Disclaimer
The content provided on this website is for informational purposes only and does not constitute professional advice. While we strive for accuracy, we make no guarantees regarding the completeness or reliability of the information. Users are encouraged to seek professional guidance for specific inquiries. By accessing this site, you agree to these terms.
The content provided on this website is for informational purposes only and does not constitute professional advice. While we strive for accuracy, we make no guarantees regarding the completeness or reliability of the information. Users are encouraged to seek professional guidance for specific inquiries. By accessing this site, you agree to these terms.
“Prigogine’s theory of dissipative structures parallels the Chronocosm claim that coherence is sustained not by equilibrium, but by continual reorganization within open, unstable systems that metabolize flux.”
Sublayer AI: Towards Brain‑Inspired Continual Learning with Offline Reflective Cognition
Lika Mentchoukov — January 15, 2026
Revised September 22, 2026
Status: Proposed architecture and research agenda. This article presents no new benchmark results. “Reflection” denotes computational replay, simulation, and evaluation; it does not imply conscious introspection.
IntroductionContinual learning asks how a system can acquire new capabilities while retaining useful prior knowledge under changing conditions. Sequential training can alter parameters that support earlier tasks, producing catastrophic forgetting. An architecture intended for long-term adaptation must therefore address both learning new information and preserving, correcting, or retiring earlier information appropriately.
Replay is already an established approach to this problem. Brain-inspired generative replay has been demonstrated on continual-learning benchmarks, including tasks with natural-image inputs. Sublayer AI builds on this research rather than claiming to introduce offline learning or generative rehearsal. [1]
The proposal combines an outer layer for interaction with an inner sublayer for scheduled consolidation and predictive simulation. Its potential contribution is the coordination of these processes: which experiences to retain, when to replay them, when imagined trajectories are useful, and how to determine whether an update should replace the deployed model.
The central hypothesis is that this coordinated cycle can improve retention and adaptation under defined memory and compute budgets. It remains to be tested against simpler replay methods and equally resourced baselines.
1. Neurocognitive Inspiration and Its Limits
1.1 Replay, Internal Activity, and Consolidation
Research linking memory replay with the default mode network motivates attention to internally generated activity during periods without immediate external demands. The cascaded memory systems account proposes connections between replay, the DMN, and memory organization; these connections should be distinguished from a fully established biological mechanism. Human experiments also investigate brain-wide activation associated with replay. [2, 3]
These findings motivate a design question: can a learning system use periods between interactions to revisit experience and assess candidate updates? They do not establish that a software replay module reproduces the DMN or that offline processing produces self-awareness.
1.2 Complementary Functions
The architecture adopts a functional distinction between retaining particular experiences and learning more general representations. A computational model by Spens and Burgess explores generative processes in memory construction and consolidation, providing a relevant example of how biological questions can inform machine-learning models. [4]
In Sublayer AI, a bounded buffer preserves selected observations, while a learned generator supports rehearsal. These are engineering roles inspired by memory research. Neither is a literal hippocampus or neocortex, and the analogy does not determine the correct implementation.
1.3 What the Analogy Does Not Establish
Human memory is not a guarantee of interference-free learning, and a biologically inspired design is not automatically effective. Replay, consolidation, internal simulation, and self-evaluation are distinct processes whose contributions should be measured separately.
The proposed controller is a scheduling mechanism. Describing its operations as “reflective” should not be taken as evidence of metacognition unless the system is explicitly evaluated for a defined capacity to monitor and regulate its own performance.
2. Architecture: Interaction, Consolidation, and Validation
Sublayer AI comprises two processing layers and an update-validation boundary. The outer layer supplies predictions or actions during interaction. The inner sublayer prepares candidate updates using retained and generated experience. A validation process determines whether a candidate is eligible for deployment.
The deployed model and the candidate model remain distinct. Learning during an interaction period may update a candidate without immediately changing the model serving users or controlling an environment. The permitted deployment cadence is a separate design choice.
