GrowNet
GrowNet explores a neuron-centric, biologically inspired alternative to heavyweight deep-learning pipelines: local event-by-event learning, interpretable slot memory, and controlled capacity growth when true novelty appears.
GrowNet is under active development and is not yet deployed in NektronAI products. Architecture details and benchmarks will continue to evolve as experiments mature.
How GrowNet works
Instead of one global training phase that nudges millions of parameters at once, GrowNet treats learning as local updates inside neurons. Each neuron maintains a small table of memory slots that specialize to recurring patterns and can create new slots when truly novel signals appear.
Local, event-driven learning
Learning happens incrementally per event tick, enabling always-on adaptation instead of periodic monolithic retraining.
Slots as inspectable memory
Slot tables are designed to be readable so researchers can inspect what each neuron learned and how it changed over time.
Controlled growth mechanics
Capacity can expand from slots to neurons, layers, and regions under explicit growth rules, rather than fixed static size.
System framing
The architecture is organized as Region → Layers → Neurons → Synapses, with execution occurring per tick. It also supports transient modulation and inhibition signals through a bus-style mechanism.
Execution model
- Discrete tick-based processing
- Per-neuron local slot updates
- Deterministic growth triggers
Research ergonomics
- Cross-language parity: Python, Java, C++, Mojo
- Comparable behavior across implementations
- Designed for repeatable benchmarking
The “Golden Rule”
GrowNet’s guiding principle is intentionally simple so adaptation and growth stay controlled and testable.
“When something truly new shows up, make room. If it isn’t truly new, improve what you already have.”
Research direction and ambition
GrowNet is positioned as a serious research program with explicit benchmark goals for accuracy, robustness, and efficiency. The ambition is publishable, evidence-backed work, targeting peer-reviewed venues once results are mature.
Adaptation
Stay responsive to distribution shifts through local online updates rather than full retraining cycles.
Efficiency
Allocate capacity only when novelty warrants it, keeping compute and memory tied to observed complexity.
Interpretability
Prefer inspectable mechanisms such as slot usage, growth events, and wiring decisions over opaque global updates.
How it connects to NektronAI products
NektronAI’s applications are developed independently today. As GrowNet matures, the long-term goal is to use validated mechanisms where they improve practical systems without forcing premature integration.
Current applications
DeepTrading.ai, InterviewHelperAI, and TagMySpend.com act as real-world proving grounds.
Research-product loop
Products surface edge cases and constraints that feed back into research, while validated research can later inform new product iterations.
Interested in collaborating?
If you’re a researcher or engineer interested in the direction—interpretability, event-driven learning, growth policies—reach out.
What helps
- Clear problem framing and evaluation plans
- Benchmark design and measurement hygiene
- Systems engineering for reproducibility