From Neural Networks to Digital Brains: The Next Leap in AI • Daniel Lütgehetmann • GOTO 2025

Daniel Lütgehetmann, CTO at inait, presented at GOTO Copenhagen 2025, addressing the fundamental gap between current AI capabilities and the natural adaptability of biological systems. While modern AI excels in tasks like language processing, it struggles with the dynamic and complex physical world, a domain where even simple animals demonstrate superior competence.

The Problem with Current AI in Physical Environments

Lütgehetmann highlights that today’s robots, despite mastering specific tasks like doing dishes or folding clothes in constrained environments, fail dramatically when taken out of these controlled settings. The current paradigm relies on:

• Extensive Data Collection: Requiring vast datasets for training.
• Large Models and Infrastructure: Throwing computational resources at problems.
• Difficulty in Error Correction: When models fail in production, the solution often involves adding more sensors, scaling data, and increasing infrastructure.
• Lack of Skill Accumulation: “Current AI does not know how to accumulate skills adding more and more skills gets more and more difficult.” This contrasts sharply with biological learning, where acquiring new skills often makes subsequent learning easier, not harder.

Inspiration from Evolution: The Brain as a Blueprint

The solution, according to Lütgehetmann, lies in studying the brain, which has evolved over millions of years to navigate complex, unpredictable environments. Intelligence in nature combines innate abilities with rapid learning. He cites examples like a turtle instinctively knowing how to survive or an octopus escaping an unfamiliar jar through on-the-spot learning. This biological inspiration suggests a path to AI that doesn’t require billions of hours of video data for autonomous navigation or swarm-like coordination in systems like cars.

Digital Brains: A Biologically Accurate Approach to AI

Daniel Lütgehetmann’s journey with renowned neuroscientist Henry Markham led to the creation of “digital brains” at inait. These are not merely abstract neural networks but “actual copies of a real brain,” developed over 20 years with 300 million euros in funding and validated by 300 scientific papers and countless lab experiments. The goal was initially neuroscience research (e.g., Alzheimer’s, autism) but now extends to AI. These digital brains replicate:

• The precise map of neurons and synapses.
• The functionality, including chemical reactions at the synapse level.

This approach aims to explore the unique aspects of the brain that enable it to navigate the complex world, running inside computers, robots, and commodity hardware.

How Digital Neurons and Synapses Differ

Lütgehetmann contrasts traditional deep learning models with their biological counterparts:

Deep Learning Neuron:

• Simplistic.
• Takes inputs, sums them up, applies a nonlinearity.

Real Neuron (in Digital Brains):

• Complex Geometry: Intricate branches connecting to other neurons in specific ways.
• Computation: Does not perform math (addition, formulas) but integrates information into a continuous state (charge/electricity).
• Event-Based Communication: Communicates only when its charge reaches a certain threshold, triggering a “spike.” This introduces a crucial time dimension and local computation.

Deep Learning Synapse:

• Simple weight (a number) indicating connection strength.

Real Synapse (in Digital Brains):

• Dynamic Behavior: Much more complex than a simple weight. For instance, a facilitating synapse strengthens its impact with more activity, while a depressing synapse weakens it, even with the same “weight.” This dynamic adaptability is crucial for the brain’s ability to react to sudden changes in input data.

Integrating Digital Brains into AI Systems

To utilize digital brains for AI, inait developed methods for input, output, and learning:

• Encoding Data (Input): Sensor data (e.g., images, video, sinusoidal curves) must be translated into electrical stimuli for the digital brain. Examples include:
  • Intensity Encoding: Applying more electricity to neurons as a sensor value increases.
  • Geometric Encoding: Mapping signal levels to specific neuron locations (e.g., top neurons for high signal, bottom for low). This requires a fundamental shift in how data is conceived and processed.
• Reading Out Actions (Output): Interpreting the brain’s recurrent activity into actionable decisions. Methods include:
  • Measuring the activity frequency of neuron populations to derive values (e.g., robot arm torque).
  • Detecting activity spikes for anomaly detection.
  • Using competing populations of neurons where the relative activity determines the output.
• The Feedback Loop: The environment provides input, which is translated into electrical stimuli for the digital brain. The brain responds, its activity is read out as an action, which then affects the environment, restarting the loop.

Biological Learning: Beyond Backpropagation

The core innovation is a proprietary learning rule, the “conceptual equivalent of backpropagation,” but inspired directly by biology. Unlike standard AI, where the entire network is optimized, in the digital brain:

• Local Learning: “Every synapse is learning by itself,” yet they collaborate as a group to achieve the task.
• No Backpropagation: The system learns through local feedback, adapting rapidly.

This learning mechanism enables systems to quickly master tasks, as demonstrated by a cartpole (balancing stick) example where the system rapidly improves its balance over episodes.

Applications and Showcases

inai’s technology is developed in two stages:

Gen I: Forecasting (Prediction)

Leveraging the brain’s mechanisms for powerful time series forecasting across various domains (sensor data, demand, financial data). The company partnered with Microsoft to bring this to businesses, and an online demo is available for users to upload Excel sheets and receive predictions with confidence scores. This technology has shown “better forecasting accuracy… than if you use normal technologies.”

Gen II: Acting (Interaction)

This stage focuses on enabling AI to learn how to interact with and change the world. A compelling demo is presented using the game Pong:

• Traditional AI: Even the best deep learning models require tens or hundreds of tries for a single hit.
• Digital Brains: Learns to hit the ball by the second game, scores by the fourth game, and wins outright by the eleventh game. This represents an “astonishing” learning speed.
• Efficiency: A 500-neuron network runs on a laptop, while competitors (20-30 times slower) require a GPU for 12 hours.
• Robotics: This speed-up in learning (30-50x or more) is being explored with Microsoft for factory floor robotics. The digital brain naturally merges different sensor inputs (image, video, robot position) into electrical activity without requiring explicit merging logic.

The Future and Key Takeaways

Inait’s vision is to build AI with digital brains, starting small (e.g., 500 neurons, an ant has 250,000 neurons). The goal is not merely scale, but effective learning and interaction. This could lead to complex behaviors like swarm intelligence and autonomous 3D navigation without massive scaling.

Key Takeaways:

• Instant Learning for Complex Interaction: Real-world complexity demands rapid, on-the-fly learning, which current data/compute scaling struggles to provide, especially at the edge.
• Groundbreaking Hybrid AI: Combining the intelligence of even tiny animals with existing AI (LLMs, vision models) would be revolutionary, augmenting capabilities rather than replacing them.
• Digital Brain AI Works Today: It outperforms deep learning in dynamic, fast-learning environments, with showcases demonstrating its effectiveness.
• Broad Applications: Beyond robotics, this approach is relevant for complex dynamical systems in digital environments like manufacturing plants or anywhere data is scarce.

Lütgehetmann challenges the audience to consider the implications of an AI that learns like a child or a dog, dynamically, on a factory floor, without a massive data center, and how this could transform businesses and software development.

• Tags:
• Robotics
• Reinforcement Learning
• Feature
• Machine Learning
• Neuroscience
• Physical AI
• Adaptive Systems
• Digital Brains
• Biologically Inspired AI
• Neuromorphic AI
• Time Series Forecasting