similarity, Optdigits.

Artificial learning systems are increasingly expected to acquire knowledge sequentially: new users appear, sensor calibrations drift, and novel categories must be learned while earlier ones remain in service. Networks trained with gradient descent handle this demand poorly. When the training distribution shifts from one task to the next, the updates that fit the new task partially overwrite the parameters that encoded the previous one, so that competence acquired earlier degrades rapidly. This failure mode, known as catastrophic interference, was demonstrated convincingly in the early connectionist literature: McCloskey and Cohen [1] showed that a network trained by back-propagation of errors [2] on one paired-associate list and then a second lost almost all retention of the first, and they argued that the problem was intrinsic to the way distributed representations are updated rather than an artifact of a particular architecture.

The most influential theoretical account of how biological brains sidestep this problem is the Complementary Learning Systems (CLS) framework of McClelland, McNaughton, and O'Reilly [3]. CLS theory posits two interacting memories: a fast, sparse hippocampal store that binds the specifics of new episodes, and a slow neocortical system that acquires the statistical structure of the environment through gradual, interleaved re-exposure to old and new material. The computational translation of this account is rehearsal: retain some record of past experience and re-present it, interleaved with new data, while learning continues. Robins [4] demonstrated that rehearsal with a small store of true exemplars — and, more remarkably, pseudorehearsal with internally generated pseudo-exemplars — substantially reduces catastrophic forgetting in small networks. Rehearsal-style buffers have since become a standard tool, most visibly in reinforcement learning, where Experience Replay (ER) stabilizes learning from non-stationary streams [5].

Contemporary continual learning research has broadened this toolkit considerably. Regularization approaches such as elastic weight consolidation identify parameters important to old tasks and penalize their modification [6]; architectural approaches allocate or grow task-specific capacity; and replay-based approaches store, condense, or generate past data and mix it into current training. Recent surveys [7], [8] and overview chapters [9] catalogue dozens of variants and report that replay methods with even modest memory remain among the most robust baselines across benchmarks. At the same time, the empirical picture is fragmented: reported comparisons typically change several design factors at once — the amount of stored data, the rule by which stored samples are drawn into training, and the relatedness of the tasks themselves — so that the individual contribution of each factor is difficult to isolate.

The study uses the 1797-sample digit subset bundled by scikit-learn [10], [11], partitioned into five tasks discriminating digit pairs 2i and 2i+1. ER draws uniformly from a class-balanced buffer including the current task; IR uses the same buffer but samples only past tasks. Thus the compared distributions also change past-task rehearsal dose and current-task duplication. Joint training provides a future-data reference, not a mathematical upper bound. Per-task input permutations manipulate layout statistics, and ten seeded replications support conditional exploratory comparisons.

The design varies nominal buffer capacity, replay sampling distribution and input layout. Shared splits, initialization and buffer snapshots are reported controls, subject to code confirmation. The evidence concerns aggregate dose response and exploratory layout effects; it neither proves a pure scheduling effect nor establishes a biological memory mechanism.

Catastrophic interference was characterized long before it became a benchmarking concern. McCloskey and Cohen [1] trained back-propagation networks [2] on paired-associate lists and observed that learning a second list drove retention of the first toward zero, in sharp contrast to human subjects, who showed moderate proactive and retroactive interference but no wholesale erasure. Their analysis located the problem in the overlap of internal representations: because the same hidden units participate in encoding both lists, gradients from the new task move weights that the old task depends on. Interference is therefore expected to scale with representational overlap between sequentially learned tasks — an expectation that directly motivates our third hypothesis.

The CLS framework [3] motivates retaining episodes and re-exposing a shared learner to them. Robins [4] studied literal and pseudo-rehearsal. The present ER-IR comparison varies current-task inclusion and past-task sample dose along with task balancing, so it tests a combined replay distribution rather than pure temporal scheduling.

Regularization methods such as elastic weight consolidation protect important parameters [6]; surveys review replay and other continual-learning families [7], [8]. A recent chapter places these methods within learning and memory [9]. Reinforcement-learning replay serves additional purposes such as decorrelation [5]. This experiment investigates exemplar memory and task sampling in a supervised task-incremental network, rather than testing all rehearsal mechanisms.

Continual learning involves sequential exposure to changing task data. Active learning concerns selective acquisition of labels and need not assume a stationary pool [12]; classical pooled prediction [13] addresses another evaluation setting. These distinctions motivate specifying the task sequence and memory-access assumptions directly.

This is a controlled empirical comparison rather than a new algorithm. Capacity, current-task inclusion, past-task sampling and layout statistics are explicit factors. Shared buffer contents improve the comparison, but differing replay doses and future-data access must be accounted for before attributing effects solely to scheduling. The present review does not establish that this particular factorization has never been studied.

