---
title: Catastrophic Memorizing in Continual Learning
url: https://www.emergentmind.com/topics/catastrophic-memorizing
type: topic
---

# Catastrophic Memorizing in Continual Learning

Searching arXiv for the cited work and closely related usage of the term.
I’ll look up recent arXiv entries on “catastrophic remembering”, “catastrophic memorizing”, and related continual learning formulations.
Catastrophic memorizing denotes a failure mode in sequential or continual learning in which a model preserves or accumulates prior structure in a way that degrades inter-task discriminability rather than merely preserving competence. In continual learning, the closest explicit formulation is catastrophic remembering: the worsening ability to discriminate between data from different tasks, so that the network overgeneralizes across tasks, exhibits a sense of false familiarity with new inputs, and confuses task identity [2102.11343]. In language modeling, a related phenomenon appears as memory collapse: predictions collapse to a small set of answers when the model is trained on low-diversity knowledge, temporarily overriding unrelated knowledge [2305.09144]. This usage is distinct from catastrophic forgetting, which is the abrupt degradation of performance on previously learned tasks after sequential updates, and it is also distinct from within-task overfitting; the pseudo-recursal literature explicitly states that “Catastrophic memorizing” is not discussed there and that pseudo-recursal targets forgetting, not overfitting [1802.03875].

## 1. Definitions and conceptual boundaries

Catastrophic forgetting and catastrophic memorizing describe different pathologies of sequential learning. Catastrophic forgetting is the loss of performance on previously mastered tasks after updating on new tasks. Catastrophic memorizing, in the continual-learning formulation, is the loss of task discrimination caused by excessive representational overlap or replay-heavy training that encourages a global superset function across tasks. The result is not simply low retention or low accuracy, but confusion about whether an input belongs to an old task, a new task, or no known task at all [2102.11343].

The language-model literature uses a related but not identical operationalization. There, memory is tracked as the ability to answer factual queries correctly over training checkpoints, and undesirable memorization appears when the model’s predictions become dominated by a narrow answer set. The paper on language-model memory dynamics does not explicitly define catastrophic memorizing as a formal term, but it identifies memory collapse as an acute form of undesirable memorization in which low-diversity knowledge temporarily overrides unrelated knowledge [2305.09144].

A useful conceptual boundary separates catastrophic memorizing from both forgetting and standard memorization. Forgetting concerns the loss of earlier task performance under later updates. Standard memorization or overfitting concerns fitting idiosyncrasies of training data within a task and failing to generalize. Catastrophic memorizing instead concerns pathological retention or sharing: old structure dominates new discrimination, or low-entropy knowledge dominates unrelated predictions. This suggests that good continual learning requires not only stability against forgetting but also preservation of inter-task boundaries.

## 2. Bayesian and representational formulations

A Bayesian account of catastrophic remembering is given through the posterior after sequentially learning $n$ tasks:

$$
\log \mathcal{P}(\theta_{1:n}\mid D_{1:n}) =
\log \mathcal{P}(D_n\mid \theta_n)
+ \sum_{i=1}^{n-1} \log \mathcal{P}(D_i\mid \theta_i)
+ \log \mathcal{P}(\theta_1)
- \mathcal{C}.
$$

In this formulation, catastrophic remembering arises when the current-task likelihood is dominated by the cumulative prior-like contribution of previous tasks:

$$
\log \mathcal{P}(D_n\mid \theta_n) \ll \sum_{i=1}^{n-1} \log \mathcal{P}(D_i\mid \theta_i).
$$

The representational interpretation is that excessive overlap or misallocation of parameter relevance causes task confusion: features useful for earlier tasks overwhelm or entangle those needed for the current task, making task identity and boundaries indistinct [2102.11343].

