B.J.W. Dudzik
info
Please Note
<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
22 records found
1
Automatic speech recognition (ASR) remains less accurate for children than for adults. This difference is poorly understood in multilingual systems. This thesis examines whether adding sparse Mixture-of-Experts (MoE) layers to a shared FastConformer-CTC encoder can narrow the adult-child word error rate (WER) gap. It also tests whether experts specialise by age, even without age labels in the training data. Dense and MoE encoders were initialised from the same pretrained checkpoint and fine-tuned in two stages: first on adult speech, then on mixed adult-child speech. Experiments covered English, Dutch, German, and Polish. Paired speaker-level bootstrap tests compared WER gaps. In the English-only experiment, MoE reduced the gap from 22.33 to 19.97 percentage points after adult-only fine-tuning and from 11.05 to 10.69-10.78 points after child-speech adaptation. It also lowered adult and child WER. Part of this advantage emerged before any child speech audio was used, suggesting an architectural contribution beyond data exposure. In the multilingual experiment, the benefit did not generalise uniformly. MoE reduced the gap significantly only for English. Dutch showed a smaller, non-significant reduction. German showed an increase in child WER and a significant widening of the adult-child gap, despite improved adult WER. The Polish results were difficult to interpret, as the corpus contained only isolated words and children with sigmatism. A routing analysis of the English models found age-correlated differences in early and late encoder regions after child-speech fine-tuning. The experts that adult and child speech were routed to produced distinct representations. An English domain-control analysis did not reproduce this pattern. Language was the dominant routing factor, and age-related specialisation was layered on top. Overall, a single shared MoE encoder can reduce the adult-child gap and develop age-related specialisation without demographic labels at inference time. Its effect depends on the coverage and domain match of the child speech data in each language. This makes the sparse MoE encoder a practical direction for more inclusive speech recognition: one model that reduces the adult-child gap without age labels, and whose routing can be examined to explain the improvement.
...
Automatic speech recognition (ASR) remains less accurate for children than for adults. This difference is poorly understood in multilingual systems. This thesis examines whether adding sparse Mixture-of-Experts (MoE) layers to a shared FastConformer-CTC encoder can narrow the adult-child word error rate (WER) gap. It also tests whether experts specialise by age, even without age labels in the training data. Dense and MoE encoders were initialised from the same pretrained checkpoint and fine-tuned in two stages: first on adult speech, then on mixed adult-child speech. Experiments covered English, Dutch, German, and Polish. Paired speaker-level bootstrap tests compared WER gaps. In the English-only experiment, MoE reduced the gap from 22.33 to 19.97 percentage points after adult-only fine-tuning and from 11.05 to 10.69-10.78 points after child-speech adaptation. It also lowered adult and child WER. Part of this advantage emerged before any child speech audio was used, suggesting an architectural contribution beyond data exposure. In the multilingual experiment, the benefit did not generalise uniformly. MoE reduced the gap significantly only for English. Dutch showed a smaller, non-significant reduction. German showed an increase in child WER and a significant widening of the adult-child gap, despite improved adult WER. The Polish results were difficult to interpret, as the corpus contained only isolated words and children with sigmatism. A routing analysis of the English models found age-correlated differences in early and late encoder regions after child-speech fine-tuning. The experts that adult and child speech were routed to produced distinct representations. An English domain-control analysis did not reproduce this pattern. Language was the dominant routing factor, and age-related specialisation was layered on top. Overall, a single shared MoE encoder can reduce the adult-child gap and develop age-related specialisation without demographic labels at inference time. Its effect depends on the coverage and domain match of the child speech data in each language. This makes the sparse MoE encoder a practical direction for more inclusive speech recognition: one model that reduces the adult-child gap without age labels, and whose routing can be examined to explain the improvement.
Reminders are often used to support people's prospective memory (PM). Research in Human-Robot Interaction (HRI) has primarily focused on when and how a robot reminder could better help humans, while the content of the reminder itself has received less systematic attention. This research investigates whether the content of a robot-delivered reminder affects users' subjective experience. Specifically, it compares episode-grounded reminders, which reference the situation in which the PM intention was originally formed, with reminders that omit this context, measuring effects on confidence in remembering, trust in the robot, and perceived usefulness. A between-subjects experiment (N = 21) was conducted on a browser-based 2D platform that uses a cooking scenario to impose concurrent cognitive demand with four PM tasks. Participants received either episode-grounded (EE1) or plain (EE0) reminders. Results showed that EE1 reminders were associated with significantly higher confidence compared to EE0 reminders (p = .034 after correction for multiple comparisons); trust and perceived usefulness did not differ between conditions. These findings contribute a theoretically grounded design procedure and a reusable experimental protocol for episode-grounded reminder content, and suggest that grounding reminders in episode context may shape users' confidence in remembering.
...
Reminders are often used to support people's prospective memory (PM). Research in Human-Robot Interaction (HRI) has primarily focused on when and how a robot reminder could better help humans, while the content of the reminder itself has received less systematic attention. This research investigates whether the content of a robot-delivered reminder affects users' subjective experience. Specifically, it compares episode-grounded reminders, which reference the situation in which the PM intention was originally formed, with reminders that omit this context, measuring effects on confidence in remembering, trust in the robot, and perceived usefulness. A between-subjects experiment (N = 21) was conducted on a browser-based 2D platform that uses a cooking scenario to impose concurrent cognitive demand with four PM tasks. Participants received either episode-grounded (EE1) or plain (EE0) reminders. Results showed that EE1 reminders were associated with significantly higher confidence compared to EE0 reminders (p = .034 after correction for multiple comparisons); trust and perceived usefulness did not differ between conditions. These findings contribute a theoretically grounded design procedure and a reusable experimental protocol for episode-grounded reminder content, and suggest that grounding reminders in episode context may shape users' confidence in remembering.
Automatic speech recognition (ASR) systems achieve very low word error rates on standard adult speech but underperform on populations whose speech deviates acoustically from the training distribution. Dysarthric speech, in which neurological injury impairs the muscular control of speech production, is among the most challenging cases: word error rates can exceed 90% on severely impaired speakers even after in-domain fine-tuning. Mixture-of-Experts (MoE) architectures, in which feed-forward sublayers are replaced by parallel “expert” sublayers and a small gating network routes each input to a subset of them, offer a candidate mechanism for inclusive ASR: in principle, the experts could come to specialize on different speaker populations and the router could allocate them where they are needed.
This thesis investigates whether MoE delivers that promise on dysarthric ASR, using the FastConformer-CTC Large encoder fine-tuned per fold on the TORGO and UASpeech corpora across 41 leave-one-speaker-out folds. Three research questions are addressed: how standard sparsely-gated MoE compares to a fine-tuned dense baseline (RQ1); what specialization the trained experts develop (RQ2); and whether either training-time supervision or inference-time router steering can push the model toward useful specialization (RQ3).
