The people behind Cerenaut

David’s Monash Masters thesis, “Improving Robust Maintenance in Transformer RL Agents”, takes the prefrontal cortex and basal ganglia working-memory framework into Transformer-based RL. A Transformer-XL agent normally evicts its oldest memories regardless of relevance, so a cue is lost once it leaves the window. David replaced this with use-based eviction (least recently used, and an exponential moving average of attention) and tested sparse attention for the readout. Across 72 configurations (360 runs) on Endless T-Maze and MiniGrid Memory, the standard age-based eviction failed whenever memory was short relative to the corridor, while the use-based rules avoided this failure: with a quarter of the memory they came close to its best scores. The cost was slower, less steady learning, and sparse attention gave no consistent benefit.

Jingwen’s Masters thesis, “Emergent Bilateral Asymmetry in Motor Control”, extended the left/right motor-control model that Jarrad Rinaldo built. It asked whether two-hemisphere controllers beat single ones for two-armed robotic control. In simulated musculoskeletal arms, modular controllers beat capacity-matched monolithic ones (87.6% vs 68.9% success) with lower energy use. Giving each controller a different training objective with fixed hand roles produced the strongest specialisation, at the cost of flexibility when roles change. Two papers from the work are on arXiv.

Guixuan’s Monash Masters thesis, “Explicit Belief Dynamics for Recurrent Grid Reasoning”, studies models that refine an answer over many steps, such as HRM, TRM and URM, and asks what the thing being refined should be. His Belief Dynamics Reasoner keeps the answer as an explicit per-cell probability draft over the nine digits, so it can be inspected while the solver is still running, and he tried updating it directly, with a learned velocity field, and with a learned energy. On Sudoku the direct version reached about 59% accuracy on blank cells, well below the TRM and URM baselines, and none of the versions solved boards reliably. The draft worked mainly as a diagnostic tool, showing whether the belief was moving, whether conflicts were falling and whether the model preferred wrong states to the target. The thesis reports the energy and velocity experiments as incomplete.

Ben’s Monash Masters thesis, “In Two Minds about Lifelong Learning”, asked whether three features of the brain can reduce catastrophic forgetting: replay of past experience, REM-like sleep, and two slightly different hemispheres. His model has two asymmetric hemispheres, each with short- and long-term memory, and a sleep-like consolidation phase between tasks. A paper with Levin Kuhlmann, Kaushik Roy and Gideon Kowadlo, posted to arXiv in 2026, reports 98.3% on Split-MNIST and 84.9% on Split-Fashion-MNIST, cutting forgetting by more than half against the strongest replay baseline.

Joel’s Monash Masters thesis extended Grant Nicholas’s bi-hemispheric RL agent. He asked whether the acting hemisphere could be chosen by its confidence, measured as the spread of each hemisphere’s action distribution, instead of blending the two outputs. On the Meta-World robot benchmark the selection agent underperformed the blending baseline, because the left hemisphere was chosen less often and so missed learning updates. It is a useful negative result, and the thesis points to forced or off-policy updates for the left hemisphere as the next step.

Haowei’s Monash Master of IT thesis, “Corpus Callosum-Inspired Bilateral Neural Networks”, began a new phase of the bilateral vision work by connecting the two hemispheres of the network the way the corpus callosum does. Two ResNet-9 networks stood in for the hemispheres, with fixed and learnable connections added at different depths. On CIFAR-100 the best method, gated feature transmission after convolutional blocks, reached 68.0% fine-grained accuracy, about a point above a single network. Learnable connections beat fixed ones and connecting earlier worked better, though the gains were small relative to run-to-run variation.

Grant’s Monash Masters thesis built a bi-hemispheric RL agent based on the Novelty-Routine Hypothesis: the right hemisphere handles novelty, and responsibility shifts to the left as a task becomes routine. A frozen, meta-trained generalist right network acts first while a left network learns the task from scratch, with control handed over by a gating schedule or a learned gate. On nine Meta-World robot tasks the agent often avoided poor early performance, and a gate with a penalty term trained a left network that matched or beat an agent trained from scratch.

Jarrad’s Monash Masters thesis applied the left/right brain idea to human motor control and its implications for robotics. In people the dominant side is better at coordinated, efficient movement and the non-dominant side at holding a position steady. He built a two-sided neural motor controller in the MotorNet simulator, gave each side a different loss function, and tested it on reaching and holding tasks. Each side came to beat the other on its favoured task, though the advantage over a single-sided controller was not statistically significant. Code is released on GitHub.

Luke’s Honours project at Monash built World Model + Replay for model-based continual RL (WMAR), on the Dreamer architecture, testing the idea that hippocampal replay builds a world model rather than a policy. Adding a long-term, distribution-matching replay memory beside the usual first-in-first-out buffer prevented catastrophic forgetting with far less memory and compute: on Atari tasks with no shared structure, average forgetting fell from 0.68 to 0.10. It is not a complete fix, since some tasks still lose performance and tasks with very different reward scales need reward scaling.

Yanfeng began his Masters in July 2022, extending the work on bilateral deep learning that Chandramouli Rajagopalan developed: deep networks with two hemispheres of different specialisation.

Abdelrahman began a PhD at UNSW in February 2022, jointly supervised by Cerenaut, on using hippocampal modelling of episodic memory to improve continual learning. He worked on it for about six months.

Chandramouli built version 2 of the bilateral networks as an intern: deep networks with two hemispheres, the left trained on local features and the right on global features, whose outputs a network head combines with weighted attention. On image classification with ResNet and VGG backbones the bilateral design beat baselines of similar capacity, except a conventional ensemble of unilateral networks trained on the same objectives. The paper, with David Rawlinson, Elkhonon Goldberg and Gideon Kowadlo, appeared at ACAIN 2024.

Amir’s Monash Masters thesis built the first left/right neural networks, in response to the 2018 Request for Research on networks inspired by our bilateral brains. A right and a left stream were trained on CIFAR-100, and the question was whether they would specialise, one on broad classes and one on narrow ones, by varying hyper-parameters alone. Sparsity of the hidden layer and how many layers and filters each side gets drove specialisation, but the specialised model did not beat a standard single-stream network.

Mohit’s Monash Masters thesis, “Improving Episodic Memory with Disentangled Representations”, built on the Artificial Hippocampal Algorithm (AHA). He replaced AHA’s long-term memory autoencoder with a Beta-VAE that learns disentangled representations, and tested one-shot, 20-way classification of unseen Omniglot characters. Long-term memory accuracy rose by about five percentage points (64% vs 59%), but the hippocampal components did not improve as hoped, so they could not yet exploit the richer representation.

An’s project used reinforcement learning to learn attentional strategies for the Recurrent Sparse Memory (RSM).

Kok-Ping, a Monash summer student, continued the self-organising project over the 2019 to 2020 summer, after Satya Borgohain completed his Masters.

Satya’s Monash Master of Data Science minor thesis (2019) produced the self-organising neural network hierarchy, in collaboration with Project AGI, and was later published at AI 2020. A meta-network of small autoencoders adjusts where each one looks in the image, using a rule that maximises the entropy of their encoded representations. On MNIST that entropy tracked classification accuracy, and randomly initialised architectures reorganised into better-performing hierarchies within a few steps, competitive with hand-designed ones. He presented it as a proof of concept.