| name | embodied-neurocomputation-framework |
| description | Embodied Neurocomputation framework for interfacing biological neural cultures with scaled task-driven validation. Systems-level approach to multi-variable optimization of encoding/decoding between silicon computing and living biology. Demonstrates that biological neural networks (BNNs) can outperform DQN agents in goal-driven navigation when encoding parameters are properly optimized.
|
| category | neuroscience |
| tags | ["biological-neural-networks","neurocomputation","MEA","encoding-decoding","bio-silicon","hybrid-computing","parameter-optimization"] |
| related_skills | ["embodied-neurocomputation-framework","neural-digital-twins-bci","neural-brain-framework","energy-based-neurocomputation"] |
| activation_keywords | ["embodied neurocomputation","biological neural network computing","MEA neurocomputation","bio-silicon computing","biological neural culture interfacing","cortical labs CL1","neural encoding optimization"] |
Embodied Neurocomputation Framework
Paper: Embodied Neurocomputation: A Framework for Interfacing Biological Neural Cultures with Scaled Task-Driven Validation
Authors: Johnson Zhou, Daniel Tanneberg, Forough Habibollahi, Alon Loeffler, Kiaran Lawson, Valentina Baccetti, Kwaku Dad Abu-Bonsrah, Candice Desouza, Finn Doensen, Bradley Watmuff, Daria Kornienko, Azin Azadi, Justin L. Bourke, Bernhard Sendhoff, Brett J. Kagan
Institutions: Cortical Labs (Australia), Honda Research Institute Europe (Germany)
arXiv: 2605.13315 (May 13, 2026)
Category: cs.ET, cs.LG, cs.NE
Overview
This paper introduces a formal Embodied Neurocomputation Framework — a systems-level approach to interfacing biological neural networks (BNNs) with conventional computers via Micro-Electrode Arrays (MEAs). The framework conceptualizes the digital-biological interface as a multi-variable optimization problem across four interdependent modules: encoding, biological transformation, decoding, and feedback. It validates this through the first large-scale parameter optimization of encoding configurations for BNN agents performing closed-loop navigation.
Core Framework
Mathematical Formulation
Neurocomputation f at time t, with parameter set theta_t, as sequential feed-forward mappings:
y_t = f(x_t; theta_t) = d(b(e(x_t; theta_e); theta_b,t); theta_d)
Where:
- e(x_t; theta_e): Encoding — transforms task information into electrical stimuli
- b(u_t; theta_b,t): Biological transformation — BNN's intrinsic dynamics
- d(v_t; theta_d): Decoding — transforms neural responses into task-relevant outputs
- r(Score; theta_r): Feedback — drives BNN adaptation toward objectives
Biological adaptation:
theta_b,t+1 = g(r(Score; theta_r); theta_b,t)
Four Key Modules
1. Encoding (theta_e = theta_task union theta_stim)
- Transforms task-specific information into stimulation matrix u_t in {0,1}^(C x tau_in)
- C: stimulation channels, tau_in: pulse delivery time steps
- Parameters: frequency, amplitude, pulse width, waveform morphology, spatiotemporal distribution
- Rate encoding: sensor value -> sequence of stimulations at interpolated frequencies
2. Biological Transformation (b)
- Non-stationary mapping depending on temporal structure of input
- Parameters theta_b evolve according to stimulation/response history, feedback, and spontaneous processes
- Produces qualitative change: reorganizes informational structure, not just scaling/filtering
- "Third-order" information-processing system: response depends on how previous outputs shaped transformation
- Non-invertible: reflects reorganization and compression of information
3. Decoding (theta_d)
- Inverse of encoding: transforms BNN responses into task-relevant formats
- Response matrix v_t in R^(C x tau_out) -> output
- Count decoding: spike counts aggregated in spatial regions, normalized against baseline spontaneous activity
- Action with highest relative spike density is executed
4. Feedback (theta_r)
- Special form of encoding designed to drive BNN adaptation
- Reinforcing (r+): structured bursts for favorable outcomes
- Plasticity-inducing (r-): random stimulation to encourage alternative mappings
Empirical Evaluation
Task: Goal-Driven Navigation
- Simulated 6x6 gridworld with barrier, food source, and odor gradient
- Agent actions: move forward, turn left, turn right
- Scalar sensor: odor strongest to left (-1), front (0), right/behind (1)
- Three evaluation modes: 30 steps/1 episode, 150 steps/1 episode, 30 steps/5 episodes
Experimental Setup
