Soundboard-Trained Dogs: Deliberate and Non-Random Two-Button Combinations
Soundboard-Trained Dogs: Non-Accidental, Non-Random, and Non-Imitative Two-Button Combinations
Introduction and Background
Early Interspecies Communication Attempts: Initial studies sought to teach language (vocal and gestural) to enculturated great apes.
Criticisms: These early efforts faced significant criticism due to:
Inconsistent and underreported methods.
Overinterpretation of animal behavior by caretakers and scientists.
Rearing conditions in human environments being potentially detrimental to apes.
Claims of linguistic productivity (e.g., Washoe signing "water bird" for a swan) were met with alternative explanations, suggesting imitation rather than symbolic comprehension.
Shift to Controlled Methods: Following criticisms, researchers moved to controlled laboratory experiments and expanded the field to include vocal imitators like parrots.
Augmentative Interspecies Communication (AIC) Devices: Introduced devices like lexigrams, magnetic chips, and buttons.
Advantages: Did not require vocal mimicry or fine motor skills, allowing for broader species applicability.
Improved Rigor: Allowed for greater separation between subjects and human trainers, leading to more rigorous training and data collection.
Evidence: Some animals (apes, dolphins, and professionally-trained dogs like Sofia and Laila) learned to use AIC for communicative requests by associating labels with effects.
Example: Sofia pressed keys for "play" and "walk" to request actions, and both Sofia and Laila were sensitive to human visual perspective in their communication.
Remaining Criticisms: Still susceptible to behavioral overinterpretation and the Clever Hans effect (unintentional human cueing).
Current Trend: Citizen Science with Dog Soundboards
Recent Phenomenon: Thousands of dog owners are now training their pets with button soundboards that produce prerecorded human words or phrases.
Owner Reports: Beyond single-button requests, owners report dogs using buttons for abstract concepts (e.g., "more," "later," "help") and producing recurring two-button sequences, resembling lexigram-trained apes.
Prior Research: Recent findings indicate soundboard-trained dogs can recognize some word labels, responding appropriately even without additional contextual cues Bastos et al., 2024
Research Question: This study investigates the nature of these button presses to determine if soundboard use by dogs reflects viable interspecies communication.
Predictions for Dog Button Presses
If dogs are using soundboards to communicate, their presses should be:
Non-accidental: Deliberate actions, not unintended results of other behaviors.
Non-random: Dogs do not indiscriminately press buttons simply for rewards based on trained commands.
Non-identical to owner's presses: Dogs do not merely repeat their owners' presses through social learning (e.g., stimulus enhancement or imitation).
Methods
Data Collection
Citizen Science Approach: Owners of soundboard-using pets manually reported button presses by themselves and their dogs using a custom mobile application.
Reported presses either as they occurred or from video annotations.
Owners instructed to report all presses and provided support.
No prescriptive instructions on button labels or concepts were given.
Ethical Considerations: Informed consent obtained from owners; study exempt from IRB approval under 45 ext{ CFR } 46.104( ext{d}).
Subject Inclusion: Only subjects with 200 or more reported soundboard interactions were included to ensure legitimate data and reasonable experience.
Dataset Overview (over 21 months):
Dog Pressing Events: 194,901 interactions by 152 pet dogs.
56,676 (29.08 ext{%}) were multi-button combinations.
Owner (Modelling) Events: 65,682 interactions.
Engagement: Individual dogs' presses recorded for a median of 98 days.
Median 10.9 presses per logged day.
Days with most presses (upper 25 ext{%} quartile) ranged from 17.8 to 90 presses per day.
Soundboard Customization: Button labels and layouts were decided by owners.
Concept Categorization: Labels were sorted into 68 broad, pre-determined concept categories (e.g., "kibble," "dinner," "food" grouped under "FOOD"). Concept categories are written in capital letters to distinguish from specific voice-recorded labels.
Social Learning Model
Purpose: To assess the effect of owner modelling on dog presses and determine if dogs merely repeated owner actions without regard for labels.
