN 3 WAYS:

@AGENIUM [available]

DNN and training data is provided to AGENIUM Space. Then AGENIUM runs reduction and optimisation for customer on AGENIUM’s ML high performance computers.

@Cloud [ETA Q2.2023]

Customer uses AGENIUM Space software on a cloud platform to perform distillation, without disclosing DNN and training data.

@Customer [ETA Q4.2022]

Secured docker image delivered on a USB-stick to be run at customer’s site. Computing options:

Customer’s own high performance machine-learning computer (HPC);
AGENIUM Space’s recommended HPC procured by customer;
AGENIUM Space’s leased HPC, with hard-drive removed after completion.

SERVICES

Helping our customers to gain the most from our products and competencies

SERVICES OFFERED:

Optimisation of DNN distillation according to specific customer priorities
Adaptation of reduced-DNN to a specific target platform (e.g. XILINX Ultrascale+ ZU9EG, Myriad, Kintex, etc.)
Support for DNN source improvements and distillation algorithm updates

THE SERVICES CAN BE PROVIDED AT: