Skip Nav Destination
Close Modal
Update search
NARROW
Format
TocHeadingTitle
Date
Availability
1-2 of 2
John Thornton
Close
Follow your search
Access your saved searches in your account
Would you like to receive an alert when new items match your search?
Sort by
Proceedings Papers
. alife2018, ALIFE 2018: The 2018 Conference on Artificial Life129-136, (July 23–27, 2018) doi: 10.1162/isal_a_00032
Abstract
PDF
The predictive processing theory of cognition and neural encoding dictates that hierarchical regions in the neocortex learn and encode predictive hypotheses of current and future stimuli. To better handle uncertainty these regions must also learn, infer, and encode the precision of stimuli. In this treatment we investigate the potential of handling uncertainty within a single learned predictive model. We exploit the rich predictive models formed by the learning of temporal sequences within a Hierarchical Temporal Memory (HTM) framework, a cortically inspired connectionist system of self-organizing predictive cells. We weight a cell’s feedforward response by the degree of its own temporal expectations of a response. We test this model on data that has been saturated with temporal or spatial noise, and show significant improvements over traditional HTM systems. In addition we perform an experiment based on the Posner cuing task and show that the system displays phenomena consistent with attention and biased competition. We conclude that the observed effects are similar to those of precision encoding.
Proceedings Papers
. ecal2017, ECAL 2017, the Fourteenth European Conference on Artificial Life114-121, (September 4–8, 2017) doi: 10.1162/isal_a_022
Abstract
PDF
It has become increasingly popular to view the brain as a prediction machine. This view has informed a number of theories of brain function, the most prominent being predictive processing, where generative hypotheses are iteratively updated by error signals. In this treatment we take a lower level approach by examining the hierarchical temporal memory framework, which views individual pyramidal cells as the primary predictive unit of a self-organizing networked sequence learning system. Within this computational framework, the cell behaviour is constrained by a number of parameters which are static and shared across all cells. To further increase the adaptability of the cells, we shift away from this paradigm by introducing the concept of dynamic thresholds. This allows for the activation threshold (the amount of activity on a distal dendrite needed to form a prediction) to be adjusted continuously and individually for each cell. As a metric we use the prior, or unconditional, probability of activity on the proximal dendrites. Our experiments show that using this metric for dynamic thresholds can improve the predictive capabilities of the system in a number of domains, including anomaly detection, where we achieve state-of-the-results on the Numenta Anomaly Benchmark.