Page Not Found
Page not found. Your pixels are in another canvas.
Sitemap
A list of all the posts and pages found on the site. For you robots out there is an XML version available for digesting as well.
Pages
Bio
Background
Posts
Vibe Coding and Levels of Abstraction
One of the most underappreciated ideas in all of science is the idea of abstraction. Abstraction serves as a practical tool for understanding, building, and communicating (see also Why Do We Abstract?). The history of computer science is arguably the history of abstraction. And right now, with the rise of AI coding tools, that history is more relevant than ever.
The Genie and the Therapist: Rethinking Bioethics for Digital Mental Health
I have become more interested in bioethics and AI governance recently, because it has become clear that these potentially present the biggest challenge to digital mental health transformation. I have come to believe that traditional bioethical frameworks like compassionate use are inadequate for digital mental health. This essay examines why existing frameworks fail and identifies key questions that new approaches must address.
Poor Economics and the Loneliness Epidemic
Other people often seem to make irrational decisions. But, as it turns out, these decisions are often rational in their frame of reference. This essay discusses this rationality in the context of economics and AI friends for loneliness.
Thinking in Extremes
This essay discusses the cognitive tool of thinking in spectra and extremes.
The Limits of Thinking
This essay discusses the mental concept of thinking in limits. It’s an incredibly powerful mental tool for analyzing a wide range of situations. Its utility is demonstrated in a situation where humans have poor intuition.
Cognitive Unit Tests
This essay introduces the concept of something I like to call the “cognitive unit tests”.
Subjective Knowledge Graphs
Some years ago I read this blog post by Tim Urban about Neuralink. You should probably go and read it first, he is a great story teller. The main image for this article’s context is the following.
Karen Horney’s Radical Idea: You Can Analyze Yourself
Karen Horney believed that ordinary people, under the right conditions, could do a meaningful version of what analysts do in a consulting room. This was a controversial claim in 1942, when she published Self-Analysis, and it remains a provocative one today. This essay examines what she actually meant, how credible her case is, and why the idea feels newly relevant in an era of digital self-reflection tools.
The User-Centric Fallacy
There is a piece of advice that gets recycled endlessly in startup culture: build something simple that anyone can use. Ship the minimum viable product. Talk to your users. Remove friction. Design for the person who has never heard of you and never wants to think about your product for more than thirty seconds. Go for the large user base.
A Taxonomy of Robust Science
I took this great course on behavioral science as a robust discipline by Amy Orben at the University of Cambridge during my PhD. It has since greatly influenced how I see the fields of psychology and neuroscience, but also science in general. In fact, Samuel Bell and I wrote a short paper on what the field of machine learning research could learn from the reproducibility crisis in psychology.
The Mental Health and Performance Continuum
Mental health and mental performance exist on a continuum, not as binary states. Like most natural processes in the world, if we plotted the population’s mental functioning (as a rough concept), we’d see a normal distribution from those struggling significantly on one end, through the average majority in the middle, to high performers on the other. The people at the left side of the spectrum suffer tremendously and are typically incapable of functioning in society, whereas the people on the right side are thriving. Commonly used mental health questionnaires such as the PHQ-9 assign individuals to a category of “severity”, such as “mild” or “severe”.
Why Do We Abstract?
The concept of abstraction is one of the building blocks to advanced understanding of our world.
Essays in Progress
I’ve decided to start publishing occasional essays—short pieces on things I’m thinking about, exploring, or trying to understand better. These won’t be polished takes or final conclusions (those are more reserved for peer-reviewed publications). Instead, I’ll treat them as living documents: open to revision, correction, and rethinking.
publications
Nora the empathetic psychologist
INTERSPEECH, 2017
Nora is a new dialog system that mimics a conversation with a psychologist by screening for stress, anxiety, and depression. She understands, empathizes, and adapts to users using emotional intelligence modules trained via statistical modelling such as Convolutional Neural Networks. These modules also enable her to personalize the content of each conversation.
Winata et al. (2018). "Nora the Empathetic Psychologist." INTERSPEECH. [PDF]
Attention-based LSTM for psychological stress detection from spoken language using distant supervision
ICASSP, 2018
The bidirectional LSTM model with attention is found to be the best model in terms of accuracy and f-score and it is shown that distant supervision fine-tuning enhances the model’s performance by 1.6% accuracy and 2.1% f-score.
Winata et al. (2018). "Attention-Based LSTM for Psychological Stress Detection from Spoken Language Using Distant Supervision." ICASSP. [PDF]
Investigating audio, video, and text fusion methods for end-to-end automatic personality prediction
ACL, 2018
It is shown that a multimodal fusion approach outperforms each single modality channel, with an improvement of 9.4% over the best individual modality (video).
