Contra assertions, feedback improves word recognition: How feedback and lateral inhibition sharpen signals over noise

Abstract

Whether top-down feedback modulates perception has deep implications for cognitive theories. Debate has been vigorous in the domain of spoken word recognition, where competing computational models and agreement on at least one diagnostic experimental paradigm suggest that the debate may eventually be resolvable. Norris and Cutler (2021) revisit arguments against lexical feedback in spoken word recognition models. They also incorrectly claim that recent computational demonstrations that feedback promotes accuracy and speed under noise (Magnuson et al., 2018) were due to the use of the Luce choice rule rather than adding noise to inputs (noise was in fact added directly to inputs). They also claim that feedback cannot improve word recognition because feedback cannot distinguish signal from noise. We have two goals in this paper. First, we correct the record about the simulations of Magnuson et al. (2018). Second, we explain how interactive activation models selectively sharpen signals via joint effects of feedback and lateral inhibition that boost lexically-coherent sublexical patterns over noise. We also review a growing body of behavioral and neural results consistent with feedback and inconsistent with autonomous (non-feedback) architectures, and conclude that parsimony supports feedback. We close by discussing the potential for synergy between autonomous and interactive approaches.

Keywords: Computational models; Language processing; Neural Networks; Spoken word recognition.

Figure A1.

Results of Simulation 1. Replication of Magnuson et al. (2018) using activations instead of response probabilities. Each point represents the outcome of simulating every word in the 212-word slex lexicon, with 10 simulations conducted with each word at each noise level greater than zero. The recognition threshold was set to 0.4, which maximized accuracy with feedback and noise set to zero.

Figure A2.

Comparing activation-based recognition time with feedback (set to 0.03) and without feedback (0.0) in Simulation 1. At each noise level greater than 0, there were 2120 simulations (10 repetitions of each word with Gaussian noise added to the input). Results are plotted only for words that were correctly recognized both with and without feedback. Items classified as “faster” were recognized more quickly (reached the threshold) more quickly with feedback than without; “equal” reached threshold at the same cycle with and without feedback; “slower” were reached the threshold later with feedback than without.

Figure 1.

Comparing the complexity of purely feedforward architecture (left) with the addition of feedback (center), as in TRACE, or sublexical decision nodes (right), as in Merge (Norris et al., 2000). Either feedback connections or decision nodes are required to account for lexical effects on sublexical decisions. The decision-node architecture requires more additional nodes and connections than feedback (dashed lines). Reproduced from Magnuson (2022b).

Figure 2.

Interactive activation example. Arrows denote excitatory connections (7 input phonemes feed forward to 3 words, which send feedback to constituent phonemes). Edges with bulb connectors indicate lateral inhibition links within layers. Reproduced from Magnuson (2022a).

Figure 3.

A schematic of iterative refinement (or signal sharpening, cf. Blank & Davis, 2016) in interactive activation. (A) Given input consistent with ‘tell’, TELL becomes strongly activated while CAT and LOAD become partially activated. (B) TELL inhibits CAT and LOAD. (C) Words send feedback to constituent phonemes, with lateral inhibition at the phoneme level enhancing the advantage for phonemes that are relatively strongly activated. (D) Over subsequent cycles of excitatory and inhibitory activation flow, the activations for TELL and its phonemes are iteratively enhanced / sharpened. Note that if random noise were added to the inputs, the same refinement / sharpening would happen so long as the target word has at least slightly higher activation than other words. The most important connections are highlighted in red in each panel, with darker red indicating greater activation. Reproduced from Magnuson (2023).