The answer to all three questions is negative. Five standard MoE configurations match but do not beat the fine-tuned dense baseline, and a parameter-equal follow-up shows that at equal feed-forward capacity, the MoE structure is actively worse than dense by up to 2.89 pp. The trained experts develop genuine functional specialization, but the axis of specialization is encoder depth, not speaker group or phonetic content: different layers prefer different expert pairs, and within any layer, routing is roughly content-independent and group-independent. Training-time supervision toward group-aware specialization collapses at inference, as the depth-based organization overrides the supervised partition. Inference-time router steering produces real per-speaker improvements under post-hoc selection of the bias strength, but neither of two deployment-realistic mechanisms for choosing that bias recovers the gain.
The MoE structure on this task does meaningful work, but not the work the task needs: it organizes the encoder’s computation along an axis the dense baseline already covers through its per-layer feed-forward weights. Routing-input variants, cross-architecture replication, and continual-learning extensions are flagged as the most direct follow-ups. ...
This thesis investigates whether MoE delivers that promise on dysarthric ASR, using the FastConformer-CTC Large encoder fine-tuned per fold on the TORGO and UASpeech corpora across 41 leave-one-speaker-out folds. Three research questions are addressed: how standard sparsely-gated MoE compares to a fine-tuned dense baseline (RQ1); what specialization the trained experts develop (RQ2); and whether either training-time supervision or inference-time router steering can push the model toward useful specialization (RQ3).
The answer to all three questions is negative. Five standard MoE configurations match but do not beat the fine-tuned dense baseline, and a parameter-equal follow-up shows that at equal feed-forward capacity, the MoE structure is actively worse than dense by up to 2.89 pp. The trained experts develop genuine functional specialization, but the axis of specialization is encoder depth, not speaker group or phonetic content: different layers prefer different expert pairs, and within any layer, routing is roughly content-independent and group-independent. Training-time supervision toward group-aware specialization collapses at inference, as the depth-based organization overrides the supervised partition. Inference-time router steering produces real per-speaker improvements under post-hoc selection of the bias strength, but neither of two deployment-realistic mechanisms for choosing that bias recovers the gain.
The MoE structure on this task does meaningful work, but not the work the task needs: it organizes the encoder’s computation along an axis the dense baseline already covers through its per-layer feed-forward weights. Routing-input variants, cross-architecture replication, and continual-learning extensions are flagged as the most direct follow-ups. ...
Automatic speech recognition (ASR) systems achieve very low word error rates on standard adult speech but underperform on populations whose speech deviates acoustically from the training distribution. Dysarthric speech, in which neurological injury impairs the muscular control of speech production, is among the most challenging cases: word error rates can exceed 90% on severely impaired speakers even after in-domain fine-tuning. Mixture-of-Experts (MoE) architectures, in which feed-forward sublayers are replaced by parallel “expert” sublayers and a small gating network routes each input to a subset of them, offer a candidate mechanism for inclusive ASR: in principle, the experts could come to specialize on different speaker populations and the router could allocate them where they are needed.
This thesis investigates whether MoE delivers that promise on dysarthric ASR, using the FastConformer-CTC Large encoder fine-tuned per fold on the TORGO and UASpeech corpora across 41 leave-one-speaker-out folds. Three research questions are addressed: how standard sparsely-gated MoE compares to a fine-tuned dense baseline (RQ1); what specialization the trained experts develop (RQ2); and whether either training-time supervision or inference-time router steering can push the model toward useful specialization (RQ3).
The answer to all three questions is negative. Five standard MoE configurations match but do not beat the fine-tuned dense baseline, and a parameter-equal follow-up shows that at equal feed-forward capacity, the MoE structure is actively worse than dense by up to 2.89 pp. The trained experts develop genuine functional specialization, but the axis of specialization is encoder depth, not speaker group or phonetic content: different layers prefer different expert pairs, and within any layer, routing is roughly content-independent and group-independent. Training-time supervision toward group-aware specialization collapses at inference, as the depth-based organization overrides the supervised partition. Inference-time router steering produces real per-speaker improvements under post-hoc selection of the bias strength, but neither of two deployment-realistic mechanisms for choosing that bias recovers the gain.
The MoE structure on this task does meaningful work, but not the work the task needs: it organizes the encoder’s computation along an axis the dense baseline already covers through its per-layer feed-forward weights. Routing-input variants, cross-architecture replication, and continual-learning extensions are flagged as the most direct follow-ups.
This thesis investigates whether MoE delivers that promise on dysarthric ASR, using the FastConformer-CTC Large encoder fine-tuned per fold on the TORGO and UASpeech corpora across 41 leave-one-speaker-out folds. Three research questions are addressed: how standard sparsely-gated MoE compares to a fine-tuned dense baseline (RQ1); what specialization the trained experts develop (RQ2); and whether either training-time supervision or inference-time router steering can push the model toward useful specialization (RQ3).
The answer to all three questions is negative. Five standard MoE configurations match but do not beat the fine-tuned dense baseline, and a parameter-equal follow-up shows that at equal feed-forward capacity, the MoE structure is actively worse than dense by up to 2.89 pp. The trained experts develop genuine functional specialization, but the axis of specialization is encoder depth, not speaker group or phonetic content: different layers prefer different expert pairs, and within any layer, routing is roughly content-independent and group-independent. Training-time supervision toward group-aware specialization collapses at inference, as the depth-based organization overrides the supervised partition. Inference-time router steering produces real per-speaker improvements under post-hoc selection of the bias strength, but neither of two deployment-realistic mechanisms for choosing that bias recovers the gain.
The MoE structure on this task does meaningful work, but not the work the task needs: it organizes the encoder’s computation along an axis the dense baseline already covers through its per-layer feed-forward weights. Routing-input variants, cross-architecture replication, and continual-learning extensions are flagged as the most direct follow-ups.
Root cause analysis (RCA) in large-scale industrial software systems requires engineers to correlate heterogeneous logs, source code, documentation, and domain knowledge, often under time pressure. This thesis studies whether agentic AI can support this work through human-agent collaboration. In a case study at ASML, we design and evaluate TalkToYieldStar, a multi-agent system for RCA on diagnostic log dumps from wafer metrology tools. The system decomposes RCA into bounded subtasks, including log inspection, source-code search, knowledge retrieval, hypothesis formation, and report generation. Evaluation through a case study with eleven professional engineers shows that agentic AI is useful primarily as an evidence-structuring investigation aid rather than an autonomous diagnostician: 8 of 11 participants reported the agent session as more efficient, and all rated willingness for daily use positively. Participants valued support for finding evidence and offloading tiring subtasks—including in unfamiliar domains and off-hours situations—but required steering, verification, and final human judgment. Gains were conditional on case type: the hardest cases, requiring deep domain-state reconstruction, resisted both human-guided and one-shot approaches, with hallucinated reasoning appearing in both baseline systems on those cases. The findings indicate that reliable industrial RCA agents need calibrated autonomy, inspectable evidence trails, domain-state representations, clearer planning checkpoints, and reduced computational overhead.
