Emeron Labs · Independent research

Harnessing neural circuit skills for frontier RL models

What would it take to give a learning agent a circuit’s capacity for pattern separation, timing, or controlled forgetting? I simulate neural circuits, capture their skills in compact phenomenological models, and test whether those skills can improve reasoning and reduce training demands in RL agents guided by active inference.

Explore the pipeline ↓
Illustrative skill · Controlled forgetting

From circuit simulation to RL integration

A selective decay rule moves from circuit to agent A circuit readout shows one weakening association and one stable association. A compact model fits the weakening trace. The fitted rule enters the agent’s belief update before policy selection. Circuit simulation Two learned associations Phenomenological model Task changes Circuit readout Fitted rule Active-inference RL agent Update beliefSelect policy Decay rule Act and observe again
Proposed experiment

Can the agent adapt faster while retaining what still works?

Compare three agents

Reverse one cue–outcome mapping and keep another stable. Compare the agent using controlled forgetting, the same agent with the decay rule disabled, and a matched RL baseline.

Measure capability and cost

Measure recovery on the changed mapping and accuracy on the stable one. Test reasoning on unfamiliar tasks, then compare interactions, training updates, and compute to reach the same performance threshold. Test at controlled scale before extending to frontier RL models.

01 / Active inference & reinforcement learning

Emeron Labs

Emergence of inference.

Can an agent learn more from each experience? I study how to capture circuit functions in compact models and incorporate them into reinforcement-learning agents through active inference.

My hypothesis is that selected functions could improve learning and reasoning while reducing training interactions, updates, and total compute. I aim to test these effects separately in controlled tasks before investigating frontier-scale RL.

WORKING HYPOTHESES · OPEN TO DISCUSSION Discuss a hypothesis ↗
01 What does neural architecture contribute to inference? +

I want to examine how circuit organization, recurrent dynamics, and interactions across timescales constrain learning and behavior. The central question is which architectural features provide useful mechanisms for active inference, and which effects can be captured by simpler phenomenological models.

02 How should an agent choose what to learn? +

My hypotheses distinguish expected information gain before a choice from learning realized afterward. Proposed comparisons examine an expected-free-energy selector alongside an RL-value-plus-information-gain hybrid, with explicit attention to exploration, task performance, and computational cost.

03 How does experience become reusable behavior? +

I am exploring how recurring sequences can become executable skills through discovery, training, effect modeling, and validation. A related question is whether offline consolidation and multi-timescale organization can support more reliable long-horizon behavior.

04 Which biological principles could improve RL training? +

The aim is to translate candidate mechanisms into controlled computational experiments: compare them with matched RL baselines, isolate their contributions, and measure learning efficiency, transfer, robustness, and training cost. Relevance to commercial frontier models is a research objective; these are hypotheses to investigate.

02 / Research journey

How does intelligence emerge?

The methods have changed. The underlying question remains: how does organized neural activity become meaningful function?

In vitro

Engineering living neural systems

Bioengineering PhD · University of Illinois

Building three-dimensional neural tissue and studying electrophysiology, development, and neuromuscular interactions grounded my questions in measurable biological systems.

Neural tissue Electrophysiology Emergent function
Computational

Simulating circuits and learning dynamics

Computational neuroscience · UT Austin

Cerebellar simulations and neural data analysis extended that work into models of learning, forgetting, and extinction, using C++, CUDA, and Python.

Circuit simulations Learning dynamics GPU computing
Independent
research · Now

Emeron Labs: emergence of inference

Active inference · Phenomenological & circuit modeling

Bringing biological architecture, Grossberg-inspired neuroscience, and active inference into a research program on neural function and the mechanisms that could improve reinforcement-learning training.

Working hypotheses Cross-scale models Reinforcement learning
03 / Experimental foundations

Emergence, observed
in living systems.

Earlier coauthored work: high-resolution microscopy, electrophysiology, and neuromuscular junctions that ground my interest in neural function.

Structure 01
Scanning-electron and fluorescence microscopy of engineered motor-neuron networks

Building living
neural systems.

Engineering three-dimensional neural tissue and examining how geometry, development, and stimulation relate to its organization and function.

Explore the microscopy ↗
Scientific Reports, 2020 · Figure 2 · Paper ↗
Dynamics 02
Measured electrophysiology traces and activity maps comparing neural stimulation conditions

From signals
to coordinated activity.

Using electrophysiology and computational analysis to investigate neural firing, synchrony, plasticity, and the effects of developmental stimulation.

Explore the recordings ↗
Scientific Reports, 2020 · Figure 4 · Paper ↗
Co-founder, Emeron Labs

Gelson J. Pagan Diaz

Portrait of Gelson J. Pagan Diaz

At 21, studying entropy and protein folding led me to a question that has stayed with me: how do physical processes give rise to something as remarkable as a thought?

That curiosity took me from mechanical engineering and theoretical-physics interests at the University of Puerto Rico to a PhD in bioengineering at Illinois, where I worked with living neural systems. Computational neuroscience at UT Austin then gave me a way to investigate circuit dynamics and learning through simulation.

I co-founded Emeron Labs with a colleague to pursue questions like these. I am especially interested in discussing active-inference hypotheses, connecting phenomenological descriptions to neurophysiological mechanisms, and identifying experiments that can tell competing explanations apart.

Request CV ↗
NSF Graduate Research Fellow UT Austin Provost’s Early Career Fellow Active Inference Institute Fellow

Which circuit function would you test?

Suggest a circuit mechanism, a compact model, or a task that could distinguish competing explanations.

Start a research conversation ↗