The study asks how nominal capacity relates to retention, how two replay distributions differ, and how per-task layout permutations affect interference. These are operational questions in a small network; the design does not equate layout correlation with semantic task similarity or measured hidden-representation overlap. RQ1 (capacity). How does replay buffer capacity, from 25 to 200 stored samples per class, govern Backward Transfer (BWT) and final average accuracy on sequential digit tasks?

H1. forgetting decreases monotonically — BWT rises toward zero — as capacity grows.

RQ2 (scheduling). Does distributing a fixed replay budget uniformly over past tasks (Interleaved Replay (IR)) reduce forgetting relative to drawing the same number of samples uniformly from the entire buffer (ER)?

H2. IR achieves higher BWT than same-budget ER, with the benefit expected at moderate buffer sizes (50– 100 samples per class).

RQ3 (similarity). Does the degree of input-layout overlap between tasks modulate interference?

H3. streams whose tasks share a common input layout (natural digit images) suffer more interference than streams whose tasks are decorrelated by per-task feature permutations, consistent with overlap-based accounts of catastrophic interference [1], [3].

The scope is one dataset subset and a compact task-incremental network. Task identity is supplied at test time. Shared controls and paired seeds support conditional comparisons, while the full pipeline and run-level statistics require artifact confirmation.

The source reports 200 runs: ten seeds 42-51, two layouts and ten conditions per layout, with 186 s runtime. NumPy [13], scikit-learn [11], SciPy [14] and Matplotlib [15] are named components. The original script and results file are needed to verify execution, controls, timings, and programmatic insertion of values.

Reported checks include first-task loss changing from 0.551 to 0.00016 and mean untrained accuracy 0.49 over non-joint runs. These are plausible sanity summaries, but pooled mean accuracy does not prove unbiased heads or absence of forward transfer on each task. Run-level loss and accuracy matrices are required to confirm the checks.

The source reports a balanced two-layout by ten-condition grid per seed, yielding 200 runs. Naive, ER/IR at four caps, and joint have ten replications per cell. Shared splits, initialization and snapshots support pairing among sequential strategies; original artifacts must confirm completeness, exclusions and exact controls. Paired two-sided Wilcoxon signed-rank tests summarize differences over ten seeds; rank-biserial effects and percentile bootstrap intervals are reported. The source uses 2000 bootstrap draws with generator 1234. The twenty-six comparisons are exploratory and unadjusted in the original tables. A Holm sensitivity calculation on the displayed rounded p-values gives minimum adjusted p = 0.052 across the complete 26-test family, so none reaches 0.05. Exact nonzero counts, ties and rounding require the raw results before assigning exact adjusted probabilities.

Tables 3 and 4 retain the twenty-six reported unadjusted paired tests: sixteen ER-IR comparisons and ten layout comparisons. One ER-IR accuracy contrast has p = 0.036, whereas all BWT contrasts have p ≥ 0.058. Holm correction on displayed values across 26 tests gives no adjusted p below 0.05 (minimum 0.052); separating the exploratory questions into 16- and 10-test families leaves no ER-IR result and only three layout contrasts below 0.05. Exact adjustment should use unrounded raw probabilities and a justified family definition.

Unadjusted layout p-values range from 0.002 to 0.020 for the seven favourable contrasts, with large reported rank-biserial magnitudes. Separate marginal bootstrap intervals are not a paired confidence interval for the contrast, and nonoverlap cannot replace a multiplicity-aware paired test. Conversely, overlapping intervals do not prove equivalence or absence of a practically relevant difference.

Ten nonzero, untied pairs permit a minimum exact two-sided signed-rank probability 2/2^10=0.001953125. Zeros or ties alter the effective sample and implementation; the raw comparisons are needed to reconcile p-values and effect sizes. No formal equivalence margin, TOST, or preregistered testing family is supplied. Small observed mean differences do not establish a bound on the true paired effect.

Memory capacity is associated with most of the measured accuracy recovery in natural-layout ER, with diminishing increments at larger nominal caps. Capacity 100 retains roughly 70 percent of the training examples; capacity 200 retains the complete available class data. These are sizeable memory fractions on a small benchmark, so they should not be described as an arbitrary 200-sample reservoir comparison or a deployment provisioning rule.

The replay distributions differ in both current-task duplication and past-task dose. In a filled class-balanced buffer after i prior tasks, ER includes the current task with probability 1/(i+1); that fraction declines as tasks accumulate. IR samples only prior tasks and assigns them equal task probabilities. At early stages and while reservoirs fill, these differences can be larger. The absence of a detectable ER-IR benefit does not test pure temporal ordering at an identical past-task dose or prove a specific complementary-learning-systems mechanism. A matched past-dose and current-duplication experiment would isolate those questions.