...
Root cause analysis (RCA) in large-scale industrial software systems requires engineers to correlate heterogeneous logs, source code, documentation, and domain knowledge, often under time pressure. This thesis studies whether agentic AI can support this work through human-agent collaboration. In a case study at ASML, we design and evaluate TalkToYieldStar, a multi-agent system for RCA on diagnostic log dumps from wafer metrology tools. The system decomposes RCA into bounded subtasks, including log inspection, source-code search, knowledge retrieval, hypothesis formation, and report generation. Evaluation through a case study with eleven professional engineers shows that agentic AI is useful primarily as an evidence-structuring investigation aid rather than an autonomous diagnostician: 8 of 11 participants reported the agent session as more efficient, and all rated willingness for daily use positively. Participants valued support for finding evidence and offloading tiring subtasks—including in unfamiliar domains and off-hours situations—but required steering, verification, and final human judgment. Gains were conditional on case type: the hardest cases, requiring deep domain-state reconstruction, resisted both human-guided and one-shot approaches, with hallucinated reasoning appearing in both baseline systems on those cases. The findings indicate that reliable industrial RCA agents need calibrated autonomy, inspectable evidence trails, domain-state representations, clearer planning checkpoints, and reduced computational overhead.
Retrospective Think-Aloud Data as a Window into Performance Trust
A Human-Agent Collaborative Game Study
Trust between users and autonomous agents is dynamic, yet most measurement instruments capture only a static end-state. Think-aloud protocol offers a window into the reasoning behind trust judgements, but no prior work has examined how retrospective think-aloud (RTA) data relates to performance trust. This paper presents a user study in which 30 participants interacted with a scripted AI agent in a collaborative box-moving game, verbalised their thoughts while watching a recording of their own gameplay (RTA), and then completed the Multidimensional Measure of Trust (MDMT). A deductive content analysis grounded in the MDMT Reliable and Capable subscales produced a 16-code scheme applied to all 30 transcripts. Negative trust language in RTA tracked the game's scripted trust trajectory, peaking during the most severe failure phase, but did not relate to participants' final MDMT scores - suggesting RTA's diagnostic value lies in the timing of this language across the session, rather than its overall volume. These findings support RTA as a method for capturing within-session trust dynamics that static questionnaires cannot surface.
...
Trust between users and autonomous agents is dynamic, yet most measurement instruments capture only a static end-state. Think-aloud protocol offers a window into the reasoning behind trust judgements, but no prior work has examined how retrospective think-aloud (RTA) data relates to performance trust. This paper presents a user study in which 30 participants interacted with a scripted AI agent in a collaborative box-moving game, verbalised their thoughts while watching a recording of their own gameplay (RTA), and then completed the Multidimensional Measure of Trust (MDMT). A deductive content analysis grounded in the MDMT Reliable and Capable subscales produced a 16-code scheme applied to all 30 transcripts. Negative trust language in RTA tracked the game's scripted trust trajectory, peaking during the most severe failure phase, but did not relate to participants' final MDMT scores - suggesting RTA's diagnostic value lies in the timing of this language across the session, rather than its overall volume. These findings support RTA as a method for capturing within-session trust dynamics that static questionnaires cannot surface.
As autonomous agents become increasingly integrated into collaborative environments, tracking and calibrating human trust in real time has become more and more crucial to prevent costly outcomes like automation complacency or disuse. However, psychometric tools, such as the Multi-Dimensional Measure of Trust (MDMT), capture only a static retrospective snapshot, missing the moment-to-moment fluctuations of interaction. This study investigates whether concurrent think-aloud (CTA) data can serve as a viable real-time proxy for performance trust. A methodological validation study (N=30) was conducted, using a cooperative game-based grid-world environment where participants worked alongside a teammate programmed to fail in controlled ways across distinct interaction phases. Qualitative protocol analysis successfully mapped transcript indicators to MDMT dimensions and revealed Defiance as a novel behavioral category capturing disuse and takeover actions. Phase-based event analysis showed high temporal sensitivity, with 49.6% of negative performance remarks occurring within a 10-second window following a robot failure. Finally, convergence analysis revealed a striking asymmetry: positive verbal indicators significantly correlated with retrospective MDMT performance trust scores (r=.539,p=.004$), whereas negative verbalizations did not (p=.066), indicating that verbalized frustration captures transient emotional spikes rather than structural drops in stable trust beliefs. These findings demonstrate that while CTA cannot replace post-hoc scales, it offers a powerful, non-invasive complementary tool providing the temporal resolution necessary to understand the dynamic process of performance trust development.
...
As autonomous agents become increasingly integrated into collaborative environments, tracking and calibrating human trust in real time has become more and more crucial to prevent costly outcomes like automation complacency or disuse. However, psychometric tools, such as the Multi-Dimensional Measure of Trust (MDMT), capture only a static retrospective snapshot, missing the moment-to-moment fluctuations of interaction. This study investigates whether concurrent think-aloud (CTA) data can serve as a viable real-time proxy for performance trust. A methodological validation study (N=30) was conducted, using a cooperative game-based grid-world environment where participants worked alongside a teammate programmed to fail in controlled ways across distinct interaction phases. Qualitative protocol analysis successfully mapped transcript indicators to MDMT dimensions and revealed Defiance as a novel behavioral category capturing disuse and takeover actions. Phase-based event analysis showed high temporal sensitivity, with 49.6% of negative performance remarks occurring within a 10-second window following a robot failure. Finally, convergence analysis revealed a striking asymmetry: positive verbal indicators significantly correlated with retrospective MDMT performance trust scores (r=.539,p=.004$), whereas negative verbalizations did not (p=.066), indicating that verbalized frustration captures transient emotional spikes rather than structural drops in stable trust beliefs. These findings demonstrate that while CTA cannot replace post-hoc scales, it offers a powerful, non-invasive complementary tool providing the temporal resolution necessary to understand the dynamic process of performance trust development.
Think-aloud data for automatic trust assessment
RTA x Morality
Trust in an AI teammate is usually measured with a single post-hoc questionnaire, which records a settled judgment but reveals little about when trust shifts or why. Think-aloud protocols are a promising way to recover what questionnaires leave out, because they expose the reasoning behind a judgment rather than only its endpoint, but little work has examined how they reflect trust, and almost none how they reflect its separate dimensions. This paper asks how retrospective think-aloud (RTA) data reflects a user's morality trust in an agent. Using 30 RTA transcripts from a cooperative game with a scripted faulty agent, the study develops a coding scheme for morality-trust markers, links them to the in-game events that elicit them, and compares the result against the morality subscale of the Multi-Dimensional Measure of Trust (MDMT). Morality trust surfaces in RTA narrowly and reactively: mainly as attributions of the agent's intent, strongly skewed negative and ordered in time. Convergence with the MDMT is real but narrow, while the humanising attributions RTA captures have no questionnaire counterpart. The two thus emerge as complementary rather than interchangeable: the questionnaire records a settled judgment, RTA recovers when and why it moved.
