Errors bounded (single pixel)
Errors independent
No artefacts
Embedded noise model
Compression ratio, avg.
Compression speed, avg.
Lossy
compression
JPEG
7:1
100MB/s
per core
Limited
error
BD3, LERC
5:1
50MB/s
per core
Visually
lossless
JPEG-2000
7:1
10MB/s
per core
Lossless
compression
LZW, FLIF
1.5:1
10MB/s
per core
Metrologically
accurate
Dotphoton
7:1
200MB/s
per core
Cost of cloud storage
& network, USD/year
CO2 emissions,
tonnes/year
Time required
to transfer on 100Gbps
Reliably emulate different acquisition modalities (less intensity, higher exposure, etc) in real-life environments, by generating images that follow the same distribution albeit with statistically correct variations.
Prevent systematic bias in data distribution that degrades ML models, as Dotphoton adds robustness to data changes in illumination, magnification, and other sensor parameters.
Track images based on the generation of the metadata to improve data accessibility.
84%
less energy used with networking & storage optimization
6:1
typical compression ratio
8:1
typical
compression
ratio
4x
more images stored & processed on the onboard RAM and flash
4:1
typical compression ratio
5:1
more images on cloud without plan upgrade
Good practices for health applications of machine learning: Considerations for manufacturers and regulators
Available from ITU website, 2023
Physical data models in machine learning imaging pipelines
Machine Learning and the Physical Sciences workshop, NeurIPS 2022, selected for a contributed talk
Data models for dataset drift controls in machine learning with optical images
TMLR (Transactions on Machine Learning Research), 2023. Presented as workshop paper at: ICML Spurious Correlations, Invariance, and Stability Workshop, 2023 • ICML Differentiable Almost Everything Workshop, 2023
Data-centric AI workflow based on compressed raw images
8th International Workshop on On-Board Payload, Athens, 26 September 2022