Lexical Feedback in the Time-Invariant String Kernel (TISK) Model of Spoken Word Recognition

1.1. The problem of sequence encoding

Sequence encoding is a fundamental challenge for models of spoken word recognition; speech unfolds over time, and representing phonological word forms entails representing temporal order (CAT vs. TACK, i.e., /kæt/ vs. /tæk/) and repeated elements (SOUL vs. SOLO, i.e., /sol/ vs. /solo/). To illustrate this challenge, consider the simple network in Figure 1. Here, the only connections are forward ones from phoneme nodes to word nodes. Note that such a network cannot encode temporal order. Any word node receiving input from /k/, /æ/, and /t/ in any order (i.e., ACT /ækt/, CAT /kæt/, TACK /tæk/, or nonwords /tkæ/, /ktæ/, or /ætk/) would be equally activated by any ordering of the three phonemes. Neither could such a network distinguish words with the same constituent phonemes but differing in repeated elements (SOUL vs. SOLO). The second /o/ in /solo/ would simply be more evidence that /o/ had occurred; the network cannot represent two instances of /o/ in different temporal positions.

Figure 1

Note that a model like this could be used to investigate many aspects of word recognition. In fact, the Merge model (Norris et al., 2000) has this structure (as well as lateral inhibition), and can simulate many important aspects of spoken word recognition, despite being unable to encode order or repeated elements. Avoiding these challenges can only be a temporary simplifying assumption, however. Ultimately, models of spoken word recognition must grapple with the representation of order and repeated elements.

The TRACE model (McClelland & Elman, 1986) takes an innovative approach to the problem. TRACE translates time to space, by creating time-specific duplicates of feature, phoneme, and word nodes. A template for CAT is maximally activated by strongly activated /k/, /æ/,² and /t/ phonemes aligned with a word node standing for CAT.

Figure 2 contains a schematic outlining the complex relationships between feature, phoneme, and word nodes in TRACE. At the bottom of the figure, black cells stand in for the distributed vector of pseudo-spectral representations used as TRACE inputs. Their horizontal extent represents their temporal extent. Although feature patterns for adjacent phonemes overlap in TRACE (providing a coarse analog to coarticulation), for the sake of simplicity, we do not attempt to depict that overlap here.

Figure 2

At each time step t in a TRACE simulation, pseudo-spectral input patterns are applied. Feature nodes aligned with input slice t (that is, time-specific feature nodes) are activated by the bottom-up input at time t. Subsequently (from step t + 1 onward), bottom-up input is not applied at slice t. However, feature detectors aligned at slice t that were activated by input continue to be active for many time steps, because their activations are a summative combination of their bottom-up input and previous activation. The latter is scaled by a decay parameter, such that a unit’s activation will eventually diminish to a defined baseline level in the absence of new input. Similarly, phoneme nodes are aligned at specific time slices, and receive input from feature nodes aligned with them in time. As long as the aligned feature nodes are active, the aligned phoneme nodes will receive bottom-up input. Phoneme nodes’ activations are a function of bottom-up input and decay-scaled prior activation, as well as lateral inhibition from other phoneme nodes with which they overlap in time, and lexical feedback (described below). Phoneme nodes send bottom-up activation to nodes corresponding to words containing them that are aligned (at least partially) in time with the phoneme node. Word nodes also send feedback to phoneme nodes that send them bottom-up input. As “time” progresses in a TRACE simulation, inputs aligned with specific time points activate aligned features, phonemes, and words. This time-specific “reduplication” strategy – aligning copies of each feature, phoneme, and word in memory with specific time points – allows TRACE to represent temporally ordered sequences, including sequences with repeated elements. Thus, given the input /dæd/ (DAD), the first and second instances of /d/ would activate independent /d/ nodes.