...
Trust in an AI teammate is usually measured with a single post-hoc questionnaire, which records a settled judgment but reveals little about when trust shifts or why. Think-aloud protocols are a promising way to recover what questionnaires leave out, because they expose the reasoning behind a judgment rather than only its endpoint, but little work has examined how they reflect trust, and almost none how they reflect its separate dimensions. This paper asks how retrospective think-aloud (RTA) data reflects a user's morality trust in an agent. Using 30 RTA transcripts from a cooperative game with a scripted faulty agent, the study develops a coding scheme for morality-trust markers, links them to the in-game events that elicit them, and compares the result against the morality subscale of the Multi-Dimensional Measure of Trust (MDMT). Morality trust surfaces in RTA narrowly and reactively: mainly as attributions of the agent's intent, strongly skewed negative and ordered in time. Convergence with the MDMT is real but narrow, while the humanising attributions RTA captures have no questionnaire counterpart. The two thus emerge as complementary rather than interchangeable: the questionnaire records a settled judgment, RTA recovers when and why it moved.
How Concurrent Think-Aloud Data Reflects a User's Morality Trust in an Agent
A Think-Aloud Study of Morality Trust in Human-Agent Collaboration
Trust is a key factor in human-agent interaction, especially when users collaborate with an AI agent whose behavior is not fully predictable. Trust is often measured after an interaction using questionnaires, but these measures do not fully show how trust changes during the interaction or which agent behaviors shape the user's judgment. To address this limitation, this paper investigates whether Concurrent Think-Aloud (CTA) data can capture participants' real-time reasoning about the agent's moral trustworthiness during interaction. The study uses a modified version of the Moving Out game, in which participants collaborate with an AI agent while verbalizing their thoughts. Trust is examined using repeated in-game pop-up questionnaires, a post-interaction questionnaire, coded CTA transcripts, game logs, and screen recordings. The results show that morality-related trust is rarely expressed through formal moral terms such as "ethical'' or "principled''. Instead, participants express it through task-specific comments about whether the agent cooperates with them, ignores them, or acts in their interest. The findings also show that CTA markers align more closely with in-game trust ratings than with post-interaction questionnaire scores. Overall, the results suggest that CTA can complement questionnaire-based trust measures by capturing situated, real-time expressions of morality trust during interaction.
...
Trust is a key factor in human-agent interaction, especially when users collaborate with an AI agent whose behavior is not fully predictable. Trust is often measured after an interaction using questionnaires, but these measures do not fully show how trust changes during the interaction or which agent behaviors shape the user's judgment. To address this limitation, this paper investigates whether Concurrent Think-Aloud (CTA) data can capture participants' real-time reasoning about the agent's moral trustworthiness during interaction. The study uses a modified version of the Moving Out game, in which participants collaborate with an AI agent while verbalizing their thoughts. Trust is examined using repeated in-game pop-up questionnaires, a post-interaction questionnaire, coded CTA transcripts, game logs, and screen recordings. The results show that morality-related trust is rarely expressed through formal moral terms such as "ethical'' or "principled''. Instead, participants express it through task-specific comments about whether the agent cooperates with them, ignores them, or acts in their interest. The findings also show that CTA markers align more closely with in-game trust ratings than with post-interaction questionnaire scores. Overall, the results suggest that CTA can complement questionnaire-based trust measures by capturing situated, real-time expressions of morality trust during interaction.
Complex machine learning models, such as XGBoost, are often used on tabular data, but explaining their predictions is a non-trivial task that requires the setup and configuration of post-hoc explainers such as SHAP or LIME. These explainers need access to the model and technical expertise, which are often unavailable when a model is proprietary or only accessible through a restricted interface. Large Language Models (LLMs) could offer an alternative, as prior work has shown that LLMs can rank feature importance without access to the model’s internals, but only on small binary classification datasets, so it is unclear when their rankings agree with established explainers.
To address this gap, this thesis investigates under what conditions LLM-generated feature importance rankings agree with SHAP and LIME. To answer this, we trained XGBoost models on three datasets across different task types and dimensionalities: California Housing, Wine Quality, and Taiwan Bankruptcy Prediction. We then prompted two open-weight LLMs, Gemma 4 (31B) and GPT-OSS (120B), to rank which features are driving 100 individual predictions per dataset. For the prompting, we used five prompt templates derived from three approaches: text-driven prompts that provide feature names and task context, data-driven prompts that provide anonymized perturbed samples, and a novel mixed neighbors prompt that combines feature names with real contrastive neighboring samples. The rankings were compared against rankings from SHAP and LIME using Spearman correlation and top-k overlap.
We found that alignment is mainly dependent on the dataset characteristics and the information included in the prompt. The best setup on California Housing, mixed neighbors with Gemma, came close to the agreement between SHAP and LIME themselves (a Spearman correlation of 0.693 compared to 0.772). On Wine Quality, where SHAP and LIME are in slight agreement (0.226 Spearman correlation), most setups surpassed their agreement, with a text-driven prompt reaching 0.645 against SHAP. On Taiwan Bankruptcy Prediction, all setups stayed below the agreement between SHAP and LIME (at best 0.305 compared to 0.414), with data-driven prompts performing best. Across all three datasets, the choice of prompt template affected agreement more than the choice of LLM.
However, text-driven rankings often missed the most important feature of individual predictions, suggesting that their agreement comes largely from the LLM’s own domain knowledge rather than the sample’s feature values. Overall, LLM-generated rankings can offer a quick, rough estimate of which features are predictive, but are not able to completely replace traditional post-hoc explainers. In practice, LLMs are best suited as a first check when model access is restricted, while high-stakes decisions still require explanations computed from the model itself.
...
To address this gap, this thesis investigates under what conditions LLM-generated feature importance rankings agree with SHAP and LIME. To answer this, we trained XGBoost models on three datasets across different task types and dimensionalities: California Housing, Wine Quality, and Taiwan Bankruptcy Prediction. We then prompted two open-weight LLMs, Gemma 4 (31B) and GPT-OSS (120B), to rank which features are driving 100 individual predictions per dataset. For the prompting, we used five prompt templates derived from three approaches: text-driven prompts that provide feature names and task context, data-driven prompts that provide anonymized perturbed samples, and a novel mixed neighbors prompt that combines feature names with real contrastive neighboring samples. The rankings were compared against rankings from SHAP and LIME using Spearman correlation and top-k overlap.
