Inhibition in the dynamics of selective attention: an integrative model for negative priming

Abstract

We introduce a computational model of the negative priming (NP) effect that includes perception, memory, attention, decision making, and action. The model is designed to provide a coherent picture across competing theories of NP. The model is formulated in terms of abstract dynamics for the activations of features, their binding into object entities, their semantic categorization as well as related memories and appropriate reactions. The dynamic variables interact in a connectionist network which is shown to be adaptable to a variety of experimental paradigms. We find that selective attention can be modeled by means of inhibitory processes and by a threshold dynamics. From the necessity of quantifying the experimental paradigms, we conclude that the specificity of the experimental paradigm must be taken into account when predicting the nature of the NP effect.

Keywords: computational modeling; connectionist models; negative priming; selective attention.

Figure 1

Example of a sequence of stimuli. Consecutive screens are shown. Either stimuli or a blank screen followed by a fixation cross is displayed. Acronyms are explained in Table 1.

Figure 2

Four different paradigms for NP. (A) The location priming paradigm reveals NP in the encoding of space. (B) The flanker task implements a stimulus response mapping. (C) Responses are given as vocalization in the voicekey paradigm. (D) The word-picture comparison paradigm has the advantage of a disentanglement of target identity and response. The examples have been adapted such that green always defines the target.

Figure 3

Interaction scheme of the different components of the GMNP. Perceived stimuli are decomposed into single features, each of which is represented in a single variable in the corresponding layer. Object identity is maintained by activations in the binding layer, associating the different features of a stimulus object. Most paradigms require a semantic evaluation of the stimuli in order to generate a response. Therefore, the semantic layer gates information flow from the relevant features to the action layer which decides on the action to perform. Parallel to the information flow from perception to action a so-called central executive steers the model behavior with regard to the current task, i.e., providing information about the target and the mapping of semantic variables to actions. According to the similarity of the percept and a memorized stimulus configuration, the memory layer feeds back information of the former trial. The similarity signal also affects the effectiveness of transmission between features, semantic layer and actions as well as between memory itself and actions, the latter inversely to the first.

Figure 4

Normalized average firing rate of the network as a response to input (applied from time step 0 to 50 indicated by the gray shaded region) and no input (blue). The firing rate is determined by binning the spikes in each time step. Normalization is performed by division by the average maximum firing rate at time 50. The fraction of two subsequent firing rates, which corresponds to the time constant in an exponential fixed-point dynamics, is shown in red. Black lines show the means of the respective red lines. The deviation of the blue curve from a purely exponential dynamics is apparent, but quite small, justifying the simplified dynamics as described in the text.

Figure 5

Distribution of membrane potentials averaged over 10,000 trials. Note that the potentials are mostly uniformly distributed, as the color map only covers values from 0.0098 to 0.0115. Nevertheless, the fine grained plot reveals the processes generating the firing rates analyzed in Figure 4: initially all neurons are pushed toward higher membrane potential by the input, leaving a relative gap that is propagated upwards. Then, assemblies of neurons that are characterized by increased membrane potentials form when the recurrent input builds up. Finally, the system relaxes and the less regular spikes rebuild a more equally distributed picture until no further spikes are generated.

Figure 6

Activation traces over time in the different layers of the GMNP in the voicekey paradigm described in Section 2.2. Different colors correspond to different variables in the respective layer. A few traces are to be highlighted: solid blue lines in both the semantic and the action layer correspond to the respective threshold variable, black in the episodic memory layer denotes the strength of the memory trace, yellow is the uncertainty region for the old-new signal which is drawn in orange. The model is in classical episodic-retrieval mode, see Section 3.1. Targets are boosted and the entire episode retrieved. Retrieval is apparent in the plots by the re-rise of formerly active variables.

Figure 7

Activation traces over time in the relevant layers of the GMNP in the comparison paradigm. For coloring see Figure 6. The model is tuned to distractor inhibition mode, see Section 3.1. Two different conditions are shown: DTs, the former target becomes the current target and the reaction switches (from no to yes in this case); and DTr, again the former distractor becomes the current target but now the reaction does not switch (yes in both prime and probe trial). This plot illustrates the difficulty of comparing theories that are developed in a different context. Distractor inhibition theory itself is not able to explain a reaction time difference between the two conditions, as it is only formulated on a semantic level. Indeed GMNP does not show a difference in the traces except in the action layer, where persistent activation and relative inhibition causes the observed effects.