arxiv: 2605.02509 · v1 · submitted 2026-05-04 · 💻 cs.LG · cs.NE

Recognition: 3 theorem links

· Lean Theorem

MPCS: Neuroplastic Continual Learning via Multi-Component Plasticity and Topology-Aware EWC

Joern Hentsch

Authors on Pith no claims yet

Pith reviewed 2026-05-08 19:05 UTC · model grok-4.3

classification 💻 cs.LG cs.NE

keywords continual learningneuroplasticityFourier encodingEWC regularizationPareto frontierMEP-BENCHstability-plasticity dilemmaneurogenesis

0 comments

The pith

MPCS integrates eleven mechanisms to reach a 94.2 normalized efficiency score on a 31-task continual learning benchmark, with Fourier encoding as the most critical component.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents MPCS, a neuroplastic continual learning architecture that combines task-driven neurogenesis, Fourier-encoded inputs, EWC regularization, meta-replay, mixed consolidation, hybrid gating, synapse pruning, Hebbian updates, task similarity routing, adaptive growth control, and neuron importance tracking. It tests this system on MEP-BENCH, a benchmark with 31 tasks spanning regression, classification, logic, and mixed domains, and measures success via a three-dimensional Pareto criterion of task performance, representation diversity, and gradient conflict rate. The evaluation across ablations shows MPCS lands on the Pareto frontier, Fourier encoding drives the largest gains, global EWC hurts results while its removal helps, and stripping EWC plus Hebbian creates a simpler variant that runs at lower cost. A reader would care because continual learning systems must acquire new skills without erasing old ones, and the work identifies which added mechanisms actually help versus add overhead in high-similarity task regimes.

Core claim

MPCS is a neuroplastic architecture that integrates eleven complementary mechanisms: task-driven neurogenesis, Fourier-encoded inputs, EWC regularization, meta-replay, mixed consolidation, hybrid gating, synapse pruning/regeneration, Hebbian updates, task similarity routing, adaptive growth control, and continuous neuron importance tracking. Evaluated on MEP-BENCH across 15 ablation configurations, it achieves a Normalized Efficiency Score of 94.2 and places on the Pareto frontier among 9 of 14 gate-passing systems. Ablations establish that Fourier encoding is the single most critical component, global EWC degrades performance while topology-local EWC reduces the penalty, and removing EWC in

What carries the argument

The MPCS architecture, which combines eleven mechanisms including neurogenesis and Fourier encoding to balance plasticity and stability during continual learning.

If this is right

Fourier encoding is the single most critical component; its removal drops performance by 30.7 percentage points and causes failure to pass the MEP gate on 14 percent of tasks.
In the high task-similarity regime, global EWC degrades results, topology-local EWC is better but still inferior to removing EWC entirely.
The Pareto frontier assessment acts as a model-compression guide, since jointly removing the two dominated components (EWC and Hebbian) produces MPCS_EFFICIENT with 0.6 pp higher performance at 4.7x lower compute.
MPCS reaches the Pareto frontier among 9 of 14 gate-passing systems under the three-dimensional criterion of performance, representation diversity, and gradient conflict rate.
MPCS_EFFICIENT runs in 127 minutes versus 602 minutes while improving task performance slightly.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The finding that EWC hurts performance under high task similarity could lead to simpler continual learning methods that skip regularization when tasks overlap substantially.
The multi-dimensional Pareto approach might be adopted in other continual learning work to evaluate efficiency beyond single metrics such as accuracy alone.
The topology-aware elements suggest that future systems could dynamically adjust network structure based on detected task similarity rather than applying uniform regularization.
This style of ablation-driven component selection could be tested in reinforcement learning settings where agents must adapt online without forgetting prior policies.

Load-bearing premise

The MEP-BENCH benchmark, its three-dimensional Pareto criterion, and the chosen high task-similarity regime are representative enough to generalize component-importance conclusions to other continual learning settings.