We found that alignment is mainly dependent on the dataset characteristics and the information included in the prompt. The best setup on California Housing, mixed neighbors with Gemma, came close to the agreement between SHAP and LIME themselves (a Spearman correlation of 0.693 compared to 0.772). On Wine Quality, where SHAP and LIME are in slight agreement (0.226 Spearman correlation), most setups surpassed their agreement, with a text-driven prompt reaching 0.645 against SHAP. On Taiwan Bankruptcy Prediction, all setups stayed below the agreement between SHAP and LIME (at best 0.305 compared to 0.414), with data-driven prompts performing best. Across all three datasets, the choice of prompt template affected agreement more than the choice of LLM.
However, text-driven rankings often missed the most important feature of individual predictions, suggesting that their agreement comes largely from the LLM’s own domain knowledge rather than the sample’s feature values. Overall, LLM-generated rankings can offer a quick, rough estimate of which features are predictive, but are not able to completely replace traditional post-hoc explainers. In practice, LLMs are best suited as a first check when model access is restricted, while high-stakes decisions still require explanations computed from the model itself.
...
Complex machine learning models, such as XGBoost, are often used on tabular data, but explaining their predictions is a non-trivial task that requires the setup and configuration of post-hoc explainers such as SHAP or LIME. These explainers need access to the model and technical expertise, which are often unavailable when a model is proprietary or only accessible through a restricted interface. Large Language Models (LLMs) could offer an alternative, as prior work has shown that LLMs can rank feature importance without access to the model’s internals, but only on small binary classification datasets, so it is unclear when their rankings agree with established explainers.
To address this gap, this thesis investigates under what conditions LLM-generated feature importance rankings agree with SHAP and LIME. To answer this, we trained XGBoost models on three datasets across different task types and dimensionalities: California Housing, Wine Quality, and Taiwan Bankruptcy Prediction. We then prompted two open-weight LLMs, Gemma 4 (31B) and GPT-OSS (120B), to rank which features are driving 100 individual predictions per dataset. For the prompting, we used five prompt templates derived from three approaches: text-driven prompts that provide feature names and task context, data-driven prompts that provide anonymized perturbed samples, and a novel mixed neighbors prompt that combines feature names with real contrastive neighboring samples. The rankings were compared against rankings from SHAP and LIME using Spearman correlation and top-k overlap.
We found that alignment is mainly dependent on the dataset characteristics and the information included in the prompt. The best setup on California Housing, mixed neighbors with Gemma, came close to the agreement between SHAP and LIME themselves (a Spearman correlation of 0.693 compared to 0.772). On Wine Quality, where SHAP and LIME are in slight agreement (0.226 Spearman correlation), most setups surpassed their agreement, with a text-driven prompt reaching 0.645 against SHAP. On Taiwan Bankruptcy Prediction, all setups stayed below the agreement between SHAP and LIME (at best 0.305 compared to 0.414), with data-driven prompts performing best. Across all three datasets, the choice of prompt template affected agreement more than the choice of LLM.
However, text-driven rankings often missed the most important feature of individual predictions, suggesting that their agreement comes largely from the LLM’s own domain knowledge rather than the sample’s feature values. Overall, LLM-generated rankings can offer a quick, rough estimate of which features are predictive, but are not able to completely replace traditional post-hoc explainers. In practice, LLMs are best suited as a first check when model access is restricted, while high-stakes decisions still require explanations computed from the model itself.
To address this gap, this thesis investigates under what conditions LLM-generated feature importance rankings agree with SHAP and LIME. To answer this, we trained XGBoost models on three datasets across different task types and dimensionalities: California Housing, Wine Quality, and Taiwan Bankruptcy Prediction. We then prompted two open-weight LLMs, Gemma 4 (31B) and GPT-OSS (120B), to rank which features are driving 100 individual predictions per dataset. For the prompting, we used five prompt templates derived from three approaches: text-driven prompts that provide feature names and task context, data-driven prompts that provide anonymized perturbed samples, and a novel mixed neighbors prompt that combines feature names with real contrastive neighboring samples. The rankings were compared against rankings from SHAP and LIME using Spearman correlation and top-k overlap.
We found that alignment is mainly dependent on the dataset characteristics and the information included in the prompt. The best setup on California Housing, mixed neighbors with Gemma, came close to the agreement between SHAP and LIME themselves (a Spearman correlation of 0.693 compared to 0.772). On Wine Quality, where SHAP and LIME are in slight agreement (0.226 Spearman correlation), most setups surpassed their agreement, with a text-driven prompt reaching 0.645 against SHAP. On Taiwan Bankruptcy Prediction, all setups stayed below the agreement between SHAP and LIME (at best 0.305 compared to 0.414), with data-driven prompts performing best. Across all three datasets, the choice of prompt template affected agreement more than the choice of LLM.
However, text-driven rankings often missed the most important feature of individual predictions, suggesting that their agreement comes largely from the LLM’s own domain knowledge rather than the sample’s feature values. Overall, LLM-generated rankings can offer a quick, rough estimate of which features are predictive, but are not able to completely replace traditional post-hoc explainers. In practice, LLMs are best suited as a first check when model access is restricted, while high-stakes decisions still require explanations computed from the model itself.
Automatic Speech Recognizers (ASR) transcribe speech to text and have become very popular. These systems need to accommodate all types of speech, and problems appear for different genders, ages and regional accents. In this paper we aim to discover what are the differences in types of recognition errors made by Dutch listeners and ASR systems in transcribing Flemish Dutch speech. This has been done through an experiment where human listeners and ASRs listened to 40 audio samples coming from speakers from different regions. Afterwards the transcriptions were collected, processed and we calculated the word error rate (WER). ASRs outperformed humans, scored lower WER, with differences appearing based on speaker region. Results show that ASR systems display similar recognition error patterns, a lot of substitutions and deletions, and very few insertions. The main difference stands in the way humans and ASRs deal with confusing audio. Humans tend to use similar words that fit the context whereas ASRs are more prone to inserting unrelated words.
...
Automatic Speech Recognizers (ASR) transcribe speech to text and have become very popular. These systems need to accommodate all types of speech, and problems appear for different genders, ages and regional accents. In this paper we aim to discover what are the differences in types of recognition errors made by Dutch listeners and ASR systems in transcribing Flemish Dutch speech. This has been done through an experiment where human listeners and ASRs listened to 40 audio samples coming from speakers from different regions. Afterwards the transcriptions were collected, processed and we calculated the word error rate (WER). ASRs outperformed humans, scored lower WER, with differences appearing based on speaker region. Results show that ASR systems display similar recognition error patterns, a lot of substitutions and deletions, and very few insertions. The main difference stands in the way humans and ASRs deal with confusing audio. Humans tend to use similar words that fit the context whereas ASRs are more prone to inserting unrelated words.
