EI Weight Lifecycle

Goal

The full lifecycle remains a proposed reduced-model study. Exp154 implements capacity and consumption-only lifetime tests, plus one fixed abstract genome, internally driven prenatal development, juvenile food learning and slower adult learning. Its animation replays the phases, native neural activity, choices, energy, deaths and births; the accompanying plot tracks prediction error and pathway weight norms. Evolution, viable weight-space mapping and genetic–environmental comparisons remain proposed. The current demonstration uses one seed and one fixed cue–energy mapping; the broader comparisons below are future work.

Investigate how evolution, development and experience jointly shape excitatory-to-inhibitory (E→I) and inhibitory-to-excitatory (I→E) synaptic strengths in a simple model organism receiving repeated object exposures. Identify how the region of weights compatible with stable, effective behaviour changes across the lifecycle, and how inherited rules and individual experience guide trajectories through that region.

Methods

1. Model organism and environment. Present a sequence of objects to a model organism that decides whether to consume or reject each object. Each object has one observable scalar sensory cue and a separate, hidden food energy change between −1 and +1 in normalised energy units, before metabolic costs. The cue predicts the energy change without directly revealing it. For the initial test, draw cues $x_t$ uniformly from $[0,1]$ and set

   $$\Delta E_t=s(2x_t-1),\qquad s\in\{-1,+1\}.$$

   Here $t$ indexes object exposures, $x_t$ is the observed scalar cue and $\Delta E_t$ is the hidden food energy change. The mapping sign $s$ is fixed within each lifetime; test both signs across lifetimes. Reveal the food energy change only through consumption; rejection reveals no outcome. Track an organism-level energy reserve outside the neural network:

   $$E_{t+1}=E_t-c+a_t\Delta E_t.$$

   $E_t$ is the reserve before exposure $t$, $c>0$ is the metabolic cost per exposure, and $a_t=1$ means consume while $a_t=0$ means reject. Start with a positive reserve $E_0$ and terminate the lifetime when it reaches zero or below. Choose the cost and initial budget so that rejecting everything cannot sustain the organism, while informed food selection can sustain it on average. Initially, the reserve is not an additional sensory input; a hunger signal is a possible later extension.

2. Neural controller. Use a small recurrent network containing excitatory and inhibitory rate neurons, initially around 16 excitatory and 4 inhibitory cells, to learn $V(x)$, the expected food energy change given observed scalar cue $x$. Consume when the prediction is positive, with occasional exploratory consumption. Fix sensory connections, the value readout and E→E/I→I weights within each lifetime; initially adapt only E→I and I→E strengths, with bounded, sign-preserving updates driven by prediction error after consumption. Gradients through the rate dynamics provide an initial engineering implementation, not a claim of biological plasticity. This makes adaptation of inhibitory feedback causally relevant rather than allowing a trainable readout to solve the task alone.

   Inhibitory homeostatic plasticity is a related modelling precedent; the proposed gradient-based learner does not establish that biological mechanism.

3. Inheritance and development. Represent the genome as inherited parameters controlling initial connectivity, synaptic strength distributions, plasticity and maturation timing. This is an abstract genetic representation rather than a simulation of specific genes. Simulate a prenatal analogue dominated by internally generated activity, postnatal development incorporating sensory experience, and adulthood with continued learning.

   Exp154 now implements a reduced version: restore inherited weights at conception, run 12 internal-drive windows minimizing per-cell activity error around 0.25, then learn from consumed food for 20 juvenile encounters before reducing the adult learning rate. Only E→I and I→E weights adapt. Each birth starts a fresh food optimizer and energy reserve; acquired weights are not inherited.

4. Evolution across generations. Evaluate each genotype across multiple lifetimes and environments. Select and mutate inherited parameters according to energy acquisition, survival and neural activity costs, without prescribing a preferred E/I ratio. Offspring develop afresh without inheriting acquired synaptic weights.

5. Lifecycle weight-space mapping. At developmental checkpoints, systematically vary E→I and I→E coupling strengths to identify regions supporting stable activity and stage-appropriate function. Distinguish immediate tolerance with plasticity frozen from recovery with plasticity active, and overlay naturally developed trajectories. Measure per-neuron excitatory and inhibitory inputs, their cancellation, prediction accuracy and firing stability: the E→I/I→E weight ratio is not itself E/I balance. Show the changing regions and trajectories in a lifecycle plot with phases marked. Define viability criteria before execution; sampled weight distributions are not themselves bounds on viable weights.

6. Genetic and environmental comparisons. Cross genotypes, experience histories and developmental random seeds to estimate their contributions and interactions. Include controls freezing plasticity in either pathway or disabling developmental activity. Restricted plasticity tests whether E/I adaptation can support learning; later allow other connections to learn to test whether evolution favours this solution when alternatives are available. Interpret the resulting contributions within the model and its sampled environments, rather than as a universal genetic–environmental percentage.

Expanded validation still proposed

The implemented single-mapping demonstration does not establish performance across conditions. Before adding evolution, test whether the restricted E/I controller can represent and learn both cue–energy mappings. Run separate lifetimes for each mapping sign with multiple initialisations, comparing the learning controller against the same network with plasticity disabled and an oracle that consumes only positive-energy objects. Match object sequences, initial energy and metabolic cost across comparisons. Log prediction error, energy, firing activity and pathway strengths. Evaluate prediction error across all presented cues using outcomes available to the simulator, while exposing outcomes to the learner only after consumption. Require improved performance across both mappings without unstable neural activity before making broader learning claims; failure first calls for checking representational capacity and the learning rule.