The list of hard problems that BrainBlocks can solve.
Each of these needs better engineering, better theory, and more robust testing before they can be demonstrated for peer review.
Hard Problems
open set recognition
(Simultaneous Classification And Novelty Detection) (SCAND) (Class Recognition with Clutter)
- recognizing known classes in presence of unknown classes not previously observed
incomplete data classification
(Incomplete Data Classification) (Partial Feature Classification) (View Invariant Features)
- missing elements of the data to be classified
abnormal class shapes
(Non-Trivial Class Boundaries)
- disconnected classes, classes with holes, manifolds
learn without retraining, no catastrophic forgetting
(Add New Knowledge While Guaranteeing Old Performance) (1. Prevent Regression when adding new knowledge) (2. Integrate and improve old results with new knowledge)
- with sufficient capacity, can continue learning extra things without forgetting the old
human explainability
(Language, Rules, or States Expressed to user) (Inspection of individual bit meanings) (Collage of meanings) (Shading of space to express numerical meanings)
- introspection to what is going on under the hood
small data sets
(Limits of small data problems) (What questions can be answered? Where is the boundary?) (What algorithms are suited?)
- doesn’t need vast data sets for training