Automatic speech recognition (ASR) systems are used in voice assistants and games. Studies have shown that these ASR systems do not work well for all everyone. E.g., speech from regional accents is recognised worse than the native standard accent and speech from older adults are recognised worse than adults.However, no research has been done in comparing humans and ASR systems on Dutch older adults' speech.This paper will fill this gap by finding out the transcription performance of humans and ASR systems (Google Telephony and a conformer model with XLSR-53 features) on Dutch older adults' speech and comparing the performances with each other.This paper also investigated whether regional accents, gender, and age had a significant effect on the performance of humans and ASR systems.The results showed that humans and ASR systems perform similar to each other and that their performance are positively correlated.The results also showed that humans also do not recognise speech equally from older adults as well. These results suggests that there are speech characteristics for which both humans and ASR systems have trouble with.
...
Automatic speech recognition (ASR) systems are used in voice assistants and games. Studies have shown that these ASR systems do not work well for all everyone. E.g., speech from regional accents is recognised worse than the native standard accent and speech from older adults are recognised worse than adults.However, no research has been done in comparing humans and ASR systems on Dutch older adults' speech.This paper will fill this gap by finding out the transcription performance of humans and ASR systems (Google Telephony and a conformer model with XLSR-53 features) on Dutch older adults' speech and comparing the performances with each other.This paper also investigated whether regional accents, gender, and age had a significant effect on the performance of humans and ASR systems.The results showed that humans and ASR systems perform similar to each other and that their performance are positively correlated.The results also showed that humans also do not recognise speech equally from older adults as well. These results suggests that there are speech characteristics for which both humans and ASR systems have trouble with.
Human vs AI: Recognising Teenage Speech
How good are humans at recognizing teenage speech samples compared to state-of-the-art AI-based automatic speech recognisers?
Automatic Speech Recognition (ASR) systems have achieved remarkable performance in recent years, however their robustness against diverse speech, such as teenage speech, remains an area of investigation and research. This study evaluates the performance of a state-of-the-art ASR system (Google Telephony) compared to human listeners in transcribing native Dutch teenage speech. Additionally, it investigates whether a listener's social exposure to teenage speech influences their recognition accuracy. A listening experiment was conducted using a dataset of 40 speech samples of Human Machine Interaction (HMI) speech from native Dutch speakers aged 14 to 16, curated from the JASMIN corpus. The audio samples were transcribed by the ASR model and a group of young adult participants (aged 20--24). Performance was evaluated using Word Error Rate (WER), with a specific focus on the impact of normalizing common Dutch contractions and clitics. The results demonstrate that the ASR system outperformed the average human listener, achieving a lower WER compared to both groups of participants, the one with exposure to teenage speech and the one without. However, human participants with regular social exposure to teenagers performed significantly better on average than those without, confirming that familiarity with the demographic improves recognition accuracy. Furthermore, the analysis reveals that orthographic inconsistencies regarding contractions significantly inflate WERs, with normalization reducing the WERs by quite an extent. These findings suggest that while current ASR models are highly robust for this demographic, human domain knowledge remains a relevant factor in understanding teenage speech speech patterns based on the lower WERs of the humans with exposure to teenage speech than those without.
...
Automatic Speech Recognition (ASR) systems have achieved remarkable performance in recent years, however their robustness against diverse speech, such as teenage speech, remains an area of investigation and research. This study evaluates the performance of a state-of-the-art ASR system (Google Telephony) compared to human listeners in transcribing native Dutch teenage speech. Additionally, it investigates whether a listener's social exposure to teenage speech influences their recognition accuracy. A listening experiment was conducted using a dataset of 40 speech samples of Human Machine Interaction (HMI) speech from native Dutch speakers aged 14 to 16, curated from the JASMIN corpus. The audio samples were transcribed by the ASR model and a group of young adult participants (aged 20--24). Performance was evaluated using Word Error Rate (WER), with a specific focus on the impact of normalizing common Dutch contractions and clitics. The results demonstrate that the ASR system outperformed the average human listener, achieving a lower WER compared to both groups of participants, the one with exposure to teenage speech and the one without. However, human participants with regular social exposure to teenagers performed significantly better on average than those without, confirming that familiarity with the demographic improves recognition accuracy. Furthermore, the analysis reveals that orthographic inconsistencies regarding contractions significantly inflate WERs, with normalization reducing the WERs by quite an extent. These findings suggest that while current ASR models are highly robust for this demographic, human domain knowledge remains a relevant factor in understanding teenage speech speech patterns based on the lower WERs of the humans with exposure to teenage speech than those without.
Comparing Human Listeners and Dutch ASR on Transcribing Child Speech
The Effect of Familiarity with Child Speech on Transcription Performance
Automatic Speech Recognition (ASR) systems are becoming increasingly common in day-to-day life. Yet, child speech remains challenging for ASR systems. This paper gives the first comparison of Dutch human listeners and Dutch ASR systems on Dutch child speech. It tests whether familiarity with child speech improves human transcription performance and sees if the age of the child speaker influences the transcription performance.
A balanced set of 40 utterances were taken from the JASMIN database (speakers aged 7-11), and were transcribed by 20 humans (10 familiar with child speech (parent/caretaker) and 10 unfamiliar). Transcripts were also gathered from two state-of-the-art ASR systems (Google Telephony and a Conformer model). These transcripts were evaluated against reference transcripts using Word Error Rate (WER). Statistical significance was tested.
Results show that overall ASR transcription performance was comparable to human performance, and in some cases slightly, but not significantly, better. Familiar listeners did not outperform unfamiliar listeners. In fact, there was no significant performance difference between the two groups of humans. Within the 7-11 age range, no clear relationship between speaker age and WER was shown, but results were sensitive to sentence difficulty outliers and "speaker effects". ...
A balanced set of 40 utterances were taken from the JASMIN database (speakers aged 7-11), and were transcribed by 20 humans (10 familiar with child speech (parent/caretaker) and 10 unfamiliar). Transcripts were also gathered from two state-of-the-art ASR systems (Google Telephony and a Conformer model). These transcripts were evaluated against reference transcripts using Word Error Rate (WER). Statistical significance was tested.
Results show that overall ASR transcription performance was comparable to human performance, and in some cases slightly, but not significantly, better. Familiar listeners did not outperform unfamiliar listeners. In fact, there was no significant performance difference between the two groups of humans. Within the 7-11 age range, no clear relationship between speaker age and WER was shown, but results were sensitive to sentence difficulty outliers and "speaker effects". ...
Automatic Speech Recognition (ASR) systems are becoming increasingly common in day-to-day life. Yet, child speech remains challenging for ASR systems. This paper gives the first comparison of Dutch human listeners and Dutch ASR systems on Dutch child speech. It tests whether familiarity with child speech improves human transcription performance and sees if the age of the child speaker influences the transcription performance.