2.1 Outer Layer: Task Performance and Experience Collection
The outer layer maps observations to predictions or actions. Each retained training record should distinguish observed inputs, actions, outcomes, available labels, and inferred or missing information.
The initial implementation should select a specific learning setting: supervised continual learning or reinforcement learning. Their loss functions, feedback, and evaluation requirements differ. A general architecture can accommodate both without assuming that one training procedure works unchanged across them.
2.2 Bounded Experience Buffer
A bounded buffer retains selected real observations or transitions. Its selection policy should be explicit: recent examples, representative coverage, rare events, or a defined mixture. Capacity must be counted in bytes as well as examples when comparing representations of different sizes.
Retention should reflect permissions and purpose. The buffer is not necessarily limited to recent experience; some older examples may be retained because they anchor evaluation or represent conditions that a generator reproduces poorly. Training records and held-out evaluation records must remain separate.
2.3 Generative Memory Module
A generative model produces rehearsal examples or internal representations associated with earlier experience. Candidate implementations include a conditional VAE or another suitable generator, selected for the task and resource budget.
The generator is itself a learned component and may forget. The design must specify how it is updated and evaluated. Possible mechanisms include retained real examples, frozen teacher checkpoints, or distillation against prior model behavior. A teacher can preserve earlier errors, so agreement with it is not sufficient evidence of correctness.
For supervised tasks, generated targets require a defined source: conditioning labels, teacher predictions, or another validated procedure. Synthetic labels must not silently become verified observations.
Training only on self-generated samples does not supply independent evidence that realism or coverage has improved. Evaluation should include real held-out observations and tests for loss of rare modes.
2.4 Latent Replay and Representation Compatibility
Replaying hidden representations can reduce the burden of synthesizing full inputs, as explored in brain-inspired replay research. [1] It also introduces a compatibility problem: a latent vector produced by one encoder may no longer have the same meaning after that encoder changes.
A Sublayer implementation should therefore specify whether it freezes the relevant encoder, regenerates representations from retained inputs, learns a tested translation between versions, or uses another compatibility mechanism. Record the encoder version associated with stored representations. Measure the consequences of representation drift instead of assuming latent replay remains valid indefinitely.
2.5 Goal-Conditioned Prediction and Imagination
Goal-Conditioned Predictive Coding (GCPC) learns trajectory representations that support goal-conditioned policy learning. Its published formulation should not be conflated with a complete simulator generating arbitrary future states, actions, and rewards. [5]
If Sublayer AI uses explicit imagined trajectories, it additionally requires an action-conditioned dynamics model, a policy or action-selection procedure, and definitions of rewards, termination, and rollout limits. Goals influence the selection of actions or trajectories; they do not ensure that a desired outcome is achievable.
Imagined transitions should remain labeled as model-generated. Their use should depend on measured predictive quality, the distance from observed experience, and uncertainty checks. Shorter rollouts, disagreement among models, or restrictions to supported regions are candidate controls, not guarantees against simulation error.
2.6 Controller and Context Gating
The controller determines when consolidation runs, allocates replay and simulation budgets, and chooses contexts to revisit. A first prototype can use a fixed schedule before testing adaptive scheduling based on drift indicators or retention checks.
Context gating may reduce interference, but it can also restrict useful transfer. The design must state whether context or task identity is supplied, inferred, or unavailable. Evaluation should include incorrect context assignment and conditions in which task boundaries are unclear.
An inferred context is a hypothesis. It should not become an unquestioned instruction that excludes relevant experience from rehearsal.
2.7 Update-Validation Boundary
The inner sublayer produces a candidate checkpoint. A separate evaluation process compares it with the deployed checkpoint using predefined criteria for retention, new-task performance, and domain-specific constraints.
A candidate that fails those criteria is rejected or revised. Promotion records should identify the training data versions, replay configuration, model versions, evaluation results, and authorized deployment decision. Recovery should preserve a known prior checkpoint while acknowledging that reverting model weights cannot undo actions already taken.
