Gamma neural model
Architectures AIMech:GammaNeuralModel DRAFT
Aliases: Gamma neural network, Gamma net, Gamma neural net
A neural architecture family that represents delayed inputs or neural outputs by weighted gamma-kernel memory traces, realized through cascades of first-order leaky integrators.
Identity Notes & Concept Scope
Compared all 442 pinned AIO and 580 native IDs, names, aliases and scoped synonyms; no normalized identity match. Ignored- and hidden-inclusive fixed-string searches of the recovery, older discovery and original Documents checkouts returned 3, 0, 0 lines. All distinct match contents were inspected. Semantic matching covered 65 AIO and 156 native definitions, with unchanged preceding rows verified and new matches read. 4 full annotated AIO node/edge/generated contexts and 4 full native records were inspected. Prior matches carry the temporal-kernel report’s pending Gamma lead. Cascaded gamma traces differ from SCRN’s separate exponential context, TKRNN’s factored parallel temporal kernels, Clockwork’s module schedule and NNPDA’s external fractional-symbol stack. Focused gamma input-memory placement is a source-defined configuration within the broader architecture. A gamma distribution in a separate probabilistic method and the unrelated consistency-learning Gamma-model cohort do not imply identity. Full recurrent-network scope supports the broader mapping; generic feedback, convolution and simple recurrent-layer contexts do not supply an exact or more appropriate broader match.
Claims & Grounded Evidence
6 Claims
Mechanism
#continuous-gamma-cascade
For positive fixed decay rate mu, the continuous kernels are g_k(t)=mu^k*t^(k-1)*exp(-mu*t)/(k-1)! for t>=0 and positive integer k. Convolutional traces with zero initial history obey d y_k/dt=mu*(y_(k-1)-y_k), where y_0 is the driving neural output. Their weighted combination supplies delayed interactions in the neural-state equation, alongside any direct term. Each basis kernel has unit integral, but arbitrary signed weight matrices do not make the combined interaction a probability distribution. A finite kernel expansion is not an exact representation of every possible temporal interaction; the source’s approximation discussion has decay and increasing-order assumptions.
Mechanism
#discrete-memory-and-delay-boundaries
The discrete memory uses x_0[n]=u[n] and simultaneous updates x_k[n]=(1-mu)*x_k[n-1]+mu*x_(k-1)[n-1]. For zero initial state and 0<mu<1, its k-stage impulse response is binomial(n-1,k-1)*mu^k*(1-mu)^(n-k) for n>=k, a dispersed rather than exact delay. At mu=1 it becomes a tapped delay line; at mu=0 no new input enters the stored traces and existing initial traces remain frozen. For the positive-kernel regime the mean-depth parameter K/mu describes the last memory stage, not a hard cutoff, exact retrieval horizon or proof of preserved information. Small mu trades temporal resolution for depth.
Limitation
#stability-scope
For fixed mu the discrete stage pole is 1-mu, giving asymptotic and bounded-input bounded-output stability of the linear memory for 0<mu<2. The endpoint mu=2 is excluded: an alternating bounded input gives a stage response whose magnitude grows linearly. Mu=0 does not forget a nonzero initial state. These boundaries qualify the earlier paper’s inclusive endpoint wording; the focused-network paper uses the strict interval. Values 1<mu<2 have alternating-sign memory weights, not positive averages. Stability of a fixed linear memory does not establish stability of an arbitrary coupled nonlinear network, an unconstrained adaptation trajectory or its optimizer.
Scope
#focused-network-configuration
The focused gamma network places the gamma memory at the input and sends its traces to a static nonlinear feedforward mapper. It is a source-defined configuration of the gamma family, not the whole family or an independently averaged set of temporal models. Its spatial mapper is feedforward, while the memory still has recurrent dynamics. The more general formulation can filter neural-state outputs, and additional placements are discussed as extensions.
Limitation
#learning-choices-and-optimality
The original formulation gives a temporal Hebbian weight rule using locally available gamma traces. The focused configuration instead trains spatial weights with ordinary backpropagation and gives a recursive sensitivity for the memory parameter, a_k[n]=(1-mu)*a_k[n-1]+mu*a_(k-1)[n-1]+x_(k-1)[n-1]-x_k[n-1], with an appropriate initial sensitivity; backpropagation through time is also discussed. These are learning choices, not one mandatory training algorithm. The displayed fixed-parameter sensitivity does not by itself guarantee exact gradients under parameter drift, successful optimization, automatic selection of integer order K or an always-optimal memory depth.
Scope
#evaluation-boundaries
The focused-network experiments concern five-step prediction on a selected EEG segment, a specified Mackey–Glass series and identification of one linear low-pass system. EEG comparisons scan discrete mu settings rather than demonstrate successful learned adaptation of mu in that experiment. The Mackey–Glass example adapts mu online under selected capacity, sample and epoch choices; the conclusion describes its performance as similar to TDNN in that short-memory setting. Benefits in the chosen linear-identification example do not establish universal superiority, always fewer required parameters, biological causality or unlimited long-term memory. These results are for focused configurations, not a validation of every gamma architecture.
⚠️ Curated Limitations & Negative Results
- The record covers the 1990 NeurIPS formulation and 1991 focused-network proceedings paper, both read completely. The later 1992 journal article was identified but its full text was not available from the attempted copy; it is not treated as inspected evidence.
- Continuous-time rates and dimensionless discrete-time recurrence coefficients have distinct conventions. No arbitrary discretization step, numerical integration scheme or executable implementation is supplied here.
- Finite order restricts the kernel family and memory is generally dispersive. Depth metrics and approximation statements do not prove exact reconstruction of an arbitrary past or finite-order exactness for every convolution.
- The strict fixed-coefficient stability interval concerns the linear memory, not all parameter adaptation or nonlinear feedback. Endpoint behavior and signed responses above mu=1 remain explicit.
- Focused placement, temporal Hebbian learning and backpropagation/RTRL choices are scoped configurations. Optimization claims in the source are not adopted as universal guarantees or evidence of automatic order selection.
- All seven theory-paper pages and all eight focused-paper pages, including figures and references, were visually read. Constructed recurrence, kernel and derivative checks involve no fitting. No author code, model training, runtime benchmark or empirical replication was run.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:RecurrentNeuralNetwork
Recurrent Neural Network
|
broad_match | The gamma family realizes temporal state through recursive memory dynamics. Even the focused configuration has recurrent memory feeding a spatially feedforward mapper. The AIO recurrent-network category is broader than this particular gamma-kernel architecture family. |
Evidence Sources & Primary Literature
2 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
devries1990
|
Bert de Vries, Jose C. Principe
|
primary paper
full_text
|
sha256:e5522677a922…
|
principe1991
|
Jose C. Principe, Bert de Vries, Jyh-Ming Kuo, Pedro Guedes de Oliveira
|
primary paper
full_text
|
sha256:696568f2a68f…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-13T06:20:04Z
Updated: 2026-10-13T06:24:08Z
Source File:
data/curated/GammaNeuralModel.yaml