What would settle it

An experiment on a new benchmark with lower task similarity or different domains where MPCS falls off the Pareto frontier or Fourier encoding no longer produces the largest performance drop when removed.

read the original abstract

Continual learning systems face a fundamental tension between plasticity -- acquiring new knowledge -- and stability -- retaining prior knowledge. We introduce MPCS (Multi-Plasticity Continual System), a neuroplastic architecture that integrates eleven complementary mechanisms: task-driven neurogenesis, Fourier-encoded inputs, EWC regularization, meta-replay, mixed consolidation, hybrid gating, synapse pruning/regeneration, Hebbian updates, task similarity routing, adaptive growth control, and continuous neuron importance tracking. We evaluate MPCS on MEP-BENCH, a multi-track benchmark spanning 31 tasks across regression, classification, logic, and mixed domains, using a three-dimensional Pareto criterion over task performance (Perf), representation diversity (RD), and gradient conflict rate (GCR). Across 15 ablation configurations (3 seeds x 4 tracks x 2000 epochs), MPCS achieves a Normalized Efficiency Score of 94.2, placing it on the Pareto frontier among 9 of 14 gate-passing systems. Key findings: (i) Fourier encoding is the single most critical component (removal drops Perf by 30.7 pp and fails the MEP gate on 14% of tasks); (ii) global EWC degrades performance (NES = -4.2); topology-local EWC reduces this penalty (NES 90.5->91.8) but does not eliminate it; removing EWC entirely yields MPCS_EFFICIENT, the highest-Perf system -- establishing a monotone relationship in the high task-similarity regime (s_bar ~= 0.95): global EWC < topology EWC < no EWC; (iii) the Pareto status assessment is predictive: removing the two Pareto-dominated components (EWC + Hebbian) jointly yields MPCS_EFFICIENT, which improves Perf by 0.6 pp at 4.7x lower compute cost (127 vs. 602 min), validating the Pareto frontier as an actionable model-compression guide.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

MPCS bundles standard continual learning pieces with a new 31-task benchmark and finds Fourier features matter most while EWC can be dropped for speed when tasks are highly similar.

read the letter

The paper combines known continual learning ingredients—task-driven neurogenesis, Fourier-encoded inputs, EWC regularization, Hebbian updates, meta-replay, and several others—into a single system called MPCS. It then tests the whole thing on MEP-BENCH, a new multi-track benchmark with 31 tasks across regression, classification, logic, and mixed domains, using a three-dimensional Pareto criterion on performance, representation diversity, and gradient conflict rate. Across 15 ablation runs it reports that Fourier encoding is the most critical component and that removing EWC plus Hebbian learning produces a simpler model that runs 4.7 times faster with slightly higher performance. The suggestion that the Pareto frontier can serve as a practical guide for model compression is a concrete, usable takeaway from the ablations. The work is honest about running many controlled variants and tracking component rankings directly from those runs. The central limitation is that every headline result comes from a single high task-similarity regime (average similarity around 0.95). In that narrow setting, stability penalties like EWC become dispensable, so the observed ordering (global EWC worse than topology EWC worse than no EWC) may not survive when tasks diverge more. The abstract gives no error bars, no statistical tests, and no external baseline comparisons, which leaves the concrete numbers (NES 94.2, 30.7 pp drop, etc.) hard to interpret without the full methods and data. There is also some circularity: the efficiency score and the decision to drop components are both derived from the same ablation set. This is for continual learning researchers who want to see how a large collection of mechanisms interact inside one architecture and who value ablation-driven compression ideas. It has enough empirical substance and a new benchmark to deserve a serious referee, though the review will need to press for lower-similarity controls and proper statistical reporting.

Referee Report

3 major / 2 minor

Summary. The paper introduces MPCS, a continual learning architecture combining eleven mechanisms (task-driven neurogenesis, Fourier-encoded inputs, EWC regularization, meta-replay, mixed consolidation, hybrid gating, synapse pruning/regeneration, Hebbian updates, task similarity routing, adaptive growth control, and neuron importance tracking). It evaluates the system on the custom MEP-BENCH benchmark of 31 tasks across regression, classification, logic, and mixed domains using a three-dimensional Pareto criterion over task performance (Perf), representation diversity (RD), and gradient conflict rate (GCR). Across 15 ablation configurations (3 seeds, 4 tracks, 2000 epochs), MPCS reports a Normalized Efficiency Score of 94.2, placing it on the Pareto frontier among 9 of 14 gate-passing systems. Key claims include Fourier encoding as the most critical component (30.7 pp Perf drop on removal), a monotone relationship in the high task-similarity regime (s_bar ≈ 0.95) where global EWC < topology-aware EWC < no EWC, and that Pareto-guided removal of EWC plus Hebbian yields MPCS_EFFICIENT with 0.6 pp higher Perf at 4.7× lower compute.