- 26 BNN cultures via Cortical Labs CL1 platform
- Distributed optimization: Optuna HPO server + multiple CL1 clients
- 1,296 parameter combinations screened
- 4,000+ hours of real-time agent-environment interactions
- Two-stage screening: Stage 1 (n=1,296 -> n=64), Stage 2 (n=64 -> n=12 top configurations)
Encoding Parameters Screened
| Parameter | Stage 1 Values | Top (n=12) |
|---|
| Min Frequency (Hz) | 2.0, 3.0, 4.0, 5.0 | 4.0 |
| Max Frequency (Hz) | 40.0, 60.0, 80.0, 100.0 | 40.0, 60.0, 80.0 |
| Amplitude (uA) | 1.0, 2.0, 2.5 | 2.5 |
| Pulse Width (us) | 40.0, 80.0, 160.0 | 40.0, 80.0 |
| Tick Rate (Hz) | 1.0, 2.0, 4.0 | 1.0, 2.0 |
| Ticks per Step | 2, 4, 8 | 4 |
Key Findings
- Maximum frequency is strongest driver: favors moderate values (40-60 Hz)
- Higher amplitude, shorter pulse width, faster interaction rates support improved performance
- BNN agents significantly outperform DQN benchmarks under equivalent training steps
- 12 configurations consistently demonstrated learning across multiple episodes
SHAP Analysis Results
- Max Frequency: strongest positive impact on top 1% performance
- Amplitude: higher values (2.5 uA) favored
- Pulse Width: shorter (40 us) preferred
- Ticks/Step: moderate (4) optimal
- Tick Rate: lower (1-2 Hz) better
- Min Frequency: moderate (4 Hz) optimal
Biological Setups Tested
- Group 1: 7 cortical/hippocampal cultures in PDMS ring
- Group 2: 9 cortical/hippocampal cultures on astrocytes, monolayer
- Group 3: 5 cortical-only cultures, monolayer (Stage 2)
- Group 4: 5 cortical-only cultures, monolayer (Stage 2)
Applications
- Hybrid bio-silicon computing: Bridging biological efficiency with silicon programmability
- Robotic control: BNN-driven adaptive decision-making for embodied agents
- Neuroscience research: Understanding biological learning mechanisms through task-driven validation
- Energy-efficient computing: Alternative to von Neumann bottleneck limitations
- Benchmarking framework: Establishing field-wide standards for neurocomputation
Framework Principles
Systems Thinking
- Each component is interconnected and highly parameterized
- Adjusting any single part changes the entire system response
- Configuration is a multi-variable optimization problem
Learning vs Training Distinction
- Learning: biological adaptation emerging from intrinsic biophysical plasticity
- Training: algorithmic optimization in artificial neural networks
- This distinction emphasizes that biological adaptation is fundamentally different from gradient-based updates
Hardware Agnosticism
- Framework encapsulates physical interactions within mapping functions
- Applicable to MEA, optogenetic, chemical, or other interfacing modalities
Key Insights for Practice
- Parameter optimization is essential: Heuristic/ad-hoc stimulation protocols are insufficient for robust BNN computing
- Biological stochasticity requires replication: Identical parameters must be evaluated across multiple cultures simultaneously
- Rate encoding with moderate frequencies (40-60 Hz max) provides the best coupling with BNN biophysics
- Feedback design matters: Structured bursts for reinforcement, random stimulation for plasticity induction
- Calibration time: 4,000+ hours of real-time interaction needed — not trivial
Comparison to Silicon-Based AI
| Aspect | BNN | Silicon DNN |
|---|
| Energy efficiency | Extremely high (mW range) | High (GPU/TPU watts) |
| Continual learning | Intrinsic | Requires specific techniques |
| Non-stationarity | High (adaptive, evolving) | Fixed after training |
| Parameter optimization | Biological (plasticity) | Gradient-based |
| Scalability | Limited by culture setup | Massive scale possible |
| Task performance | Surpasses DQN (this work) | Superior on complex tasks |
Pitfalls
- Biological variability: Each culture has unique dynamics; parameters optimized for one may not transfer
- Parameter space is vast: 6 encoding parameters x 4-5 values each = 1,296 combinations minimum
- Real-time constraint: Experiments run at biological speed (not accelerated)
- Hardware limitations: MEA platforms are expensive and require specialized expertise
- Decoding simplicity: Current count decoding is rudimentary; more sophisticated methods needed
- Feedback design: Fixed feedback regimen may not be optimal for all parameter configurations
References
- Cortical Labs CL1 platform: https://corticallabs.com/cloud
- Optuna HPO framework
- Count decoding from Cortical Labs prior work
- Related: Cortical Labs "DishBrain" (Pong-playing neural culture)
