Characterizing Neural-Network Classification Uncertainty Through Bayesian Statistics

The team proposes a strategy, through probability modeling, for the systematic characterization of a deep neural network's (DNN's) classification uncertainty due to invariable under-sampling. In many practical classification applications, the labeled samples used for training a DNN is a small subset of the practically infinite (but countable) population set. The likelihood probability distribution, or likelihood in short, of sample x belonging to class y, i.e., p(x|y), derived from the set of labeled samples thus inevitably differ from that of the population. Moreover, during the labeling of the samples, the subject domain experts instinctively avoid controversial samples, leading to narrower and more peaked likelihood. The uncertainty caused by such factors is difficult to estimate since the population likelihood is often unknown or unknowable. Thus, the objectives of this work are to: 1) model the likelihood of the labeled samples with a theoretical probability distribution, 2) perturb the likelihood in the (assumed) direction of better representativeness, 3) generate synthetic labeled samples based on the perturbed likelihood, 4) evaluate the impact of the perturbations on classification performance using the generated synthetic samples, and thereby 5) characterize the DNN's classification uncertainty. This will be demonstrated with cloud-type classification using satellite imagery. A variational auto-encoder/decoder approach will be taken, using the encoder for dimensionality reduction, the latent normal distributions for probability modeling, and the decoder for synthetic sample generation.

Advance Earth system science knowledge through the identification, development, and demonstration of innovative information systems technologies

This is early/mid-stage (TRL 2) — the most realistic path in is NASA SBIR/STTR, which funds small businesses and research institutions to develop technology aligned with NASA's needs (equity-free, phased funding). Check whether a current SBIR/STTR solicitation topic overlaps with this project's technology area, or contact the project directly (above) to ask.

None of these are guaranteed paths for this specific project — TechPort itself doesn't have an "apply" button. Reaching out to the contact(s) above with a specific question is usually the fastest way to find out what's actually open.