This reduplication strategy is frequently criticized. Indeed, McClelland and Elman (1986) discussed plausibility concerns (p. 77). Some have argued that this scheme is simply implausible (e.g., Grossberg & Kazerounian, 2011; Norris, 1994), largely because of the numbers of nodes and connections it would take to implement a realistic phoneme inventory and lexicon. Magnuson (2015) presents a case for the TRACE architecture as a kind of echoic memory. Hannagan et al. (2013) estimate how many nodes and connections a realistically-sized version of TRACE would require, and estimate that a version with 40 phonemes and 20,000 words would require ~1.3 million nodes and more than 40 billion connections. Given estimates that the human brain contains approximately 86 billion neurons and 150 trillion synapses (Azevedo et al., 2009), it is not clear that we can rule out the TRACE solution based on intuitions about the plausibility of numbers of units and connections required. However, it does raise the question of whether a more compact representation might be possible, which leads us to a discussion of the TISK model.

1.2. Origins and innovations of TISK

The idea of TISK originally came from discussions between Jonathan Grainger and TH, and eventually included JM. The aim was to keep the explanatory power of the TRACE model while dispensing with its duplicated time-specific units. Hannagan et al. (2013), inspired by models of visual word recognition developed by Grainger and others using open bigram codes (Whitney, 2001; Grainger & van Heuven, 2003; Dehaene et al., 2005), asked whether a simpler interactive activation model of spoken word recognition could be implemented with a variant of open diphone coding. Open diphones are adjacent or non-adjacent phoneme pairs that occur in a string. For example, the (ordered) open diphones of ACT (/ækt/) are /æk/, /kt/, and /æt/ (see Table 1 for several more examples). As it turns out, such lists are highly distinctive. To encode the lists in a length-independent fashion, we can create a phoneme × phoneme matrix (corresponding to all possible diphones),³ and simply enter the count of each diphone for a word. This then is a kind of string kernel⁴ for words: we can manipulate or compare representations of words of any size through vector/matrix operations (i.e., the operations are identical since they are computed over matrices).

TISK’s architecture is presented schematically in Figure 3. Time-specific phoneme input nodes feed to time-invariant N-phone nodes (corresponding single phone and diphone nodes), but via what Hannagan et al. dubbed a symmetry network (in recognition of prior work on the topic by Shawe-Taylor, 1993). The symmetry network does not activate all open diphones equally. It privileges ordered diphones and activation is inversely proportional to distance between diphone members (e.g., /st/ would be less activated by SPOT than STOP). This followed work by Dandurand, Hannagan and Grainger (2013) showing that weight gradients can emerge in models of visual word recognition trained to be invariant to the location of the word input on a simulated retina. It also built on work by Hannagan and Grainger (2012), who noticed the similarity between N-gram schemes for visual word recognition, and a versatile technique called “string kernels” that has been used in text classification (Lodhi et al. 2002) and computational biology (Leslie & Kuang, 2004). Building on these two strands of work, the TISK symmetry network uses weight gradients as well as gating connections to accurately activate N-phone nodes, even in the presence of repeated phonemes. For more details about the symmetry network and TISK more generally, see Hannagan et al. (2013). Note that the full code for TISK is freely available (You & Magnuson, 2018; https://github.com/maglab-uconn/TISK1.0; also, updated code from this project is also available at https://github.com/maglab-uconn/TISK_FEEDBACK).

Figure 3

TISK thus may be viewed as a potential successor to or extension of TRACE that addresses the critique of time-specific nodes. However, TISK has not been tested on the entire broad range of results that TRACE accounts for (Magnuson & Crinnion, 2022). Hannagan et al. (2013) focused on a subset of particularly critical phenomena in spoken word recognition (the time course of phonological competition, and the relations between a variety of lexical dimensions and recognition time in TISK vs. TRACE) to establish initial plausibility of the model. However, they did not consider a broad class of phenomena in spoken word recognition that have particular relevance for ongoing theoretical debates: apparent top-down lexical effects. Our primary goal here is to address this gap.