Suggested architecture-figure caption: The outer layer interacts with the environment and records selected experience. The inner sublayer uses bounded replay, generative rehearsal, and optional predictive simulation to train a candidate checkpoint. Held-out evaluation governs promotion to the deployed model; imagined experience is not treated as verified observation.
3. Training Protocol
3.1 Phase One: Online Interaction
The deployed model performs the authorized task. Selected observations enter a staging process that checks record quality, permissions, and available labels before admitting them to the training buffer.
The candidate learner may train on incoming observations together with replayed examples. The article does not assume that replay guarantees retention; the replay ratio, sampling policy, and their effects must be measured.
3.2 Phase Two: Offline Consolidation
During a scheduled consolidation period, the candidate learner combines retained observations with generated rehearsal examples. An optional predictive module supplies imagined transitions for tasks where model-based learning is appropriate.
“Offline” here means training without live interaction. It differs from the technical setting of offline reinforcement learning, in which a policy is learned from a previously collected dataset. A Sublayer implementation may use offline RL methods, but the two meanings should remain explicit.
A possible task-specific objective has the schematic form:
[
\mathcal L_{\mathrm{candidate}}=
\lambda_n\mathcal L_{\mathrm{new}}+
\lambda_r\mathcal L_{\mathrm{replay}}+
\lambda_i\mathcal L_{\mathrm{imagined}}+
\lambda_c\mathcal L_{\mathrm{constraint}}.
]
The nonnegative weights and each loss must be defined for the chosen implementation. The imagined-data term can be omitted. This expression organizes training components; it is not a convergence theorem or a guarantee that their objectives are compatible.
Regularization may help preserve selected prior behavior but can impede adaptation. Its contribution must be isolated experimentally, including the storage and computation needed to maintain its constraints.
3.3 Phase Three: Evaluation and Synchronization
Evaluate the candidate on retained-task tests, new-task tests, and scenarios reflecting the intended deployment. Tests should include rare cases and uncertainty where relevant. Acceptance tolerances must be declared rather than chosen after seeing the results.
Repeated checkpoint selection can overfit a fixed validation set. Separate development validation from final evaluation, document test reuse, and refresh evaluation appropriately. A successful promotion supports only the tested conditions.
Following promotion, monitor observed performance and compare it with the expected changes. Update the generator and stored representations under explicit compatibility rules rather than assuming all components remain synchronized automatically.
4. Failure Modes and Design Responses
Revised September 22, 2026
Status: Proposed architecture and research agenda. This article presents no new benchmark results. “Reflection” denotes computational replay, simulation, and evaluation; it does not imply conscious introspection.
IntroductionContinual learning asks how a system can acquire new capabilities while retaining useful prior knowledge under changing conditions. Sequential training can alter parameters that support earlier tasks, producing catastrophic forgetting. An architecture intended for long-term adaptation must therefore address both learning new information and preserving, correcting, or retiring earlier information appropriately.
Replay is already an established approach to this problem. Brain-inspired generative replay has been demonstrated on continual-learning benchmarks, including tasks with natural-image inputs. Sublayer AI builds on this research rather than claiming to introduce offline learning or generative rehearsal. [1]
The proposal combines an outer layer for interaction with an inner sublayer for scheduled consolidation and predictive simulation. Its potential contribution is the coordination of these processes: which experiences to retain, when to replay them, when imagined trajectories are useful, and how to determine whether an update should replace the deployed model.
The central hypothesis is that this coordinated cycle can improve retention and adaptation under defined memory and compute budgets. It remains to be tested against simpler replay methods and equally resourced baselines.
1. Neurocognitive Inspiration and Its Limits
1.1 Replay, Internal Activity, and Consolidation
Research linking memory replay with the default mode network motivates attention to internally generated activity during periods without immediate external demands. The cascaded memory systems account proposes connections between replay, the DMN, and memory organization; these connections should be distinguished from a fully established biological mechanism. Human experiments also investigate brain-wide activation associated with replay. [2, 3]
These findings motivate a design question: can a learning system use periods between interactions to revisit experience and assess candidate updates? They do not establish that a software replay module reproduces the DMN or that offline processing produces self-awareness.
