My main research focus is on bridging the gap between symbolic AI and neural networks. Symbolic AI produces knowledge that is directly readable by humans, but is brittle and hard to scale; neural networks are robust and broadly applicable, but the knowledge they encode is hard to access. By combining both approaches, I aim to build AI systems whose knowledge humans can read, check, and build on. My current work ranges from foundation models that induce logical rules from data, to neural networks that learn when to backtrack during reasoning.

Publications Grants, talks, and background

Foundation Models for Logical Rule Induction

Symbolic rule learners such as decision trees, RIPPER, and inductive logic programming (ILP) are interpretable, but they have to be retrained from scratch on every new dataset, and transfer nothing from one problem to the next. Rather than training a model for each task, we want a model that learns how to discover rules. To this end, we proposed the Neural Rule Inducer (NRI), a foundation model for logical rule induction (Phua, 35th International Joint Conference on Artificial Intelligence, 2026). NRI is trained once on randomly generated logic problems, and then induces human-readable rules from the examples of a new dataset without retraining. The resulting rules describe which combinations of features lead to an outcome. Since they are executable logic, they can be inspected by humans and reused in other systems.

Scaling to Larger Problems through Symmetry

Renaming the variables of a logic problem does not change the problem itself. By exploiting symmetries inherent in logic such as this one, we proposed G-NRI, which lets NRI transfer what it learned on small problems to much larger ones (Phua, 20th International Conference on Neurosymbolic Learning and Reasoning, 2026, accepted). A model trained on problems with 6–12 variables was applied without retraining to problems with up to 1,024 variables, improving on the original NRI. We are now investigating how far generalization can be pushed by building the structure of logic into the learning itself.

Learning When to Backtrack

When reasoning involves search, a model can follow a long reasoning trace and still fail to notice that it has reached a dead end and needs to go back. In collaboration with Tony Ribeiro, Tuan Nguyen, and Katsumi Inoue, we study how neural networks can learn this decision (Can Transformers Learn to Verify During Backtracking Search?, 2026 preprint). We found that the information needed for the decision can get buried in the trace, and that earlier search steps can distract the model from the state it is actually in.

In our approach, we make the current state explicit, and use Selective State Attention (SSA) to restrict attention to that state and the problem at hand. Across logic, planning, and parsing tasks, this helps models learn to backtrack in response to contradictions, instead of simply continuing the trace. We aim to extend this toward systems that can notice and recover from their own mistakes.

Choosing Experiments to Recover Gene Regulatory Rules

In collaboration with Foo Wei Ten, we are working on AI that not only learns from experiments, but also chooses which experiment to perform next. ABLE (Active Boolean Learning Engine) proposes candidate rules for gene regulation, checks them against observations, and selects the experiment that would distinguish the remaining candidates (Phua and Ten, ICML 2026 AI for Science Workshop, 2026). We evaluate this loop on simulations of biological Boolean models. Rather than returning a single rule and stopping, the system uses what is still unknown to decide what to investigate next.

Learning Logic Programs from Dynamic Systems

My main research focus has been on a neuro-symbolic method for learning transition rules from time series data. By feeding the time series into an LSTM (Long-Short Term Memory) and using the corresponding logic program as the hidden states, the LSTM learned representations of the logic program and used them to predict the next state — showing that a single neural network can classify time series originating from different dynamic systems and predict the next state from historical data alone (Phua, et al. 27th International Conference on Inductive Logic Programming, 2017). We then showed that a neural network can also identify time series that come from the same dynamic system, and revised our method to take background knowledge into account (Phua, et al. IfCoLog Journal of Applied Logics, 2019). Training the LSTM to identify time series from the same system further improved the accuracy of our method.

Building on these results, and leveraging the strength of neural networks for pattern recognition, we proposed a method that learns logical rules directly from state transition data (Phua, et al. 29th International Conference on Inductive Logic Programming, 2019). By classifying the rules inherent in the transitions, the network outputs a logic program that represents the dynamic system. We added a constraint on the learned rules so the network can classify all possible rules without modifying its architecture, and proved that the resulting method is robust against noise. This work was recognized by the community, and we received the Best Student Paper award. Next, by exploiting invariant properties inherent in logic systems — in particular, invariance to the order of the state transition input — we suppressed the combinatorial explosion that comes from trying to classify all possible rules (Phua, et al. 1st International Joint Conference on Learning & Reasoning, 2021). We proposed a neural network architecture that bakes this invariance in, and by considering multiple logical rules in a single inference, we also reduced the network's memory requirements. This let us apply our method to larger dynamic systems, bringing it a step closer to real world applications.

We have since extended our method to delayed dynamic systems, demonstrated robustness against missing data, and generalized it from the synchronous semantics of all our previous work to asynchronous semantics.

Class-Incremental Learning using Generative Models

Statistical machine learning methods like neural networks are known to suffer from catastrophic forgetting: training a model on a new task tends to erase what it learned on previous ones. The field that studies this phenomenon and its mitigation is called incremental learning. One branch of it considers classification, where at each training step only a limited number of classes are available. Many methods have been proposed with various degrees of success, and the one we consider combines knowledge distillation — training a model to retain knowledge from another model — with replay memory, a fixed-size buffer of examples from previous classes. Previous methods populate this buffer with instances from the training set, but the limited memory budget caps how many instances can be retained, which limits performance. In this work, we proposed replacing the replay memory with widely available generative models, and devised a method for prompting the generative model to obtain relevant images. This improved existing methods by 1~3%, and we also found there is a sweet spot between the final accuracy and the number of images sampled from the generative model.

Results of this work have been published in the workshop paper (Jodelet, et al. 1st Workshop on Visual Continual Learning, 2023), which was chosen for the best paper award. Our subsequent journal papers include Memory augmented using diffusion model for class-incremental learning (Image and Vision Computing, 2025) and Future-proofing class-incremental learning (Machine Vision and Applications, 2025), with Quentin Jodelet, Xin Liu, and Tsuyoshi Murata.

Consistent Identification of Gene Sets from Cell Sequencing Data

In collaboration with the Berlin Institute of Health, we proposed a method that identifies important gene sets for different cell types from single cell sequencing data. Biologists work under time pressure when analyzing such data, and quickly identifying gene sets can accelerate the development of important drugs. Previous methods focused only on whether gene sets could be extracted at all, but the randomness inherent to statistical machine learning meant that each run could produce a different set of genes, complicating the analysis. By utilizing an ensemble of resVAEs, we were able to stabilize the identified gene sets. We further show that the choice of activation function has large implications for stability, and thus chose one that provides even greater stability.

Results of this work were published in the journal paper (Ten, et al. Frontiers in Cell and Developmental Biology 11, 2023).

Testing Consciousness Theories on Artificial Agents

I am also interested in what theories of consciousness can tell us about intelligence. In this work, we built artificial agents with mechanisms inspired by these theories, and removed individual components to observe what changes (Can We Test Consciousness Theories on AI? Ablations, Markers, and Robustness, 2025 preprint). The experiments measure how the agents share information, judge their own confidence, and cope with internal noise. This allows ideas about cognition to be examined on systems that we can directly experiment with.

Have questions, or want to know more about any of this work? Feel free to email me at phua [at] comp.isct.ac.jp.

« Back to home