1.3. Feedback and theories of spoken word recognition

A particularly salient point of disagreement in theories of spoken word recognition concerns top-down feedback from words to sublexical representations. TRACE (McClelland & Elman, 1986) is an interactive-activation model with arguably the deepest and broadest coverage of spoken word recognition (cf. Magnuson et al., 2012, Magnuson & Crinnion, 2021). Top-down effects in TRACE emerge from lexical-phonemic feedback. In contrast, Norris, McQueen, and Cutler (2000; see also 2016) have argued that purely feedforward systems can do anything a feedback system can do, so long as they include a mechanism for post-perceptual behavior consistent with top-down influences (e.g., via post-lexical integration of phonemic input and lexical knowledge). As TISK is a derivative of TRACE, our goal here is to make TISK more comparable to TRACE and assess the possibility of adding feedback to TISK. Without feedback, top-down effects are out of scope for TISK. We will return briefly to theoretical disagreements concerning feedback in the Discussion.

Consider two important top-down effects in spoken word recognition. First, there is the Ganong (1980) effect, where phoneme identification is influenced by lexical status. For example, compared to a nonword continuum between iss and ish, where participants are asked to identify the final consonant, identification shifts towards /s/ if the continuum is instead between a word and nonword pair like kiss-*kish, but towards /∫/ given *fiss-fish. Thus, either lexical context modulates phonetic perception (the interactive or feedback assumption), or it has a post-perceptual influence on responses (the feedforward assumption). Another fundamental top-down effect in spoken word recognition is phoneme restoration (Samuel, 1981a, 1981b, 1996, 1997; Warren, 1970). If a phoneme in a word is replaced by silence, it leaves a salient gap, and participants have no trouble reporting that the word is not intact and can identify which phoneme is missing. In contrast, when a phoneme is replaced by noise, participants typically report that the word is intact but has noise added to it. They have difficulty specifying which phoneme the noise is aligned with, and report perception consistent with lexical context (e.g., if noise, denoted as #, replaces a phoneme in the word after, the noise is heard as /t/ in /æf#^r/ but as /f/ in /æ#t^r/). This implies that noise provides enough bottom-up support for the missing phoneme to be filled in, either perceptually via lexical feedback or via post-perceptual lexical integration.

While such top-down effects are quite salient, a less obvious benefit of feedback is to make models more robust to noise. Top-down feedback (in concert with lateral inhibition in TRACE; Magnuson et al., 2024) promotes accuracy and faster processing given noisy inputs (Magnuson et al., 2018). While more subtle, this may be the more important impact of feedback.

2.1. Simulation 1: Time course and Lexical Dimensions

2.1.1. Parameters

We used a trial-and-error process for parameter exploration. We began with a value of positive feedback from words to their constituent N-phones. We assessed mean accuracy over the 211-word (original TRACE) lexicon, and if accuracy was lower than approximately 80%, we examined errors for clues as to what was impeding accuracy. If we found a parameter setting that would allow reasonable accuracy, we then examined the model’s ability to simulate top-down effects (with phenomena like those discussed below). If feedback was not strong enough for plausible top-down effects, or if error patterns implied parameter changes were needed, we would adjust parameters and retest. We iterated this process, gradually increasing our accuracy threshold.

After a few iterations, we determined that there were three key parameters that could be adjusted to provide the full complement of desired outcomes (a–e above). First, of course, we needed positive feedback from words to constituent N-phones. Second, feedback tended to cause resonance between word and N-phone layers that would lead to the activation of too many words. For example, given the input /dal/ (DOLL), the lexical node for DOLL would send feedback to /d/, /a/, /l/, /da/, /dl/, and /al/ nodes at the N-phone level. These would enhance activation of doll, but also any word containing any of these elements (e.g., SADDLE and DRILL would contain /dl/), allowing them to send feedback to elements that had not occurred. We discovered that we could avoid “runaway” activation by both increasing decay at the N-phone level and by including a small amount of negative feedback to a word’s non-constituents (i.e., a small amount of inhibition to every N-phone or single phone that is not part of a word, similar to top-down inhibition in early interactive activation models, e.g., McClelland & Rumelhart, 1981). Table 2 lists key parameters we considered altering, with the three parameters that were ultimately altered in bold font. We have not searched the parameter space exhaustively. However, our explorations suggest that stable performance requires ratios among parameters similar to those in Table 2.