Significance. If the empirical results hold under broader validation, the work offers concrete evidence on the relative value of plasticity mechanisms (Fourier encoding) versus stability mechanisms (EWC) in high task-similarity continual learning, together with a Pareto-frontier approach to component pruning that could guide efficient model design. The multi-component ablation study and explicit compute-accuracy trade-off quantification are strengths that could inform future neuroplastic architectures, provided the findings are shown to be robust beyond the specific MEP-BENCH protocol.

major comments (3)

[Abstract and §4] Abstract and §4 (Experiments): The headline numerical claims (NES = 94.2, 30.7 pp Perf drop on Fourier removal, 4.7× compute reduction, NES values for EWC variants) are reported without error bars, standard deviations across the 3 seeds, or any statistical significance tests, undermining confidence in the component rankings and Pareto-frontier status.
[§5.2] §5.2 (Ablation Studies): The monotone relationship 'global EWC < topology EWC < no EWC' and the identification of Fourier encoding as the single most critical component are extracted from the identical set of 15 ablation runs that define the Normalized Efficiency Score and the three-dimensional gate; this creates circularity because the same data both construct the metric and validate the ordering.
[§3.2 and §4.1] §3.2 (Task Similarity) and §4.1 (Benchmark): The central claim that EWC becomes dispensable rests on the high-similarity regime s_bar ≈ 0.95; no control experiments are reported for lower task-similarity regimes where catastrophic forgetting is stronger, so the component-importance conclusions and the recommendation to remove EWC lack evidence of robustness outside the chosen MEP-BENCH slice.

minor comments (2)

[§3] The formal definition of the Normalized Efficiency Score and the precise weighting of the three-dimensional gate (Perf, RD, GCR) should appear as equations in the main text rather than being deferred to the appendix.
[Figures 3-5] Figure captions and axis labels for the Pareto plots should explicitly state the number of seeds and whether shaded regions represent standard error.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below, indicating where revisions will be made to the manuscript.

read point-by-point responses

Referee: [Abstract and §4] Abstract and §4 (Experiments): The headline numerical claims (NES = 94.2, 30.7 pp Perf drop on Fourier removal, 4.7× compute reduction, NES values for EWC variants) are reported without error bars, standard deviations across the 3 seeds, or any statistical significance tests, undermining confidence in the component rankings and Pareto-frontier status.

Authors: We agree that reporting variability and statistical measures is necessary to support the numerical claims and component rankings. The experiments were conducted with 3 seeds, allowing computation of standard deviations. In the revised manuscript, we will update all reported metrics in the abstract, §4 tables, and figures to include mean ± standard deviation, add error bars to plots, and include paired t-test results for key comparisons (e.g., EWC variants and Fourier ablation), while noting the limited power due to small n. revision: yes
Referee: [§5.2] §5.2 (Ablation Studies): The monotone relationship 'global EWC < topology EWC < no EWC' and the identification of Fourier encoding as the single most critical component are extracted from the identical set of 15 ablation runs that define the Normalized Efficiency Score and the three-dimensional gate; this creates circularity because the same data both construct the metric and validate the ordering.

Authors: We acknowledge the valid concern about using the same ablation set for both metric computation and deriving orderings. The Pareto criteria (Perf, RD, GCR) and NES normalization are defined independently prior to running ablations. The observed relationships are empirical outcomes from applying the metric. We will revise §5.2 to explicitly separate the a priori metric definition from the post-hoc component analysis and clarify that the monotone relationship is an observation within this specific benchmark rather than a general validation. revision: partial
Referee: [§3.2 and §4.1] §3.2 (Task Similarity) and §4.1 (Benchmark): The central claim that EWC becomes dispensable rests on the high-similarity regime s_bar ≈ 0.95; no control experiments are reported for lower task-similarity regimes where catastrophic forgetting is stronger, so the component-importance conclusions and the recommendation to remove EWC lack evidence of robustness outside the chosen MEP-BENCH slice.

Authors: We agree that the findings on EWC dispensability and the monotone relationship are tied to the high task-similarity regime (s_bar ≈ 0.95) of MEP-BENCH, with no controls for lower-similarity settings where forgetting effects are stronger. This limits the generalizability of the component recommendations. We will add explicit discussion in §3.2, §5, and the conclusion to scope the claims to high-similarity continual learning and identify lower-similarity robustness as an important direction for future work. revision: partial

standing simulated objections not resolved

Robustness of EWC-related conclusions and component pruning recommendations to lower task-similarity regimes, as no control experiments were performed outside the MEP-BENCH high-similarity slice.