A balanced set of 40 utterances were taken from the JASMIN database (speakers aged 7-11), and were transcribed by 20 humans (10 familiar with child speech (parent/caretaker) and 10 unfamiliar). Transcripts were also gathered from two state-of-the-art ASR systems (Google Telephony and a Conformer model). These transcripts were evaluated against reference transcripts using Word Error Rate (WER). Statistical significance was tested.
Results show that overall ASR transcription performance was comparable to human performance, and in some cases slightly, but not significantly, better. Familiar listeners did not outperform unfamiliar listeners. In fact, there was no significant performance difference between the two groups of humans. Within the 7-11 age range, no clear relationship between speaker age and WER was shown, but results were sensitive to sentence difficulty outliers and "speaker effects".
A balanced set of 40 utterances were taken from the JASMIN database (speakers aged 7-11), and were transcribed by 20 humans (10 familiar with child speech (parent/caretaker) and 10 unfamiliar). Transcripts were also gathered from two state-of-the-art ASR systems (Google Telephony and a Conformer model). These transcripts were evaluated against reference transcripts using Word Error Rate (WER). Statistical significance was tested.
Results show that overall ASR transcription performance was comparable to human performance, and in some cases slightly, but not significantly, better. Familiar listeners did not outperform unfamiliar listeners. In fact, there was no significant performance difference between the two groups of humans. Within the 7-11 age range, no clear relationship between speaker age and WER was shown, but results were sensitive to sentence difficulty outliers and "speaker effects".
Master thesis
(2025)
-
C.W. Ning, M.L. Tielman, Carolina Centeio Jorge, M.A. Neerincx, B.J.W. Dudzik
As robots and virtual agents are increasingly envisioned as long-term companions rather than simply tools, it becomes essential to ensure that human–robot relationships are grounded in appropriate forms of trust. This study investigates how cognitive and affective dimensions of trust develop differently over time in social human–robot interaction. We conducted a 2 (social attitude: social, baseline) × 3 (time: t1, t2, t3) mixed-design user study using a novel, card-based conversational task designed to encourage trust formation. Results show that while cognitive trust remained stable over time, affective trust increased gradually across repeated interactions. Moreover, social cues enhance both cognitive and affective trust. These findings provide empirical support for the theoretical distinction between cognitive and affective trust, offering new evidence that affective trust develops more slowly, consistent with interpersonal trust theories.
...
As robots and virtual agents are increasingly envisioned as long-term companions rather than simply tools, it becomes essential to ensure that human–robot relationships are grounded in appropriate forms of trust. This study investigates how cognitive and affective dimensions of trust develop differently over time in social human–robot interaction. We conducted a 2 (social attitude: social, baseline) × 3 (time: t1, t2, t3) mixed-design user study using a novel, card-based conversational task designed to encourage trust formation. Results show that while cognitive trust remained stable over time, affective trust increased gradually across repeated interactions. Moreover, social cues enhance both cognitive and affective trust. These findings provide empirical support for the theoretical distinction between cognitive and affective trust, offering new evidence that affective trust develops more slowly, consistent with interpersonal trust theories.
Counterfactual explanations can be applied to algorithmic recourse, which is concerned with helping individuals in the real world overturn undesirable algorithmic decisions. They aim to provide explanations to opaque machine learning models. Not all generated points are equally faithful to the model, nor equally plausible. On the other hand, predictive uncertainty quantification is used to measure the degree of certainty a model has in its predictions. Previously, it has been shown that it is possible to generate more plausible counterfactual explanations utilising predictive uncertainty. This work investigates this further by using multiple models innately supporting uncertainty quantification and comparing the produced counterfactual explanations to those produced by the models' ordinary counter-part. Predictive uncertainty tends to enhance the plausibility of the counterfactuals on visual datasets. Furthermore, we are positive that predictive uncertainty correlates proportionally with plausibility. This correlation has important implications for both research and real-world applications, as it suggests that integrating uncertainty quantification in model development can improve the quality and trustworthiness of algorithmic explanations.
...
Counterfactual explanations can be applied to algorithmic recourse, which is concerned with helping individuals in the real world overturn undesirable algorithmic decisions. They aim to provide explanations to opaque machine learning models. Not all generated points are equally faithful to the model, nor equally plausible. On the other hand, predictive uncertainty quantification is used to measure the degree of certainty a model has in its predictions. Previously, it has been shown that it is possible to generate more plausible counterfactual explanations utilising predictive uncertainty. This work investigates this further by using multiple models innately supporting uncertainty quantification and comparing the produced counterfactual explanations to those produced by the models' ordinary counter-part. Predictive uncertainty tends to enhance the plausibility of the counterfactuals on visual datasets. Furthermore, we are positive that predictive uncertainty correlates proportionally with plausibility. This correlation has important implications for both research and real-world applications, as it suggests that integrating uncertainty quantification in model development can improve the quality and trustworthiness of algorithmic explanations.
In recent years, the need for explainable artificial intelligence (XAI) has become increasingly important as complex black-box models are used in critical applications. While many methods have been developed to interpret these models, there is also potential in enhancing the models themselves to improve their inherent explainability. This paper investigates various techniques aimed at improving the explainability of black-box models. Through a systematic literature review, these techniques are categorized, and their impact on predictive uncertainty, adversarial robustness, and generative capacity is analyzed to understand how these factors contribute to the overall explainability. The snowballing methodology is used for the systematic literature review, starting with papers retrieved from four databases: IEEExplore, Scopus, ArXiv, and the ACM Digital Library to form the initial set. This process continued with backward and forward snowballing through four iterations, resulting in a total of 50 papers reviewed. Only papers focused on improving model explainability are included in the review. Due to time limitations, additional search constraints are applied for feasibility. The initial set of papers is filtered to those published since 2013. These constraints and their possible impacts are considered when interpreting the results. Findings reveal that techniques such as Bayesian approaches and variational inference, adversarial robustness, model compression and distillation, uncertainty and ensembles, regularization, self-explaining models, and hybrid techniques are used for advancing model explainability. The paper concludes with a discussion on the implications of these techniques for future research.
...
In recent years, the need for explainable artificial intelligence (XAI) has become increasingly important as complex black-box models are used in critical applications. While many methods have been developed to interpret these models, there is also potential in enhancing the models themselves to improve their inherent explainability. This paper investigates various techniques aimed at improving the explainability of black-box models. Through a systematic literature review, these techniques are categorized, and their impact on predictive uncertainty, adversarial robustness, and generative capacity is analyzed to understand how these factors contribute to the overall explainability. The snowballing methodology is used for the systematic literature review, starting with papers retrieved from four databases: IEEExplore, Scopus, ArXiv, and the ACM Digital Library to form the initial set. This process continued with backward and forward snowballing through four iterations, resulting in a total of 50 papers reviewed. Only papers focused on improving model explainability are included in the review. Due to time limitations, additional search constraints are applied for feasibility. The initial set of papers is filtered to those published since 2013. These constraints and their possible impacts are considered when interpreting the results. Findings reveal that techniques such as Bayesian approaches and variational inference, adversarial robustness, model compression and distillation, uncertainty and ensembles, regularization, self-explaining models, and hybrid techniques are used for advancing model explainability. The paper concludes with a discussion on the implications of these techniques for future research.