Table 2

Original (without feedback) parameters for TISK, and parameters that promote high performance with feedback. Parameters in the ‘optimized without feedback’ column that differ from original parameters are in bold. Parameters in the ‘optimized with feedback’ column that differ from parameters in the ‘optimized without feedback’ and/or ‘original TISK’ columns are also in bold.

PARAMETER	ORIGINAL TISK	OPTIMIZED WITHOUT FEEDBACK	OPTIMIZED WITH FEEDBACK
Input phoneme decay	0.010	0.001	0.001
N-phone decay	0.001	0.001	0.100
Word decay	0.010	0.050	0.050
Phoneme to N-phone	1.000	0.100	0.100
Diphone to word	0.050	0.050	0.050
Single phone to word	0.010	0.010	0.010
Word to word inhibition	–0.005	–0.005	–0.010
Positive word to N-phone feedback			0.150
Negative word to N-phone feedback			–0.050

We also considered that the original TISK parameters might not provide the best possible performance in noise without feedback. We therefore explored the parameter space without feedback with the aim of finding parameters that would allow the model to continue to exhibit fundamental target behaviors described below while maximizing performance in noise. We present details of our parameter space exploration for models with and without feedback in Appendix 1. For now, because the most robust parameters for the model without feedback differ from the original TISK parameters, we will present results in the following simulations using the new parameter set (while noting that the original TISK model and the version with feedback and parameters optimized for graceful degradation differ only slightly and qualitatively in the following simulations – with the exception, of course, of the final graceful degradation simulations).

Before turning to top-down effects, let us consider whether TISK performs similarly with and without feedback on the tasks evaluated by Hannagan et al. (2013). Figure 4 addresses this by first examining mean activation over time for different categories of potential phonological relatives. To conduct this comparison, we conducted 211 simulations with TRACE and with two versions (with and without feedback) of TISK. For each model, there were 211 simulations (one for each word in the original TRACE lexicon). For every target word, we tracked target activation over time, as well as the mean activation of every item in two categories of phonological relatives (cohorts and rhymes) over time (e.g., for /dal/, the activation of every word beginning /da/ would be included in the [onset] cohort mean, and every three-phoneme word ending in /al/ would be in the rhyme category). If a word had no relatives in a category, it would not contribute to the mean for that category. As a baseline reference, we simply tracked the mean activation of all words; given 211 words, this mean approaches the minimum possible activation value. Although the mean values are somewhat damped when feedback is added to TISK, the crucial consideration is that the rank ordering of competitors is similar for all three models.⁵

Figure 4

Figure 5 extends our examination of how similar the performance of TISK is (with and without feedback) to TRACE by comparing item-specific recognition times (RTs) for each model. Recognition time was operationalized as the cycle at which the target word exceeded all other word’s activations by at least 0.05 and then continued to exceed all others by that amount for at least 10 cycles (cf. Hannagan et al., 2013), and subsequently remained the most activated word until the end of the simulation. Mean accuracies were 100% for TRACE, 99% for the original TISK without feedback (TISK), and 97% for TISK with feedback (TISKfb). As can be seen in Figure 5, item-specific RTs for correctly recognized items were remarkably similar for the three models.

Figure 5

Figure 6 goes deeper by examining how item-specific RTs in the three models (plus a fourth variant: TISK without feedback with parameters optimized for accuracy in noise, as described in Simulation 5) relate to several lexical dimensions: word length (in phonemes), number of embeddings (words embedded in the target, e.g., CAT has AT embedded within it), number of cohort (onset) competitors (overlapping in the first two phonemes), number of “ex-embeddings” (words the target embeds within, e.g., CAT embeds within CATALOG), number of “DAS” neighbors (i.e., words differing from the target by a single phonemic deletion, addition, or substitution; Luce & Pisoni, 1998), and number of “rhyme” items (words differing from the target only in first position, whether by deletion, addition, or substitution; e.g., CAT’s rhymes include SCAT, BAT, MAT, SAT, and AT). The dimensions are ordered according to the sign and magnitude of their prediction on RT; longer words are recognized more slowly, having more embeddings or cohorts is associated with slower RT, and having more ex-embeddings, neighbors or rhymes is associated with faster RT. The potential reasons for these relationships is beyond the scope of this paper (Magnuson, in preparation, discusses this in detail); our focus is instead the similarities between models. All models show the similar patterns, and are even generally similar in the strength of each correlation (although TISK without feedback with parameters optimized for performance in noise [third row] differs for ex-embeddings, neighbors and rhymes, as we discuss in Section 2.5).