This perspective motivates the Optimal Overlap Hypothesis. Zero overlap is inefficient because it wastes capacity and blocks transfer between related tasks. Unconstrained overlap is also suboptimal because it yields overgeneralization and catastrophic remembering. The proposed optimum is task-specific, controlled overlap: reduce overlap for unrelated tasks and increase it for similar tasks. Relevance Mapping Networks implement this by learning task-specific relevance mappings $\mathbf{r}^{(k)} \in [0,1]^d$ that gate shared parameters $\boldsymbol{\theta}$ through the Hadamard product $\boldsymbol{\theta}\odot \mathbf{r}^{(k)}$ [2102.11343].

The mask parameterization is

$$
\mathbf{r}^{(k)} = \sigma\big(\beta(\mathbf{m}^{(k)} - \tfrac{1}{2})\big),
$$

with $\mathbf{m}^{(k)}$ drawn from a Logit-normal mixture and $\beta>0$ acting as a learnable tightening parameter. As $\beta \to \infty$, the relevance map approaches binary rounding. The learning objective combines task loss, overlap regularization, and sparsity:

$$
\min_{\boldsymbol{\theta},\{\mathbf{r}^{(k)}\}_{k=1}^K}
\sum_{k=1}^K \mathbb{E}_{(x,y)\sim \mathcal{D}_k}
\Big[\mathcal{L}_{\text{task}}(f(x;\boldsymbol{\theta}\odot \mathbf{r}^{(k)}),y)\Big]
+ \lambda \sum_{1\le i<j\le K} R(\mathbf{r}^{(i)},\mathbf{r}^{(j)})
+ \gamma \sum_{k=1}^K \|\mathbf{r}^{(k)}\|_1.
$$

An overlap score consistent with the regularizer is

$$
O(\mathbf{r}^{(i)},\mathbf{r}^{(j)}) =
\frac{\langle \mathbf{r}^{(i)},\mathbf{r}^{(j)}\rangle}
{\|\mathbf{r}^{(i)}\|_2\,\|\mathbf{r}^{(j)}\|_2},
$$

or, in unnormalized form, $\langle \mathbf{r}^{(i)},\mathbf{r}^{(j)}\rangle$. Increasing $\lambda$ reduces overlap and favors separation; reducing $\lambda$ permits more sharing. In this framework, catastrophic memorizing is the regime in which overlap becomes too high for the task geometry.

## 3. Diagnostics and empirical signatures

Catastrophic memorizing is diagnosed through cross-task confusion rather than solely through end-task accuracy. In the RMN framework, two unsupervised diagnostics are central. The first is New Task/Data Detection, based on a relevance-modified statistical test on the KL divergence between prior and posterior distributions for an incoming mini-batch $x$:

$$
S(x)=D_{\text{KL}}\big(p(\boldsymbol{\theta}\mid \text{prior})\;\|\;p(\boldsymbol{\theta}\mid \text{posterior on }x)\big).
$$

A relevance-modified Welch’s t-test on batch-wise scores determines whether a task switch occurred. High detection accuracy indicates that the model can distinguish old from new data distributions and therefore exhibits low catastrophic remembering [2102.11343].

The second diagnostic is Unsupervised Task Inference. When task identity is unknown at test time, the model evaluates all masked pathways and returns

$$
\hat{y}=\arg\max_{k\in\{1,\dots,K\}} f\big(x;\boldsymbol{\theta}\odot \mathbf{r}^{(k)}\big).
$$

The percentage of inputs routed to the correct task measures inter-task discriminability. Confusion rate, defined as the fraction of test inputs misassigned to the wrong task during unsupervised task inference, provides a direct empirical signature of catastrophic memorizing [2102.11343].

Standard continual-learning metrics remain informative but are not sufficient on their own. Backward Transfer is defined as

$$
\text{BWT}=\frac{1}{K-1}\sum_{k=1}^{K-1}\big(A_{K,k}-A_{k,k}\big),
$$

where $A_{i,j}$ is the accuracy on task $j$ after learning task $i$. Large negative BWT indicates catastrophic forgetting; sustained negative transfer from later tasks may also correlate with catastrophic memorizing if representations collapse. Overlap scores $O(\mathbf{r}^{(i)},\mathbf{r}^{(j)})$ diagnose representational entanglement, and optional inter-task separability measures such as Fisher’s discriminant ratio, Maximum Mean Discrepancy, or inter-task KL divergences quantify how well task representations remain separated [2102.11343].