1.2 Complementary Functions
The architecture adopts a functional distinction between retaining particular experiences and learning more general representations. A computational model by Spens and Burgess explores generative processes in memory construction and consolidation, providing a relevant example of how biological questions can inform machine-learning models. [4]
In Sublayer AI, a bounded buffer preserves selected observations, while a learned generator supports rehearsal. These are engineering roles inspired by memory research. Neither is a literal hippocampus or neocortex, and the analogy does not determine the correct implementation.
1.3 What the Analogy Does Not Establish
Human memory is not a guarantee of interference-free learning, and a biologically inspired design is not automatically effective. Replay, consolidation, internal simulation, and self-evaluation are distinct processes whose contributions should be measured separately.
The proposed controller is a scheduling mechanism. Describing its operations as “reflective” should not be taken as evidence of metacognition unless the system is explicitly evaluated for a defined capacity to monitor and regulate its own performance.
2. Architecture: Interaction, Consolidation, and Validation
Sublayer AI comprises two processing layers and an update-validation boundary. The outer layer supplies predictions or actions during interaction. The inner sublayer prepares candidate updates using retained and generated experience. A validation process determines whether a candidate is eligible for deployment.
The deployed model and the candidate model remain distinct. Learning during an interaction period may update a candidate without immediately changing the model serving users or controlling an environment. The permitted deployment cadence is a separate design choice.
2.1 Outer Layer: Task Performance and Experience Collection
The outer layer maps observations to predictions or actions. Each retained training record should distinguish observed inputs, actions, outcomes, available labels, and inferred or missing information.
The initial implementation should select a specific learning setting: supervised continual learning or reinforcement learning. Their loss functions, feedback, and evaluation requirements differ. A general architecture can accommodate both without assuming that one training procedure works unchanged across them.
2.2 Bounded Experience Buffer
A bounded buffer retains selected real observations or transitions. Its selection policy should be explicit: recent examples, representative coverage, rare events, or a defined mixture. Capacity must be counted in bytes as well as examples when comparing representations of different sizes.
Retention should reflect permissions and purpose. The buffer is not necessarily limited to recent experience; some older examples may be retained because they anchor evaluation or represent conditions that a generator reproduces poorly. Training records and held-out evaluation records must remain separate.
2.3 Generative Memory Module
A generative model produces rehearsal examples or internal representations associated with earlier experience. Candidate implementations include a conditional VAE or another suitable generator, selected for the task and resource budget.
The generator is itself a learned component and may forget. The design must specify how it is updated and evaluated. Possible mechanisms include retained real examples, frozen teacher checkpoints, or distillation against prior model behavior. A teacher can preserve earlier errors, so agreement with it is not sufficient evidence of correctness.
For supervised tasks, generated targets require a defined source: conditioning labels, teacher predictions, or another validated procedure. Synthetic labels must not silently become verified observations.
Training only on self-generated samples does not supply independent evidence that realism or coverage has improved. Evaluation should include real held-out observations and tests for loss of rare modes.
2.4 Latent Replay and Representation Compatibility
Replaying hidden representations can reduce the burden of synthesizing full inputs, as explored in brain-inspired replay research. [1] It also introduces a compatibility problem: a latent vector produced by one encoder may no longer have the same meaning after that encoder changes.
A Sublayer implementation should therefore specify whether it freezes the relevant encoder, regenerates representations from retained inputs, learns a tested translation between versions, or uses another compatibility mechanism. Record the encoder version associated with stored representations. Measure the consequences of representation drift instead of assuming latent replay remains valid indefinitely.
2.5 Goal-Conditioned Prediction and Imagination
Goal-Conditioned Predictive Coding (GCPC) learns trajectory representations that support goal-conditioned policy learning. Its published formulation should not be conflated with a complete simulator generating arbitrary future states, actions, and rewards. [5]
If Sublayer AI uses explicit imagined trajectories, it additionally requires an action-conditioned dynamics model, a policy or action-selection procedure, and definitions of rewards, termination, and rollout limits. Goals influence the selection of actions or trajectories; they do not ensure that a desired outcome is achievable.