Figure 6

The results from Simulation 1 demonstrate that we can add feedback to TISK without disrupting the model’s similarity to TRACE. The time course of different kinds of phonological competition are quite similar, and TISK retains its high similarity to TRACE in item-specific RTs with feedback on, and there are only very subtle quantitative differences in item-specific RTs between TISK with and without feedback apparent in our examination of how a variety of lexical dimensions relate to recognition time. With this fundamental consideration of prior results resolved, we can turn to the details of specific top-down effects.

2.2. Simulation 2: Ganong effect

For Simulation 2, we compared the ability of TISK with and without feedback to simulate the Ganong effect (Ganong, 1980). In the Ganong paradigm, we begin with a continuum from one phoneme to another (e.g., changing gradually from /s/ to /∫/, e.g., ess to esh) and establish a baseline identification pattern across the continuum (e.g., rate of “s” [vs. “sh”] responses at each step). If we add context such that the continuum changes from a word to a nonword (e.g., from bus /b^s/ to *buhsh /b^∫/, or from *russ /r^s/ to rush /r^∫/), human listeners’ identification rates will change. Specifically, they will make more responses consistent with the lexical endpoint, typically shifting the category boundary away from the lexical endpoint (e.g., for /b^s to /b^∫/, they will make more “s” responses, and the shift to “sh” responses will happen closer to the unambiguous /∫/ endpoint).

To simulate the Ganong effect with TISK, we selected ten 4-phoneme words from the lexicon (appeal, box, boost, duty, greet, least, school, shield, screw, and ugly). For each word, we conducted a Ganong simulation at each position by creating nonwords replacing the phoneme at the critical position, and then creating a continuum blending from the original phoneme to the replacement. For example, for /duti/ (duty), we created continua between /duti/ and four nonwords: /buti/, /d^ti/, /duri/, and /dut^/. So in Figure 7, for the Position 4 panel, the relevant simulation for /duti/ would be the /duti/ to /dut^/ continuum. We aggregated results by averaging activations of the lexically-consistent phoneme and its nonword replacement, and calculating ‘predicted proportion of choices’ as the ratio of the peak activations for those two phonemes. In Figure 7, we observe robust Ganong effects (lexical restoration) at each position in the model with feedback, with stronger effects in later positions (consistent with TRACE simulations of phoneme restoration in TRACE reported by Magnuson, 2015). This increase of the effect at later positions is a result of greater lexical activation as more bottom-up input is received. Thus, feedback allows TISK to simulate the Ganong (lexically-driven phoneme restoration) effect.

Figure 7

2.3. Simulation 3: Retroactive effects of feedback

In Simulation 3, we focus on retroactive influences of lexical feedback on the activation of phonemes given ambiguous input, where the disambiguating lexical context only emerged at the final phoneme (so-called right-context effects; see simulations described by McClelland & Elman [1986] on the following pages for related results: pp. 27, 29, 30 [their Figures 8, 11]; pp. 66–69). For this simulation, we used the lexical items plug and blush. If we replace the onsets of these items with a stimulus halfway between /p/ and /b/ (denoted by /#/), we create an ambiguity that will be sustained until the final phoneme is presented. We conducted simulations where the inputs were either the clear lexical inputs /pl^g/ or /bl^∫/ to establish baseline activations for /p/ and /b/ (we added blush to the TRACE lexicon for this simulation; note also that plush was not in the lexicon). Then we conducted simulations where the input was /#l^g/ (disambiguated as plug at the final phoneme) or /#l^∫/ (disambiguated as blush at the final phoneme).