Kampman et al. (2018). "Investigating Audio, Video, and Text Fusion Methods for End-to-End Automatic Personality Prediction." ACL. [PDF]
Adapting a Virtual Agent to User Personality
International Workshop on Spoken Dialogue Systems Technology, 2018
The study suggests that especially the Openness user personality trait correlates with a stronger preference for agents with more gentle personality, and people also sense more empathy and enjoy better conversations when agents adapt their personalities.
Kampman et al. (2018). "Adapting a Virtual Agent to User Personality." International Workshop on Spoken Dialogue Systems Technology. [PDF]
Towards universal end-to-end affect recognition from multilingual speech by ConvNets
arXiv, 2019
We propose an end-to-end affect recognition approach using a Convolutional Neural Network (CNN) that handles multiple languages, with applications to emotion and personality recognition from speech.
Bertero et al. (2019). "Towards Universal End-to-End Affect Recognition from Multilingual Speech by ConvNets." arXiv. [PDF]
Variational inference for neural network matrix factorization and its application to stochastic blockmodeling
ICML Workshop on Learning and Reasoning with Graph-Structured Data, 2019
Neural network matrix factorization with variational inference.
Kampman and Heaukulani. (2019). "Variational inference for neural network matrix factorization and its application to stochastic blockmodeling." ICML Workshop on Learning and Reasoning with Graph-Structured Data. [PDF]
Perspectives on machine learning from psychology’s reproducibility crisis
ICLR Workshop of Science and Engineering of Deep Learning, 2021
This short paper presents select ideas from psychology’s reformation, translating them into relevance for a machine learning audience.
Bell and Kampman (2021). "Perspectives on Machine Learning from Psychology's Reproducibility Crisis." ICLR Workshop of Science and Engineering of Deep Learning. [PDF]
Modeling the machine learning multiverse
NeurIPS, 2022
This work model the multiverse with a Gaussian Process surrogate and apply Bayesian experimental design to efficiently explore high-dimensional and often continuous ML search spaces, and synthesize conflicting research on the effect of learning rate on the large batch training generalization gap.
Bell et al. (2022). "Modeling the Machine Learning Multiverse." NeurIPS. [PDF]
Properties of V1 and MT motion tuning emerge from unsupervised predictive learning
Journal of Vision, 2022
Unsupervised predictive learning creates neural-like single-unit tuning, population tuning statistics, and integration of locally-ambiguous motion signals, and provides an interrogable model of why motion computations take the form they do.
Storrs et al. (2022). "Properties of V1 and MT motion tuning emerge from unsupervised predictive learning." Journal of Vision. [PDF]
Time-varying functional connectivity as Wishart processes
Imaging Neuroscience, 2024
The WP outperformed a sliding window approach with adaptive cross-validated window lengths and a dynamic conditional correlation-multivariate generalized autoregressive conditional heteroskedasticity (MGARCH) baseline on the external stimulus prediction task, while being less prone to false positives in the TVFC null models.
Kampman et al. (2024). "Time-varying functional connectivity as Wishart processes." Imaging Neuroscience. [PDF]
SEACrowd: A multilingual multimodal data hub and benchmark suite for Southeast Asian languages
EMNLP, 2024
This work introduces SEACrowd, a comprehensive resource center that fills the resource gap by providing standardized corpora in nearly 1,000 SEA languages across three modalities, and assesses the quality of AI models on 36 indigenous languages across 13 tasks.
Lovenia et al. (2024). "SEACrowd: A Multilingual Multimodal Data Hub and Benchmark Suite for Southeast Asian Languages." EMNLP. [PDF]
A multi-agent dual dialogue system to support mental health care providers
arXiv, 2024
A general-purpose, human-in-the-loop dual dialogue system to support mental health care professionals and found that the proposed responses matched a reasonable human quality in demonstrating empathy, showing its appropriateness for augmenting the work of mental health care providers.
Kampman et al. (2024). "A Multi-Agent Dual Dialogue System to Support Mental Health Care Providers." arXiv. [PDF]
Crowdsource, crawl, or generate? Creating SEA-VL, a multicultural vision-language dataset for Southeast Asia
ACL, 2025
SEA-VL, an open-source initiative dedicated to developing high-quality, culturally relevant data for SEA languages, aims to bridge the representation gap in SEA, fostering the development of more inclusive AI systems that authentically represent diverse cultures across SEA.
Cahyawijaya et al. (2025). "Crowdsource, Crawl, or Generate? Creating SEA-VL, a Multicultural Vision-Language Dataset for Southeast Asia." ACL. [PDF]
Conversational self-play for discovering and understanding psychotherapy approaches
AI4X Conference, 2025
This paper explores conversational self-play with LLMs as a scalable approach for analyzing and exploring psychotherapy approaches, evaluating how well AI-generated therapeutic dialogues align with established modalities.
