Cognitive computing: algorithm design in the intersection of cognitive science and emerging computer architectures

Abstract

For the first time in decades computers are evolving into a fundamentally new class of machine. Transistors are still getting smaller, more economical, and more power-efficient, but operating frequencies leveled off in the mid-2000's. Today, improving performance requires placing a larger number of slower processing cores on each of many chips. Software written for such machines must scale out over many cores rather than scaling up with a faster single core. Biological computation is an extreme manifestation of such a many-slow-core architecture and therefore offers a potential source of ideas for leveraging new hardware. This dissertation addresses several problems in the intersection of emerging computer architectures and biological computation, termed Cognitive Computing: What mechanisms are necessary to maintain stable representations in a large distributed learning system? How should complex biologically-inspired algorithms be tested? How do visual sensing limitations like occlusion influence performance of classification algorithms?

Neurons have a limited dynamic output range, but must process real-world signals over a wide dynamic range without saturating or succumbing to endogenous noise. Many existing neural network models leverage spatial competition to address this issue, but require hand-tuning of several parameters for a specific, fixed distribution of inputs. Integrating spatial competition with a stabilizing learning process produces a neural network model capable of autonomously adapting to a non-stationary distribution of inputs.

Human-engineered complex systems typically include a number of architectural features to curtail complexity and simplify testing. Biological systems do not obey these constraints. Biologically-inspired algorithms are thus dramatically more difficult to engineer. Augmenting standard tools from the software engineering community with features targeted towards biologically-inspired systems is an effective mitigation.

Natural visual environments contain objects that are occluded by other objects. Such occlusions are under-represented in the standard benchmark datasets for testing classification algorithms. This bias masks the negative effect of occlusion on performance. Correcting the bias with a new dataset demonstrates that occlusion is a dominant variable in classification performance. Modifying a state-of-the-art algorithm with mechanisms for occlusion resistance doubles classification performance in high-occlusion cases without penalty for unoccluded objects.