Imagined transitions should remain labeled as model-generated. Their use should depend on measured predictive quality, the distance from observed experience, and uncertainty checks. Shorter rollouts, disagreement among models, or restrictions to supported regions are candidate controls, not guarantees against simulation error.
2.6 Controller and Context Gating
The controller determines when consolidation runs, allocates replay and simulation budgets, and chooses contexts to revisit. A first prototype can use a fixed schedule before testing adaptive scheduling based on drift indicators or retention checks.
Context gating may reduce interference, but it can also restrict useful transfer. The design must state whether context or task identity is supplied, inferred, or unavailable. Evaluation should include incorrect context assignment and conditions in which task boundaries are unclear.
An inferred context is a hypothesis. It should not become an unquestioned instruction that excludes relevant experience from rehearsal.
2.7 Update-Validation Boundary
The inner sublayer produces a candidate checkpoint. A separate evaluation process compares it with the deployed checkpoint using predefined criteria for retention, new-task performance, and domain-specific constraints.
A candidate that fails those criteria is rejected or revised. Promotion records should identify the training data versions, replay configuration, model versions, evaluation results, and authorized deployment decision. Recovery should preserve a known prior checkpoint while acknowledging that reverting model weights cannot undo actions already taken.
Suggested architecture-figure caption: The outer layer interacts with the environment and records selected experience. The inner sublayer uses bounded replay, generative rehearsal, and optional predictive simulation to train a candidate checkpoint. Held-out evaluation governs promotion to the deployed model; imagined experience is not treated as verified observation.
3. Training Protocol
3.1 Phase One: Online Interaction
The deployed model performs the authorized task. Selected observations enter a staging process that checks record quality, permissions, and available labels before admitting them to the training buffer.
The candidate learner may train on incoming observations together with replayed examples. The article does not assume that replay guarantees retention; the replay ratio, sampling policy, and their effects must be measured.
3.2 Phase Two: Offline Consolidation
During a scheduled consolidation period, the candidate learner combines retained observations with generated rehearsal examples. An optional predictive module supplies imagined transitions for tasks where model-based learning is appropriate.
“Offline” here means training without live interaction. It differs from the technical setting of offline reinforcement learning, in which a policy is learned from a previously collected dataset. A Sublayer implementation may use offline RL methods, but the two meanings should remain explicit.
A possible task-specific objective has the schematic form:
[
\mathcal L_{\mathrm{candidate}}=
\lambda_n\mathcal L_{\mathrm{new}}+
\lambda_r\mathcal L_{\mathrm{replay}}+
\lambda_i\mathcal L_{\mathrm{imagined}}+
\lambda_c\mathcal L_{\mathrm{constraint}}.
]
The nonnegative weights and each loss must be defined for the chosen implementation. The imagined-data term can be omitted. This expression organizes training components; it is not a convergence theorem or a guarantee that their objectives are compatible.
Regularization may help preserve selected prior behavior but can impede adaptation. Its contribution must be isolated experimentally, including the storage and computation needed to maintain its constraints.
3.3 Phase Three: Evaluation and Synchronization
Evaluate the candidate on retained-task tests, new-task tests, and scenarios reflecting the intended deployment. Tests should include rare cases and uncertainty where relevant. Acceptance tolerances must be declared rather than chosen after seeing the results.
Repeated checkpoint selection can overfit a fixed validation set. Separate development validation from final evaluation, document test reuse, and refresh evaluation appropriately. A successful promotion supports only the tested conditions.
Following promotion, monitor observed performance and compare it with the expected changes. Update the generator and stored representations under explicit compatibility rules rather than assuming all components remain synchronized automatically.
4. Failure Modes and Design Responses
These controls define experiments to perform. They are not claims that the corresponding failure modes have been eliminated.
5. Evaluation: What Would Demonstrate Value?
5.1 Baselines and Resource Accounting
Compare sequential fine-tuning, bounded-buffer replay, generative replay, relevant regularization methods, and the complete Sublayer design. Include a baseline given the same additional optimization budget as the offline phase.