Counterfactual Explanations (CE) are essential for understanding the predictions of black-box models by suggesting minimal changes to input features that would alter the output. Despite their importance in Explainable AI (XAI), there is a lack of standardized metrics to assess the plausibility and faithfulness of these explanations. This paper reviews evaluation procedures in literature and proposes novel formal metrics for evaluating the plausibility and faithfulness of counterfactual explanations, addressing the existing limitations. Plausibility is defined as the coherence of explanations with the true data-generating process, while faithfulness refers to the accuracy of explanations in representing the model's reasoning. We discuss the shortcomings of existing evaluation procedures and metrics for measuring plausibility and faithfulness and consequently compare our proposed metrics with existing ones, highlighting their advantages and disadvantages. The proposed metrics are then empirically validated through experiments across multiple models and datasets, demonstrating their model-agnostic nature and reliability. Our findings indicate that the proposed metrics provide a correct and reliable means to quantify the plausibility and faithfulness of counterfactual explanations, thereby allowing one to gauge their feasibility and trustworthiness consistently.
...
Counterfactual Explanations (CE) are essential for understanding the predictions of black-box models by suggesting minimal changes to input features that would alter the output. Despite their importance in Explainable AI (XAI), there is a lack of standardized metrics to assess the plausibility and faithfulness of these explanations. This paper reviews evaluation procedures in literature and proposes novel formal metrics for evaluating the plausibility and faithfulness of counterfactual explanations, addressing the existing limitations. Plausibility is defined as the coherence of explanations with the true data-generating process, while faithfulness refers to the accuracy of explanations in representing the model's reasoning. We discuss the shortcomings of existing evaluation procedures and metrics for measuring plausibility and faithfulness and consequently compare our proposed metrics with existing ones, highlighting their advantages and disadvantages. The proposed metrics are then empirically validated through experiments across multiple models and datasets, demonstrating their model-agnostic nature and reliability. Our findings indicate that the proposed metrics provide a correct and reliable means to quantify the plausibility and faithfulness of counterfactual explanations, thereby allowing one to gauge their feasibility and trustworthiness consistently.
Adversarial Training has emerged as the most reliable technique to make neural networks robust to gradient-based adversarial perturbations on input data. Besides improving model robustness, preliminary evidence presents an interesting consequence of adversarial training -- increased explainability of model behaviour. Prior work has explored the effects of adversarial training on gradient stability and interpretability, as well as visual explainability of counterfactuals. Our work presents the first quantitative, empirical analysis of the impact of model robustness on model explainability by comparing the plausibility of faithful counterfactuals for both robust and standard networks. We seek to determine whether robust networks learn more plausible decision boundaries and representations of the data than regular models, and whether the strength of the adversary used to train robust models affects their explainability. Our finidngs indicate that robust networks for image data learn more explainable decision boundaries and representations of data than regular models, with more robust models producing more plausible counterfactuals. Robust models for tabular data, however, only conclusively exhibit this phenomenon along decision boundaries and not for its overall data representations, possibly due to its high robustness-accuracy trade-off and the difficulties associated with traditional adversarial training due to its innate properties. We believe our work can help guide future research towards improving the robustness of machine learning models keeping their explainability in mind.
...
...
Adversarial Training has emerged as the most reliable technique to make neural networks robust to gradient-based adversarial perturbations on input data. Besides improving model robustness, preliminary evidence presents an interesting consequence of adversarial training -- increased explainability of model behaviour. Prior work has explored the effects of adversarial training on gradient stability and interpretability, as well as visual explainability of counterfactuals. Our work presents the first quantitative, empirical analysis of the impact of model robustness on model explainability by comparing the plausibility of faithful counterfactuals for both robust and standard networks. We seek to determine whether robust networks learn more plausible decision boundaries and representations of the data than regular models, and whether the strength of the adversary used to train robust models affects their explainability. Our finidngs indicate that robust networks for image data learn more explainable decision boundaries and representations of data than regular models, with more robust models producing more plausible counterfactuals. Robust models for tabular data, however, only conclusively exhibit this phenomenon along decision boundaries and not for its overall data representations, possibly due to its high robustness-accuracy trade-off and the difficulties associated with traditional adversarial training due to its innate properties. We believe our work can help guide future research towards improving the robustness of machine learning models keeping their explainability in mind.
Counterfactual explanations (CEs) can be used to gain useful insights into the behaviour of opaque classification models, allowing users to make an informed decision when trusting such systems. Assuming the CEs of a model are faithful (they well represent the inner workings of the model), an explainable model generates plausible CEs (i.e. CEs fitting the real-world distribution of the data). This raises the question of whether classifiers explicitly designed to model the distribution of the data, such as energy-based models, are inherently more explainable. This work focuses on the evaluation of joint energy-based models (JEMs) in combination with the Energy-Constrained Conformal Counterfactuals (ECCCo) generator, with the goal of identifying if the generative capability of a model influences its explainability. Since ECCCo has been designed specifically to generate more faithful CEs, it makes it possible to use the CEs plausibility as a proxy of the model explainability. Two experiments have been performed to evaluate the effect of variations of generative capability within the same JEM architecture and the difference between JEMs and classically trained classifiers. Despite the experiments not having established a clear correlation between generative capability and explainability of a model, various research avenues are still open to explore in future works
...
Counterfactual explanations (CEs) can be used to gain useful insights into the behaviour of opaque classification models, allowing users to make an informed decision when trusting such systems. Assuming the CEs of a model are faithful (they well represent the inner workings of the model), an explainable model generates plausible CEs (i.e. CEs fitting the real-world distribution of the data). This raises the question of whether classifiers explicitly designed to model the distribution of the data, such as energy-based models, are inherently more explainable. This work focuses on the evaluation of joint energy-based models (JEMs) in combination with the Energy-Constrained Conformal Counterfactuals (ECCCo) generator, with the goal of identifying if the generative capability of a model influences its explainability. Since ECCCo has been designed specifically to generate more faithful CEs, it makes it possible to use the CEs plausibility as a proxy of the model explainability. Two experiments have been performed to evaluate the effect of variations of generative capability within the same JEM architecture and the difference between JEMs and classically trained classifiers. Despite the experiments not having established a clear correlation between generative capability and explainability of a model, various research avenues are still open to explore in future works