Initiated by Isaac Newton, models have been developed in the Euclidean space RN based on physical world measurements which give rise to the set R of real numbers. Likewise (statistical) models were lately extended to deep-learning ones implemented by neural computing architectures which can process vast data fast. However, when humans are involved then non-numerical data also emerge, e.g. logical propositions /symbols and other, representing semantics. Therefore, any human-machine interaction model restricted in RN is inherently deficient. The Lattice-Computing (LC) paradigm has been proposed as a modeling paradigm shift to a mathematical lattice data domain, including RN, where partial order may represent semantics. This work focuses on the lattice F1 of Intervals’ Numbers (INs), where an IN may represent either a real number or a fuzzy number or a probability distribution, hence F1⊃R. A novelty here is the introduction of a Hilbert space, namely space of Generalized Intervals’ Numbers (GINs) G1, where G1⊃F1, within which F1 is a convex cone. Conventional deep-learning transformers can potentially be enhanced by, first, processing information granules represented by INs and, second, replacing numerical neural link weights with parametric monotone real functions thus tuning the number of tunable parameters. As a “proof-of-concept”, preliminary computational experiments demonstrate the capacity of a 3-layer IN Neural Network (INNN) in cascade with a YOLO deep-learning model, in a pattern recognition agricultural problem regarding the recognition of grapes in a vineyard. The far-reaching potential of the proposed techniques is discussed.