Predictive neural fields for improved tracking and attentional properties

Abstract

Abstract-Predictive capabilities are added to the competition mechanism known as the Continuum Neural Field Theory, in order to improve and extend its attentional properties. In order to respect the distributed and bio-inspired nature of the model, the prediction is introduced as an internal stimulation, directly determined by the past field activity. Building on mathematical developments and optimization techniques, performance is ascertained on a 2D tracking application where the system must robustly focus on a target despite rapid movements, noise and distracters. In addition to a consistent gain on previously observed capabilities, the extended model can also track stimuli with full occlusions and static obstacles on the trajectory.

Spatiotemporal pattern discrimination using predictive dynamic neural fields

“…The research presented in this abstract extends the initial experimental results and mathematical accuracy proof obtained with a single predictor [1] to a set of predictors. This distributed model is grounded on the Continuum Neural Field Theory (CNFT) that uses global inhibition and local excitation to implement competition [2].…”

A unified dynamic neural field model of goal directed eye movements

“…Nevertheless, the system is able to adapt and converge on a limit cycle attractor, tracking the target. The equation can be modified to further improve the tracking accuracy by biasing the peak dynamics in a given direction, either by making the lateral connectivity kernel asymmetric (Cerda & Girau, 2010), or by adding internal projections of activity to the stimulation (Quinton & Girau, 2011). The latter is chosen and extended in this paper, as we model the tracking of rapidly moving stimuli, as well as the dynamic transitions between the different types of eye-movements.…”

“…In order to progressively introduce the components of the ANF model, this section starts from the dynamic neural fields initially proposed by Wilson and Cowan (1973), developed by Amari (1977), and later applied to visual attention directed to moving targets (Rougier & Vitay, 2006). We then introduce the PNF model with a single projection that can bias the dynamics of the neural fields (Quinton & Girau, 2011), then with many projections (Quinton & Girau, 2012), before including eye movements. The increase in model complexity goes hand in hand with extended attentional capabilities, from the detection of fixed targets in the visual field to the generation of tracking eye movements toward a moving target.…”

“…A set of independent neural fields have been previously used to cover the range of expected motions, but kernels could also be dynamically weighted and combined to reflect the probability of the expected motions (model) to fit the actual dynamics (observation). Here, in order to facilitate transitions between trajectories, and to permit their combination into novel trajectories with which the system can synchronize, we rather proceeded on the decomposition of the stimulation term (Quinton & Girau, 2011. From a neural perspective, these approaches may all break down to adjusting synaptic weights on the long term (to adapt to new forms of motion), and synaptic gating on the short term (to exploit the repertoire of expected motion).…”

Bio-inspired heterogeneous architecture for real-time pedestrian detection applications

“…Incorporating an evolution model of the target dynamics, DNFs can even further improve detection reliability by reducing the FPR. The following paragraphs rely on the classical stationary equation for explanatory purpose, but the reader can refer to [29] for details on the predictive version using a linear model of movement.…”