Model Type: Bayesian negative binomial model (using
brmsin R).Model Equation: ext{Dog Presses} ext{ ~ } ext{Modelling Events} * ext{Concept} + (1+ ext{Modelling Events} * ext{Concept} | ext{Subject})
Rationale: Negative binomial model used for count data where variance was substantially higher than the mean, violating Poisson distribution assumptions.
Randomness Index (RI) Calculation
Purpose: To determine the extent to which dogs' multi-button presses were random.
Method: RI determines non-randomness in networks based on the negative correlation between a node's Local Clustering Coefficient and its degree Meghanathan, 2017.
Network Generation: For each subject, three-or-more-button sequences were split into constituent neighboring pairs (e.g., "want" + "food" + "outside" yields "want" + "food" and "food" + "outside"). Order was disregarded. This created an undirected, unweighted combination network (buttons as nodes, combinations as links).
Comparison: RIs of real networks were compared to 1,000 randomly simulated networks with the same number of nodes and links.
Prediction: If non-random, real RIs should be at least 0.2 smaller than the average of random networks.
Statistical Test: Paired Bayesian t-test (using BayesFactor package in R).
Two-Button Concept Combination Model
Purpose: To investigate if specific button concept combinations appeared repeatedly at the population level, and if multi-button presses were non-random.
Focus: Analyzed the 16 most commonly provided concepts across all dogs' soundboards.
Hypothesis: Accidental stepping would lead to a uniform distribution of combinations, while non-accidental pressing of meaningful buttons would show population-wide preferences for certain combinations.
Model Type: Bayesian negative binomial model (using
brmsin R).Model Equation: ext{Combination Frequency} ext{ ~ } (1+ ext{Combination ID}+ ext{offset(rel prob)} | ext{Subject} + (1| ext{Combination ID}))
Partial Pooling: Random-intercepts model used due to the large number of different combinations; partial pooling benefits from shrinking estimates for groups with few data points towards the mean.
Control for Individual Probability: The model included the product of the relative probabilities of each individual button in a combination being pressed. This prevented conflating high combination frequency with individual buttons simply being pressed frequently.
Results
Dog Presses and Routine Activities
The buttons most frequently pressed by dogs typically related to their routine activities and needs (Fig. 1).
Minimal Association with Owner Presses
A mixed-effects model revealed only a minimal association between the identity of individual buttons pressed by dogs and those pressed by their owners.
eta_{ ext{modelling}}=0.014 , with a 95 ext{%} Confidence Interval of [0.011, 0.018] .
This suggests dogs do not simply repeat their owners' presses or attend to their preferred buttons.
Non-Random Two-Button Sequences
Randomness Index (RI) Findings: The RIs for dogs' real combination networks were significantly less random than those calculated for equivalent randomly generated bootstrapped networks.
Bayesian paired t-test: ext{BF}=72.73 ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ext{ } ( <0.01 ext{%} error).
Network Structure: The mode RI value for real networks was -0.84, indicating non-randomness due to a strong negative correlation between node degree and Local Clustering Coefficient.
This suggests a structured network where central "hub" nodes are surrounded by many neighboring nodes that are not interconnected, unlike random networks (average RI around 0).
Individual Variation: While most dogs showed non-random RIs, some subjects' values deviated less significantly from random expectations (Fig. 2).
Population-Level Preferences for Two-Button Concept Combinations
Even after controlling for the relative probabilities of individual button concepts being pressed, certain button combinations occurred more frequently than others at the population level (Fig. 3).