Circularity Check

2 steps flagged

Ablation-derived NES, component rankings, and Pareto 'predictions' all reduce to the same 15 MEP-BENCH runs

specific steps

fitted input called prediction [Abstract, key findings (i)-(iii)]
"MPCS achieves a Normalized Efficiency Score of 94.2, placing it on the Pareto frontier among 9 of 14 gate-passing systems. Key findings: (i) Fourier encoding is the single most critical component (removal drops Perf by 30.7 pp and fails the MEP gate on 14% of tasks); (ii) global EWC degrades performance (NES = -4.2); topology-local EWC reduces this penalty (NES 90.5->91.8) but does not eliminate it; removing EWC entirely yields MPCS_EFFICIENT, the highest-Perf system -- establishing a monotone relationship in the high task-similarity regime (s_bar ~= 0.95): global EWC < topology EWC < no EWC; "

NES, Pareto frontier membership, and the monotone EWC ordering are all computed directly from the 15 ablation configurations. The 'finding' that Fourier is critical or that EWC is removable is the numerical outcome of those runs, not a prediction tested on held-out data or external benchmarks.
fitted input called prediction [Abstract, key finding (iii)]
"(iii) the Pareto status assessment is predictive: removing the two Pareto-dominated components (EWC + Hebbian) jointly yields MPCS_EFFICIENT, which improves Perf by 0.6 pp at 4.7x lower compute cost (127 vs. 602 min), validating the Pareto frontier as an actionable model-compression guide."

The claim that 'Pareto status is predictive' and that removal yields improvement is verified on the identical ablation data used to assign the original Pareto ranks and NES values. The validation loop is internal to the same experimental set.

full rationale

The paper defines Normalized Efficiency Score (NES) and the three-dimensional Pareto gate (Perf, RD, GCR) from its own ablation suite on MEP-BENCH (31 tasks, s_bar≈0.95, 2000 epochs). It then reports component importance (Fourier critical, EWC dispensable) and claims the Pareto frontier is 'predictive' because removing the low-NES components improves the same metrics. These conclusions are direct outputs of the defining experiments rather than independent tests, satisfying the fitted-input-called-prediction pattern. No cross-benchmark or lower-similarity controls are provided to break the loop. Score 7 reflects one central load-bearing reduction without full self-definition or self-citation chains.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The abstract contains no mathematical derivations, stated axioms, or new postulated entities; the work is an empirical architecture proposal whose claims rest on standard supervised-learning assumptions and the unstated validity of the chosen benchmark and metrics.

pith-pipeline@v0.9.0 · 5667 in / 1459 out tokens · 123270 ms · 2026-05-08T19:05:07.353358+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

Foundation/AlphaCoordinateFixation.lean (cosh/sinh log-coordinate machinery) costAlphaLog_fourth_deriv_at_zero — RS Fourier/cosh structure is functional-equation-forced, not random-feature kernel approximation unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Raw inputs x are mapped to a higher-dimensional space via random Fourier features: ϕ(x) = [sin(W_f x), cos(W_f x)] (Rahimi & Recht 2007).
Foundation/BranchSelection.lean branch_selection — RS Pareto/branch-selection arguments operate on functional-equation combiners, not empirical hypervolume scores unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Three objectives are evaluated... Pareto membership is determined in the (Perf, −RD, −GCR) space. NES normalizes each system's Pareto volume contribution.

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

18 extracted references · 6 canonical work pages · 1 internal anchor

[1]

Academic Press,

Michael McCloskey and Neal J Cohen.Catastrophic Interference in Connectionist Net- works: The Sequential Learning Problem, volume 24, pages 109–165. Academic Press,
[2]

URLhttps: //www.sciencedirect.com/science/article/pii/S0079742108605368

ISBN 0079-7421. doi: https://doi.org/10.1016/S0079-7421(08)60536-8. URL https://www.sciencedirect.com/science/article/pii/S0079742108605368

work page doi:10.1016/s0079-7421(08)60536-8
[3]

Catastrophic forgetting in connectionist networks.Trends in cognitive sciences, 3:128–135, 1999

Robert M French. Catastrophic forgetting in connectionist networks.Trends in cognitive sciences, 3:128–135, 1999

1999
[4]

Overcoming catastrophic forgetting in neural networks.Proceedings of the National Academy of Sciences, 114(13): 3521–3526, 2017

James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, An- drei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks.Proceedings of the National Academy of Sciences of the Unite...

work page doi:10.1073/pnas.1611835114 2017
[5]

Memory aware synapses: Learning what (not) to forget

Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuyte- laars. Memory aware synapses: Learning what (not) to forget. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors,Computer Vision – ECCV 2018, pages 144–161. Springer International Publishing, 2018. ISBN 978-3-030-01219-9

2018
[6]

Experi- ence replay for continual learning

David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experi- ence replay for continual learning. In H Wallach, H Larochelle, A Beygelzimer, F d Alché-Buc, E Fox, and R Garnett, editors,Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. URL https://proceedings.neurips.cc/paper_files/ p...