Kampman, Onno P. (2025). "Conversational Self-Play for Discovering and Understanding Psychotherapy Approaches." AI4X Conference. [PDF]
Harnessing digital phenotyping to advance university student mental health (Brightline) in Singapore: Study protocol for a prospective observational study
BMJ Open, 2025
This study will employ an observational study design over a 6-month period, recruiting 500 students from a major public university in Singapore, to identify the digital biomarkers associated with depression, anxiety, stress, loneliness and affect among university students.
Ito et al. (2025). "Harnessing digital phenotyping to advance university student mental health (Brightline) in Singapore: study protocol for a prospective observational study." BMJ Open. [PDF]
Mind the Gap: Aligning knowledge bases with user needs to enhance mental health retrieval
NeurIPS GenAI4Health Workshop, 2025
This paper introduces a gap-informed framework for expanding mental health knowledge bases, using naturalistic data like forum posts to identify underrepresented topics and prioritize which content to add. In a case study across four RAG pipelines, this directed approach reached ~95% of an exhaustive reference corpus’s performance with expansions as small as 42%, compared to the 232–763% required by random augmentation.
Chan et al. (2025). "Mind the Gap: Aligning Knowledge Bases with User Needs to Enhance Mental Health Retrieval." NeurIPS GenAI4Health Workshop. [PDF]
Mind the Language Gap: Building an Inclusive AI Future for Southeast Asia
Tech For Good Institute, 2025
Perspective piece arguing that closing the linguistic gap in Southeast Asian AI systems — through initiatives like SEACrowd — is essential for equitable digital inclusion and for preventing harmful misinterpretations in critical sectors like healthcare.
Kampman & Lovenia (2025). "Mind the Language Gap: Building an Inclusive AI Future for Southeast Asia." Tech For Good Institute. [Read]
Worker skills associated with outcomes in suicidal-related youth chat sessions
OSF, 2026
Introduction: Text-based chatlines have become preferred entry points for youth seeking mental health support, yet most research examines dedicated crisis services rather than general chatlines where suicide emerges alongside diverse concerns. This study compared suicidal-related and non-suicidal sessions within a general youth chatline to identify session characteristics and worker skills associated with positive outcomes.Methods: We analyzed 1,710 chat sessions (202,336 messages) from QuickChat, a Singapore youth chatline between 2016 and 2020. Large language models classified sessions as suicidal or self-harm related(n=406, 24%) or non-suicidal (n=1,304, 76%).User-reported outcomes measured service quality and coping ability. Twelve therapeutic skills were coded from 79,587 worker messages. Multilevel regression models examined skill-outcome associations.Results: Suicidal-related sessions were significantly longer, contained more messages, and yielded lower outcomes. Suicidal ideation was most prevalent (85%), followed by self-harm (43%). In suicidal sessions, normalization demonstrated the strongest associations with all outcomes, followed by teaching/psychoeducation and making strengths explicit. These patterns differed substantially from non-suicidal sessions.Conclusion: Suicidal-related sessions within general chatlines demand greater engagement and differentiated responses from workers. Normalization and psychoeducation emerge as effective techniques for improving outcomes in suicidal chats. These findings provide actionable guidance for training frontline workers in general youth services.
Chung, Lim, and Kampman. (2026). "Worker Skills Associated With Outcomes In Suicidal-Related Youth Chat Sessions." OSF. [PDF]
The let’s talk Digital Peer Support Forum for Youth Mental Health and Wellbeing in Singapore: A Three-Year Process Evaluation and Framework Description
JMIR Mental Health Preprint, 2026
let’s talk is a Singaporean mental health forum. This article describes it Theory of Change and discusses a process evaluation of its first three years.
Weng et al. (2026). "The let’s talk Digital Peer Support Forum for Youth Mental Health and Wellbeing in Singapore: A Three-Year Process Evaluation and Framework Description." JMIR Mental Health Preprint. [PDF]
Language-centered mental healthcare with large language models: A comprehensive survey across the care continuum
Preprint, 2026
Survey paper on the use of LLMs in mental healthcare.
Wu et al. (2026). "Language-centered mental healthcare with large language models: A comprehensive survey across the care continuum." Preprint. [PDF]
teaching
DSA4262 Sense-making Case Analysis: Health and Medicine (2026)
Undergraduate Year 4 Course, National University of Singapore, Department of Statistics and Data Science, 2026
I am teaching DSA4262 Sense-making Case Analysis: Health & Medicine in 2026 (AY2025/26 Semester 2).