Account for all model parameters, teacher checkpoints, replay storage, simulation steps, training compute, evaluation overhead, and inference latency. If exact budget matching is impractical, report performance across resource budgets rather than attributing every gain to architecture.
5.2 Retention, Adaptation, and Transfer
For task-based evaluation, let A_t,j denote a consistently defined higher-is-better performance score on task j after learning through stage t. Report the complete performance matrix where practical.
A simple average retained-performance measure after T tasks is:
[
\overline A_T=\frac{1}{T}\sum_{j=1}^{T}A_{T,j}.
]
One nonnegative forgetting summary for earlier tasks is:
[
F_T=\frac{1}{T-1}\sum_{j=1}^{T-1}
\max\left{0,\max_{k=j,\ldots,T}A_{k,j}-A_{T,j}\right},
\qquad T>1.
]
This convention measures decline from the best observed score, including the final evaluation, and should be named explicitly. It can be affected by noisy maxima. Report variability and individual tasks rather than relying on this number alone.
Evaluate new-task learning speed, backward transfer, and forward transfer against declared baselines. For task-free streams or changing distributions, use time-resolved evaluations appropriate to the setting instead of forcing artificial task labels.
5.3 Tests of the Proposed Contribution
5. Evaluation: What Would Demonstrate Value?
5.1 Baselines and Resource Accounting
Compare sequential fine-tuning, bounded-buffer replay, generative replay, relevant regularization methods, and the complete Sublayer design. Include a baseline given the same additional optimization budget as the offline phase.
Account for all model parameters, teacher checkpoints, replay storage, simulation steps, training compute, evaluation overhead, and inference latency. If exact budget matching is impractical, report performance across resource budgets rather than attributing every gain to architecture.
5.2 Retention, Adaptation, and Transfer
For task-based evaluation, let A_t,j denote a consistently defined higher-is-better performance score on task j after learning through stage t. Report the complete performance matrix where practical.
A simple average retained-performance measure after T tasks is:
[
\overline A_T=\frac{1}{T}\sum_{j=1}^{T}A_{T,j}.
]
One nonnegative forgetting summary for earlier tasks is:
[
F_T=\frac{1}{T-1}\sum_{j=1}^{T-1}
\max\left{0,\max_{k=j,\ldots,T}A_{k,j}-A_{T,j}\right},
\qquad T>1.
]
This convention measures decline from the best observed score, including the final evaluation, and should be named explicitly. It can be affected by noisy maxima. Report variability and individual tasks rather than relying on this number alone.
Evaluate new-task learning speed, backward transfer, and forward transfer against declared baselines. For task-free streams or changing distributions, use time-resolved evaluations appropriate to the setting instead of forcing artificial task labels.
5.3 Tests of the Proposed Contribution
Run multiple seeds and task orders, report uncertainty, and include negative results. Improved retention alone does not establish explainability, ethical alignment, or safe deployment.
6. Applications as Prospective Research Settings
6.1 Adaptive Robotics
A simulator or bounded laboratory task can test whether consolidation preserves earlier skills while incorporating new observations. Imagined rollouts may help explore uncommon conditions, but model errors must be checked against independent environment behavior. Results in one simulator do not establish readiness for autonomous vehicles or unrestricted robotics.
6.2 Clinical Research
Retrospective, appropriately governed datasets could support studies of distribution changes and retention across patient groups or imaging conditions. The architecture should not be described as a validated diagnostic system. Simulated treatment outcomes require causal assumptions and suitable evidence; generating plausible trajectories does not establish treatment effects.
Explanations also require separate evaluation. Inspecting latent activity does not automatically provide a faithful account of a recommendation.
6.3 Policy Modeling
A predictive model can support conditional scenario exploration when its assumptions and uncertainty are visible. Forecasts under alternative goals are not automatically causal estimates of interventions. Researchers must assess whether the underlying data and model support the proposed comparison.