Combinations More Frequent Than Expected by Chance:
"FOOD" + "PLAYALL" (eta=0.423, CI{2.5}=0.106, CI_{97.5}=0.732)
"GO OUTSIDE" + "OTHER" (eta=0.398, CI{2.5}=0.085, CI{97.5}=0.705)
"HELP" + "OTHER" (eta=0.357, CI{2.5}=0.077, CI{97.5}=0.609)
Combinations Less Frequent Than Expected by Chance:
"LATER" + "LOVE YOU" (eta=-0.442, CI{2.5}=-0.805, CI{97.5}=-0.100)
"NOW" + "WANT" (eta=-0.461, CI{2.5}=-0.792, CI{97.5}=-0.135)
Specific Examples (Fig. 3): "FOOD" + "TREAT" and "OWN NAME" + "WANT" were common, with "OWN NAME" + "WANT" being frequent despite its constituent buttons being among the least commonly pressed individually.
Discussion
Key Findings
Deliberate and Non-Random Presses: Owner-trained dogs use soundboards to make deliberate, non-random button presses.
Not Mere Repetition: These presses are not simply repetitions of their owners' presses.
Population-Level Preferences: Soundboard-trained dogs, at the population level, produce certain two-button concept combinations more often than others, despite individual variations in soundboard layouts.
Implication: These findings suggest dogs differentiate between soundboard buttons and, given the emergence of specific combinations, associatively ascribe different meanings to them.
Supporting Arguments Against Alternative Explanations
Against Owner Cueing/Prompting: The large number of dog pressing events relative to owner modelling events (194,901 vs. 65,682) makes individual cueing unlikely.
Against Random Pressing for Reward: If presses were merely trained commands for rewards, dogs would press randomly, ignoring labels or positions. The non-randomness and preferred combinations contradict this.
Against Indirect Reinforcement: It is unlikely that all owners across diverse households would consistently respond to and interpret the same (random) combinations in a way that would reinforce these specific patterns.
Against Simple Imitation: The finding that owner presses do not meaningfully predict dog presses for the same category indicates that dogs are not simply imitating preferred buttons.
Limitations and Future Directions
Experienced Dogs Focus: The study focused on soundboard-experienced dogs ( >200 interactions). Social learning may still play a primary role for novice users.
Owner Training: Owners often train by "modelling" (demonstrating outcomes and performing actions). Anecdotally, modelling decreases with dog competence. Ongoing work investigates the extent, effectiveness, and decline of modelling.
Comprehension of Labels: The issue of whether dogs truly comprehend the labels (i.e., establish reference) is beyond this paper's scope but is an ongoing investigation.
Prerequisite: Proving that dogs consistently and correctly associate buttons with relevant outcomes (e.g., "food" leads to bowl filling) is critical but not sufficient for demonstrating full communication.
Contextual Responses: Previous work has shown soundboard-trained dogs perform contextually appropriate behavioral responses to labels Bastos et al., 2024.
Individual Variation in Comprehension: The observed variability in network randomness suggests different levels of comprehension among dogs. Future work should explore if learning is universal or limited to a subset of "gifted word-learner" dogs (similar to dogs learning hundreds of object names).
Larger Sample Sizes: Future studies with more subjects could investigate predictive traits like breed, age, or training background for soundboard engagement or communicative use.
Potential Reporting Bias: Owners might report "interesting" presses more often. While this could alter individual dog data, it is less likely to systematically bias population-level trends given diverse soundboards and interpretations.
Mitigation: Current work is gathering continuous video/audio recordings for a subset of participants, annotated by trained researchers to ensure complete records.
Investigating Communicative Intent: Future studies should examine dogs' behaviors during soundboard use to infer communicative intent, such as:
Not barking simultaneously with presses (to be heard).
Social referencing their owners after presses.
Matching body language to contextually appropriate presses (e.g., play-bowing for "play" buttons).
Conclusion
The results indicate owner-trained dogs can press soundboard buttons in a non-accidental and non-random manner, and do not merely repeat owner presses.
Dogs differentiate between buttons, and the emergence of specific two-button concept combinations at the population level suggests dogs associatively ascribe meanings to different buttons.
The observed patterns across a large number of dogs propose that soundboard use is deliberate, warranting further investigation into dogs' pressing behaviors and comprehension.