Twincher is a new class of AI systems, conceptually distinct from neural networks and other well-established approaches. Instead of learning patterns from data, twinchers operate by exploring abstract models of reality in a way that enables their use in interpreting real-world observations, even in the presence of noise, distortions, or any other factors that are not accounted for by those models of reality. Since the approach is not reliant on indicative patterns in the data, twinchers demonstrate high reliability and accuracy in generic regression tasks, such as the inference of 3D object parameters from monoscopic, stereoscopic, or LiDAR images.
The model of reality explored by a twincher can be any transformation of a parameter vector p into an idealized approximation of reality given by a vector s of higher dimensionality (e.g., a 3D object, a physics simulation, or an ML model). Twinchers first explore and then use a given model to iteratively find p* that provides the best explanation for any given s*. They achieve two distinct properties. First, for idealized s, they always converge to the p that yields given s. Second, for real s*, the error for p is bounded above by a cap proportional to the difference between the real and idealized s.
In such a way, the errors arise only from the discrepancy between idealized and real s and provably never from the interpretation failures. Note that complete elimination of such interpretation errors is usually unachievable for gradient-based methods due to the non-zero probability of getting stuck in a local minimum of the loss function. Moreover, twinchers can explicitly learn generic or provided models of these discrepancies, progressively widening the range of bounded-error operation and tightening error guarantees.
Due to conceptual differences, twinchers possess other unusual properties. They can actively study given models of reality, identify subtle cases and learn them more thoroughly – mitigating the curse of dimensionality and enabling a new form of generalization. Their learning principles naturally extend to ill-posed problems, permitting multi-component adversarial routines suitable for learning the ability to infer action plans within self-generated models of reality. Together, these properties open a clear path toward high-impact applications, including industrial diagnostics, autonomous systems and robotics.