6.4 Creative and Personal Assistance
Bounded studies could examine retention of authorized preferences, adaptation after explicit corrections, and consistency across projects. The ability to forget or retire obsolete preferences is as important as preserving them. Generative recombination may support creativity, but coherence and usefulness require task-specific evaluation.
7. Reflective Computation and Accountability
Offline simulation can produce candidate failure cases and help prioritize review. It cannot expose every flaw, especially when the simulator and deployed policy share mistaken assumptions.
The system should preserve a distinction among observed experience, generated rehearsal, hypothetical trajectories, and evaluation findings. Reports should identify which source supports a claimed improvement or detected risk.
Active-inference approaches offer a related research direction for explicit generative models and interpretable decision processes. The cited work proposes an explainability architecture; it does not establish that Sublayer AI already possesses transparent introspection. [6]
If the design claims metacognitive monitoring, it should specify what the system estimates about its own performance and test the accuracy and usefulness of those estimates. A controller that schedules replay should not be credited with knowing its limitations solely because it can initiate another training cycle.
8. Conclusion and Future Directions
Sublayer AI proposes a coordinated cycle of experience collection, replay, predictive simulation, and validated model updates. Its brain-inspired language motivates the design while leaving biological mechanisms and computational implementations distinct.
The immediate research priority is a bounded prototype with explicit record provenance, generator maintenance, representation compatibility, and candidate-checkpoint validation. Hierarchical replay, adaptive scheduling, and richer predictive models are subsequent experiments whose costs and benefits should be compared with simpler alternatives.
The proposal’s value will depend on whether it improves retention and adaptation under controlled resource budgets while keeping new failure modes visible. It does not claim to eliminate forgetting, reproduce the DMN, or create ethical intelligence through offline computation. Its ambition is to make learning between interactions a process that can be inspected, measured, and corrected.
Selected References
[1] van de Ven, G. M., Siegelmann, H. T., & Tolias, A. S. (2020). Brain-inspired replay for continual learning with artificial neural networks. Nature Communications, 11, 4069. https://doi.org/10.1038/s41467-020-17866-2.
[2] Kaefer, K., Stella, F., McNaughton, B. L., & Battaglia, F. P. (2022). Replay, the default mode network and the cascaded memory systems model. Nature Reviews Neuroscience, 23, 628–640. https://doi.org/10.1038/s41583-022-00620-6.
[3] Huang, Q., et al. (2024). Replay-triggered brain-wide activation in humans. Nature Communications, 15, 7185. https://doi.org/10.1038/s41467-024-51582-5.
[4] Spens, E., & Burgess, N. (2024). A generative model of memory construction and consolidation. Nature Human Behaviour, 8, 526–543. https://doi.org/10.1038/s41562-023-01799-z.
[5] Zeng, Z., Zhang, C., Wang, S., & Sun, C. (2023). Goal-Conditioned Predictive Coding for Offline Reinforcement Learning. Advances in Neural Information Processing Systems, 36. Primary publication.
[6] Albarracin, M., Hipólito, I., Tremblay, S. E., Fox, J. G., René, G., Friston, K., & Ramstead, M. J. D. (2023). Designing explainable artificial intelligence with active inference: A framework for transparent introspection and decision-making. arXiv:2306.04025. Manuscript. This reference cites the identifiable preprint version rather than the incorrect journal attribution in the original draft.
Evidence scope: These sources support component methods, biological research, or related proposals. None independently validates the complete Sublayer AI architecture presented here. The evaluation definitions and implementation requirements above specify the proposed research program.
6. Applications as Prospective Research Settings
6.1 Adaptive Robotics
A simulator or bounded laboratory task can test whether consolidation preserves earlier skills while incorporating new observations. Imagined rollouts may help explore uncommon conditions, but model errors must be checked against independent environment behavior. Results in one simulator do not establish readiness for autonomous vehicles or unrestricted robotics.
6.2 Clinical Research
Retrospective, appropriately governed datasets could support studies of distribution changes and retention across patient groups or imaging conditions. The architecture should not be described as a validated diagnostic system. Simulated treatment outcomes require causal assumptions and suitable evidence; generating plausible trajectories does not establish treatment effects.
