A neural ecosystem that grows,
prunes itself, and sleeps

Research Proposal

Three things prior research all missed

NEAT · Progressive Networks · DEN · EWC — attempts at "networks that grow" and "networks that never forget"

Over the past decades, many studies tried to build networks that grow and networks that never forget, yet all halted before the same wall: the clash between stability and plasticity, and the impossibility of differentiating through structural change. Neutom Network turns to the three things they all left out — things a living brain has, and today’s artificial networks lack.

Research Proposal

An ecosystem of four elements

Not just nodes and weights — a place where computation, connection, care, and mobile regulatory data live together

Research Proposal

Forgetting is not a flaw but a function

The balance of memory and forgetting — the heart of this design

Keep what should be kept; let go of what should be let go.

In today’s neural networks, “forgetting” means catastrophe — catastrophic forgetting, where new knowledge overwrites and topples the old. Neutom Network sees forgetting differently. Depending on resource conditions, a Curia lossily compresses concepts and puts weak connections into dormancy (lulling them rather than cutting them), tidying the network’s structure in idle time. When forgetting becomes a controlled function, a network can grow without limit on finite resources.

Unverified Hypothesis

From primitive information to concepts — folding

From indivisible root information to a single concept Neutom

Neutom Network starts from indivisible primitive information — the number of kinds is not fixed but a tentative figure on the order of a few hundred, varying with research. Used together again and again, these form local Neutom networks, and a local network that has hardened enough is folded and referenced as a single concept Neutom — much as a new headword is added to a dictionary. And when needed, that concept can be unfolded again (plasticity) and relearned.

Unverified Hypothesis

Five operating modes — and the sleep of Neutom Network

Awake · Default · Maintenance · Regeneration · Deep sleep

Neutom Network does not stay in one state but moves among five modes: Awake, working alertly; Default, wandering loosely; Maintenance and Regeneration in idle time; and Deep sleep, which lowers resources to a minimum. Maintenance calms overexcitation, renormalizes the sum of Sunya strengths, and reclaims long-dormant connections; Regeneration reactivates and recombines memory paths and searches for the seeds of generalization. Deep sleep keeps the state and leaves only the least circuitry needed to wake by itself — which is why the floor is not zero. The imbalances that pile up while awake, it mends itself while asleep.

Unverified Hypothesis

The network decides for itself when to sleep

Not an outside schedule — the network’s own condition calls for sleep

Attempts to give artificial neural networks a time of regeneration and tidying have already been many. In most of them, though, when to do it is decided from outside — a fixed period, or the external partition between one task and the next. This is the place where Neutom Network wagers differently. It puts the trigger of sleep in the network’s internal state.

A Curia reads five things — accumulated prediction error, imbalance in the sum of weights, the proportion of dormant connections, resource pressure, and the recent relearning failure rate. All five are obtained inside the network, so no outside schedule has a place in this procedure. The strength of external input is not among the five — that is a separate condition returning the network to Awake from any mode, and it does not enter where it is decided whether to fall asleep.

Entering because performance has degraded is not the same as entering because resource pressure has built up. A design that calls maintenance on a signal of rising loss, and one that calls it when resources and structure reach their limit, are measuring different things even when they do the same work. The Neutom Network aims at the latter.

The separation of three layers holds this design up. What flows along a connection is divided into three — activation signals, information objects, and regulatory signals — and one requirement is stated explicitly: with topology and weights left as they are, the operating mode must be shiftable by the composition of Naruon alone. If that requirement does not stand, no ablation experiment that follows can say what it was that had been removed.

Unverified Hypothesis

Naruon pharmacology — a testbed for AI psychopharmacology

Agonist · Antagonist · Tolerance · Withdrawal

Likening Naruon to neurotransmitters, we can borrow the conceptual framework that biological pharmacology has built and handle the network’s behavior far more finely. With a Naruon that raises arousal and a paired Naruon that lowers it, not only injection but also release becomes possible, and regulation turns bidirectional.

Once tolerance and withdrawal are implemented, we can observe phenomena such as “this network has formed a dependence on a particular Naruon.” This is not a side effect but a new testbed for diagnosing the network’s internal state and studying alignment and safety — what does it do when you inject this, how does it waver when you cut that. Pharmacological intervention becomes a handle for looking inside Neutom Network.

Unverified Hypothesis

Pain and curiosity — a mind that wants to learn on its own

Intrinsic motivation, as Naruon

Balancing these two signals so they do not tip into avoidance or distracted exploration is still an open problem. But the direction is clear — the fuel of Neutom Network is not an external whip, but pain and curiosity arising from within.

Unverified Hypothesis

Honest open problems, and the larger picture

This design is not a finished feature but an open research question. How to partition the scope of influence at a Naruon branch; how to adapt the pruning threshold; how to run asynchronous, stateful packets efficiently on today’s hardware; how to keep a self-prescribing policy network from destabilizing against the main network — we do not hide the questions we cannot yet answer. Until positive results are in, we do not speak of hypotheses as achievements.

Neutom Network does not stand alone. The language that describes how information and concepts form and transform upon it is JIDEL; the theory that sees the mind as “a structure in which information flows and changes among elements” is BOSIFWEM; and the ultimate direction where these three converge is ECONAS — an evolving, conscious and conscientious artificial system.