2019
[7]

Learning to learn without forgetting by maximizing transfer and minimizing interference

Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, , and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. InInternational Conference on Learning Representations, 2019. URL https: //openreview.net/forum?id=B1gTShAct7

2019
[8]

Progressive Neural Networks

Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks, 2022. URLhttps://arxiv.org/abs/1606.04671

work page internal anchor Pith review arXiv 2022
[9]

Lifelong learning with dynamically expandable networks

Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. InInternational Conference on Learning Representations. International Conference on Learning Representations, ICLR, 2018

2018
[10]

Progress & compress: A scalable framework for continual learning

Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress & compress: A scalable framework for continual learning. In Jennifer Dy and Andreas Krause, editors,Proceedings of the 35th International Conference on Machine Learning, volume 80, pages 4528–4537. PMLR, 4 2018. ...

2018
[11]

Measuring catastrophic forgetting in neural networks.Proceedings of the AAAI Conference on 7 Artificial Intelligence, 32, 4 2018

Ronald Kemker, Marc McClure, Angelina Abitino, Tyler Hayes, and Christopher Kanan. Measuring catastrophic forgetting in neural networks.Proceedings of the AAAI Conference on 7 Artificial Intelligence, 32, 4 2018. doi: 10.1609/aaai.v32i1.11651. URL https://ojs.aaai. org/index.php/AAAI/article/view/11651

work page doi:10.1609/aaai.v32i1.11651 2018
[12]

An Empirical Investigation of Catastrophic Forgetting in Gradient-Based Neural Networks

Ian J Goodfellow, Mehdi Mirza, Da Xiao, Aaron Courville, and Yoshua Bengio. An empirical investigation of catastrophic forgetting in gradient-based neural networks, 2015. URL https: //arxiv.org/abs/1312.6211

work page Pith review arXiv 2015
[13]

Continual learning through synaptic intelligence

Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In Doina Precup and Yee Whye Teh, editors,Proceedings of the 34th International Conference on Machine Learning, volume 70, pages 3987–3995. PMLR, 4 2017. URL https: //proceedings.mlr.press/v70/zenke17a.html

2017
[14]

Continual learning with deep generative replay

Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. In I Guyon, U V on Luxburg, S Bengio, H Wallach, R Fergus, S Vishwanathan, and R Garnett, editors,Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. URL https://proceedings.neurips.cc/paper_files/paper/ 2017/file...

2017
[15]

Hebb.The Organization of Behavior A NEUROPSYCHOLOGICAL THEORY

Donald O. Hebb.The Organization of Behavior A NEUROPSYCHOLOGICAL THEORY. John Wiley & Sons, 1949

1949
[16]

Journal of Mathematical Biology15(3), 267–273 (1982)

Erkki Oja. Simplified neuron model as a principal component analyzer.Journal of Mathematical Biology, 15:267–273, 1982. ISSN 1432-1416. doi: 10.1007/BF00275687. URL https: //doi.org/10.1007/BF00275687

work page doi:10.1007/bf00275687 1982
[17]

Random features for large-scale kernel machines

Ali Rahimi and Benjamin Recht. Random features for large-scale kernel machines. In J Platt, D Koller, Y Singer, and S Roweis, editors,Advances in Neural Information Processing Sys- tems, volume 20. Curran Associates, Inc., 2007. URL https://proceedings.neurips.cc/ paper_files/paper/2007/file/013a006f03dbc5392effeb8f18fda755-Paper.pdf

2007
[18]

Avalanche: An end-to-end library for continual learning

Vincenzo Lomonaco, Lorenzo Pellegrini, Andrea Cossu, Antonio Carta, Gabriele Graffieti, Tyler L Hayes, Matthias De Lange, Marc Masana, Jary Pomponi, Gido M van de Ven, Martin Mundt, Qi She, Keiland Cooper, Jeremy Forest, Eden Belouadah, Simone Calderara, German I Parisi, Fabio Cuzzolin, Andreas S Tolias, Simone Scardapane, Luca Antiga, Subutai Ahmad, Adri...

2021