Explanations also require separate evaluation. Inspecting latent activity does not automatically provide a faithful account of a recommendation.
6.3 Policy Modeling
A predictive model can support conditional scenario exploration when its assumptions and uncertainty are visible. Forecasts under alternative goals are not automatically causal estimates of interventions. Researchers must assess whether the underlying data and model support the proposed comparison.
6.4 Creative and Personal Assistance
Bounded studies could examine retention of authorized preferences, adaptation after explicit corrections, and consistency across projects. The ability to forget or retire obsolete preferences is as important as preserving them. Generative recombination may support creativity, but coherence and usefulness require task-specific evaluation.
7. Reflective Computation and Accountability
Offline simulation can produce candidate failure cases and help prioritize review. It cannot expose every flaw, especially when the simulator and deployed policy share mistaken assumptions.
The system should preserve a distinction among observed experience, generated rehearsal, hypothetical trajectories, and evaluation findings. Reports should identify which source supports a claimed improvement or detected risk.
Active-inference approaches offer a related research direction for explicit generative models and interpretable decision processes. The cited work proposes an explainability architecture; it does not establish that Sublayer AI already possesses transparent introspection. [6]
If the design claims metacognitive monitoring, it should specify what the system estimates about its own performance and test the accuracy and usefulness of those estimates. A controller that schedules replay should not be credited with knowing its limitations solely because it can initiate another training cycle.
8. Conclusion and Future Directions
Sublayer AI proposes a coordinated cycle of experience collection, replay, predictive simulation, and validated model updates. Its brain-inspired language motivates the design while leaving biological mechanisms and computational implementations distinct.
The immediate research priority is a bounded prototype with explicit record provenance, generator maintenance, representation compatibility, and candidate-checkpoint validation. Hierarchical replay, adaptive scheduling, and richer predictive models are subsequent experiments whose costs and benefits should be compared with simpler alternatives.
The proposal’s value will depend on whether it improves retention and adaptation under controlled resource budgets while keeping new failure modes visible. It does not claim to eliminate forgetting, reproduce the DMN, or create ethical intelligence through offline computation. Its ambition is to make learning between interactions a process that can be inspected, measured, and corrected.
Selected References
[1] van de Ven, G. M., Siegelmann, H. T., & Tolias, A. S. (2020). Brain-inspired replay for continual learning with artificial neural networks. Nature Communications, 11, 4069. https://doi.org/10.1038/s41467-020-17866-2.
[2] Kaefer, K., Stella, F., McNaughton, B. L., & Battaglia, F. P. (2022). Replay, the default mode network and the cascaded memory systems model. Nature Reviews Neuroscience, 23, 628–640. https://doi.org/10.1038/s41583-022-00620-6.
[3] Huang, Q., et al. (2024). Replay-triggered brain-wide activation in humans. Nature Communications, 15, 7185. https://doi.org/10.1038/s41467-024-51582-5.
[4] Spens, E., & Burgess, N. (2024). A generative model of memory construction and consolidation. Nature Human Behaviour, 8, 526–543. https://doi.org/10.1038/s41562-023-01799-z.
[5] Zeng, Z., Zhang, C., Wang, S., & Sun, C. (2023). Goal-Conditioned Predictive Coding for Offline Reinforcement Learning. Advances in Neural Information Processing Systems, 36. Primary publication.
[6] Albarracin, M., Hipólito, I., Tremblay, S. E., Fox, J. G., René, G., Friston, K., & Ramstead, M. J. D. (2023). Designing explainable artificial intelligence with active inference: A framework for transparent introspection and decision-making. arXiv:2306.04025. Manuscript. This reference cites the identifiable preprint version rather than the incorrect journal attribution in the original draft.
Evidence scope: These sources support component methods, biological research, or related proposals. None independently validates the complete Sublayer AI architecture presented here. The evaluation definitions and implementation requirements above specify the proposed research program.