The results are plotted in Figure 8. Left panels show results with the _lug context (either /pl^g/ when it is intact, or /#l^g/ when it is ambiguous); right panels show results with the _lush context (/bl^∫/ when it is intact, or /#l^∫/ when it is ambiguous). Upper panels show results with feedback; lower panels show results without feedback. In each panel, we plot activations for /p/ and /b/ given intact vs. ambiguous inputs. Without feedback (lower panels), lexical contexts have no effects, and the ambiguous stimuli drive equivalent activations of /p/ and /b/. With feedback (upper panels), the initial phase of activation is identical for both contexts because it is driven purely by the bottom-up input. As more context arrives, we see changes primarily in diminished decay of lexically-consistent phonemes (e.g., ambiguous /p/ in the upper left panel). However, the effects are different for the two ambiguous contexts, with differences emerging around cycle 20. The initial differences are stronger activation of /p/ than /b/ prior to disambiguation. This occurs because there are more words that begin with /p/ than /b/ in the model’s lexicon. The effects are stronger for the _lush context; this emerges because there are 4 items with the diphone /^S/ in the lexicon, but 7 with /^g/. Since the items activated by feedback will compete with the ambiguous onset position, having a smaller number of items sharing the pattern leads to greater ultimate activation. Thus, Simulation 3 shows clear retroactive effects of feedback.

Figure 8

2.4. Simulation 4: Phoneme restoration

In Simulation 4, we turn to another classic top-down effect using an analog to the phoneme restoration paradigm (Samuel, 1981a,b, 1996, 1997; Warren, 1970). In a phoneme restoration paradigm, a phoneme is replaced either with noise or with silence (typically in a lexical context where there is only one possible completion for the replaced phoneme, e.g., #uxury or _uxury [where # indicates noise and _ indicates silence] can only be restored as luxury). The two kinds of replacement yield very different effects. If a phoneme is replaced by noise, a listener typically reports hearing all the phonemes in the word, and will likely have difficulty identifying the precise location of the noise. If a phoneme is replaced by silence, the gap is salient, and listeners can report the precise location of the silence and which specific phoneme is missing. Another difference is that noise-replaced phonemes can drive selective adaptation (Samuel, 1997), as though the actual phoneme had been repeated, but silence cannot. The interpretation of this pattern is that noise provides sufficient bottom-up activation that the missing phoneme is “filled in” by feedback. As a result, the listener not only cannot reliably report which phoneme has been replaced, but is uncertain of the position of the noise. This means that the critical pattern a model must be able to simulate is (a) robust activation of a lexically-consistent phoneme when it is replaced with noise, but (b) weak or absent activation when it is replaced with silence (see Grossberg & Kazerounian, 2011, 2016 and Magnuson, 2015, for a debate about how phoneme restoration should be modeled).

To test TISK’s ability to simulate phoneme restoration with and without feedback, we used the same ten 4-phoneme words from Simulation 2. For each item, we conducted 48 simulations; 2 models (feedback of no feedback) × 4 phoneme positions × 6 input types (intact phoneme, silence replacement [the phoneme replaced by zero input], or noise replacement [silence plus noise with standard deviation of 0.2, 0.3, 0.4, or 0.8]). We examined the activations of the “expected” phonemes each position (e.g., /d/, /u/, /t/, and /i/ for duty) when they were intact versus when they were replaced with silence or increasingly strong noise. Again, in a successful simulation, replaced phonemes should be robustly activated given sufficient noise input, but should be activated weakly or not at all given replacement with silence.

The results are shown in Figure 9. First, consider the results without feedback (top row). There is no activation whatsoever of the replaced phoneme given silence replacement or noise with standard deviation of 0.2, and slightly graded activations given noise, very near the level of activation observed for intact phonemes.⁶ In contrast, large lexical effects are readily apparent with feedback (bottom row). Noise with SD greater than 0.2 drives robust activation of the ‘expected’ phoneme on average, but activations from noise are much lower than activations from intact phonemes. Thus, TISK with feedback generates a plausible pattern of results that are transparently linkable to results with human subjects.