Input permutations reduce task-average image correlation and are associated with improved retention in some exploratory conditions. Gradient conflict is a possible explanation, but no gradient dot products, hidden-state overlaps, or parameter-importance trajectories are measured. Invertible task encodings preserve information but can change optimization and representation geometry; both may contribute. The evidence therefore supports a benchmark-specific association with layout rather than a verified interference mechanism.

The layout contrast is not significant for naive training, where seed variation is largest. A formal interaction test is needed before concluding that layout effects require partial memory. Separate significances across capacities do not by themselves establish an interaction.

This task-incremental study compares exemplar replay distributions only. It does not benchmark against regularization methods [6], generative replay, larger networks or class-incremental inference. Shared reported controls improve comparison, but implementation verification and the replay-dose distinction remain necessary for strong internal-validity claims.

The experiment separates nominal memory capacity, a combined replay-distribution treatment, and coordinate layout. It finds a marked capacity association and no multiplicity-supported ER-IR advantage. Joint future-data access has a different information regime, so its performance cannot identify biological interleaving as the causal mechanism. Layout is a useful controlled benchmark variable, while generalization and mechanistic attribution require new measurements.

On this small digit subset, a nominal cap 100 retains about 70 percent of the training stream and yields ER accuracy 0.984 versus joint-reference 0.996; retaining all class training examples at cap 200 yields 0.988. These fractions and task-identity assumptions limit direct use for edge-memory provisioning. Per-task remapping may affect optimization and interface requirements, so a deployment recommendation cannot follow from this single layout experiment.

The learner is small and uses task-specific heads with known task identity. The dataset is the 1797-example scikit-learn subset rather than the full UCI corpus. Coordinate permutations modify layout statistics without changing task labels, but do not remove every shared representation. Nominal cap 200 exceeds available class training examples. The ER-IR contrast changes past-task sample dose and current-task inclusion, and the first-task fallback needs confirmation.

The seed-level tests are exploratory, contain ties and zeros, and lack a preregistered family, paired-effect interval, power analysis or equivalence margin. Multiplicity sensitivity limits strong significance claims. The source reports 1000 updates per task, around 111 current-sample exposures in epoch units, with additional replay work; normalization and actual counts require code. Hyperparameters are shared, and no regularization or generative baselines are evaluated.

The source specifies code/experiment.py, results/results.json and code/build_docx.py as the generation workflow, with seeds 42-51 and bootstrap seed 1234. It reports Python 3.12.14, NumPy 2.1.3, SciPy 1.14.1, scikit-learn 1.5.2 and Matplotlib 3.9.2, and 186 s runtime. A stable repository or supplied code/result package is required to verify these commands, environments and generation claims; it is not identified here.

The public dataset is the bundled scikit-learn digit subset [10], [11]. Exact splits, permutations, reservoirs, fallback batches and the claimed 200 accuracy matrices require the seeded original script and machine-readable output. No new human or animal participants are described, but applicable dataset-use conditions and declarations should be confirmed by the authors.

The reported task-incremental experiment compares exemplar replay on a small real digit subset. Natural-layout ER has improving mean retention with larger nominal capacity, from BWT -0.170 without replay to - 0.010 when the available training examples are retained, while final accuracy rises from 0.861 to 0.988 against a joint future-data reference of 0.996. These means are descriptive of the stated protocol rather than a monotonicity theorem.

No reported BWT contrast detects an ER-IR advantage at unadjusted 0.05, and none survives whole-family multiplicity control. Because past-task dose and current-task inclusion differ, this is a comparison of sampling distributions rather than a pure interleaving test. Layout permutations have favourable exploratory contrasts, but no whole-family Holm significance or measured gradient mechanism is established. The results provide a bounded empirical baseline whose reproduction and broader relevance require code, raw matrices and larger, matched-treatment studies.

The study uses the 1797-example load_digits subset of UCI Optical Recognition of Handwritten Digits. Splits and permutations are experimental transformations; their exact reconstruction requires the original seed-controlled code.

The study describes no newly recruited human or animal participants. Dataset permissions and any applicable review requirements must be confirmed by the authors.

This work received no external funding.

none declared.

Actual contributions and author approval must be confirmed by the named researchers. A program or AI tool cannot establish the author roles or scientific accountability.

The source reports AI-assisted drafting. The authors must confirm the tools and tasks actually used, the origin of experiment code and data, and responsibility for independent scientific verification.