Figure 9

2.5. Simulation 5: Graceful degradation

The obvious impact of including feedback in a model is that it can provide a mechanism for simulating (and explaining) top-down effects. A less obvious but crucial consideration is that feedback promotes graceful degradation: gradual rather than catastrophic declines in performance given noise or parameter changes (see Magnuson, Mirman, Luthra, Strauss & Harris, 2018, for extended discussion as well as demonstrations that feedback in the TRACE model promotes higher accuracy and faster word recognition given noisy inputs). We tested TISK with and without feedback for graceful degradation with series of full-lexicon simulations (that is, one simulation for every word in the original 211-word TRACE lexicon) while gradually increasing the amount of Gaussian noise added to input patterns. At each of 15 levels of noise (SD 0.01 to 0.15 in steps of 0.01), we conducted 15 full-lexicon runs (with SD > 0, the noise would vary and therefore performance might as well; multiple runs allow us to establish more stable performance estimates).

However, there is no reason to suspect that the default TISK parameters represent the best possible performance without feedback; these parameters were originally chosen without any consideration for performance under noise. To ensure we were putting the autonomous (no feedback) and feedback versions of TISK on maximally equal footing, we explored the parameter space more fully both with and without feedback. The details of these parameter space explorations are presented in Appendix 1. These explorations led to the ‘optimized’ parameters with and without feedback presented in Table 2 above.

We present results in Figure 10 for accuracy and recognition time. With optimized parameters, TISK exhibits graceful degradation with or without feedback; that is, with a gradual decline in accuracy as noise increases, rather than a collapse (as we see for the original parameters without feedback). However, we do observe a significant advantage from feedback in terms of accuracy.

Figure 10

It is also notable that the optimized feedforward variant of TISK differs markedly from the other models in Figure 6, where we plot model RTs relative to various lexical dimensions. Specifically, it shows weaker associations with Neighbors and Rhymes, and a reversed relationship with ex-embeddings. We have not attempted to determine why this model differs from the others in these ways, as we expect the theoretical gain from such inquiry would be slight at best.

We noted earlier that Magnuson et al. (2018) conducted similar explorations with TRACE. Magnuson et al. (2018) observed catastrophic degradation for TRACE without feedback, and graceful degradation with feedback. They also observed a recognition time advantage for feedback even without noise (see Magnuson et al., 2024, for a replication using raw TRACE activations rather than response probabilities). Curiously, as can be seen in the right panel of Figure 10, recognition times in TISK tend to be longer with feedback until we reach the highest levels of noise. What might explain this difference? The most notable difference is that the default parameters for TRACE were optimized for running the model with feedback. When Magnuson et al. compared TRACE with and without feedback, it was a matter of removing feedback from the feedback-optimized parameters. We took a different tack here, in terms of finding maximally robust parameters without feedback. A question for future research is whether better performance might be possible with TRACE without feedback.

However, another possibility is that the RT differences could be related to the accuracy differences. Specifically, the words that the model with feedback is able to recognize but are not recognized by the model without feedback could be particularly challenging items, and that could substantially raise the mean RT for the model with feedback. To investigate this possibility, we restrict the means to only include words that both models (with and without feedback) recognize. The results are showing in Figure 11. Here we see a more modest disadvantage for feedback, and a much earlier switch to a feedback advantage (when noise > 0.6).

Figure 11

To probe this further, we created scatter plots for one model run (Figure 12) and all 15 runs combined (Figure 13). These plots only include points for words that were recognized by both models at a particular noise level on a specific run. Each panel also includes annotations indicating how many valid cases (i.e., recognized words) there were with and without feedback, how many valid pairs could be plotted (those are words that were recognized by both models at that level of noise), and what proportion of words were recognized more quickly with or without feedback. Red squares in each panel plot mean RT for the two models. This mean value tends to be very close to the identity line even when RT is lower without feedback for a majority of items. Eventually, when noise level reaches 0.07, the mean point rises above the identity line (indicating faster mean RT with feedback than without). Compare this to Figure 10, where the RT lines cross later (at noise = 0.11). Thus, while the impact of feedback on RT is more modest than Magnuson et al. (2018) observed with the TRACE model, it seems that the tendency for higher RTs with feedback in Figure 10 is largely driven by the more challenging words that the model with feedback is able to continue recognizing at higher noise levels.

Figure 12

Figure 13