In language models, the analogous empirical object is the forgetting curve. Performance $A_t(s)$ is measured as top-1 answer accuracy for a target relation $t$ after training step $s$. Forgetting is analyzed through the trajectory of $A_t(s)$ across time steps and learning cycles. Memory collapse appears as singularities in these curves, where predictions across unrelated prompts converge to a narrow answer set [2305.09144].

## 4. Relevance Mapping Networks and controlled overlap

Relevance Mapping Networks are a strict continual-learning method: no replay, no multi-heads, and a single model. Training proceeds by initializing $\boldsymbol{\theta}$, adding a new relevance mapping for each task, optimizing weights and the current relevance map jointly, pruning small mask entries below threshold $p$, freezing parameters with $\mathbf{r}^{(i)}\approx 1$ from previous tasks, and increasing $\beta$ near the end of training so that masks become nearly binary. At inference time, the model computes task-wise activations for routing; no additional classifier heads are introduced [2102.11343].

The practical role of the three principal regularization parameters is explicit. The overlap term $\lambda$ controls representational sharing, the sparsity term $\gamma$ governs capacity allocation and stabilization, and the tightening parameter $\beta$ determines the hardness of the masks. Ablation results are correspondingly interpretable: removing the overlap term increases cross-task mask sharing and raises confusion in task inference; excessive sparsity limits capacity, reducing forward transfer and increasing catastrophic forgetting; and insufficient or premature mask tightening degrades learning dynamics [2102.11343].

Empirically, RMNs achieve state-of-the-art average accuracies without replay, multi-heads, or additional models: 97.73 ± 0.07% on Permuted MNIST, 99.5 ± 0.2% on Split MNIST, 85.33 ± 1.7% on Sequential Omniglot, 70.02 ± 2.5% on Split CIFAR-100 (10 tasks), and 80.01 ± 0.9% on ResNet-18 Split CIFAR-100 (20 tasks). In unsupervised continual learning, they report 97.73 ± 0.1% on P-MNIST, 99.5 ± 0.2% on Split MNIST, 85.33 ± 1.7% on S-Omniglot, and 99.1 ± 0.5% on Fuzzy S-MNIST, compared with 93.22 ± 0.07% for CN-DPM on the same noisy-transition setting. The paper characterizes RMNs as the only strict CL method that simultaneously alleviates catastrophic forgetting and preserves the discriminability required for unsupervised task inference [2102.11343].

Memory overhead is one mask per task over $d$ parameters, with near-binary masks that can be stored compactly at approximately $Kd$ bits if binarized. Computationally, the method adds the cost of learning masks and evaluating masked forward and backward passes, but it avoids replay overhead. This architecture is directly motivated by the claim that catastrophic memorizing is not solved by retention alone; it requires learning where to reuse parameters and where not to.

## 5. Relation to catastrophic forgetting, replay, and pseudo-recursal

The catastrophic memorizing problem is most visible when compared with methods designed primarily for catastrophic forgetting. Replay-based methods, rehearsal buffers, pseudo-rehearsal, distillation, and Fisher-based regularization target the preservation of earlier task performance. The continual-learning critique is that replay methods alleviate catastrophic forgetting but commonly aggravate catastrophic remembering by forcing the network to overgeneralize across all tasks, diminishing the ability to detect novel inputs or task changes [2102.11343].

Pseudo-recursal provides a clear example of a forgetting-oriented solution. In that framework, a classifier $f_\theta$ is trained on a mixture of current real data and GAN-generated pseudo-data labeled by the previous classifier snapshot, while a DCGAN is trained recursively on current images plus replayed pseudo-images so that a single generator retains the ability to sample across previously seen tasks. The classifier objective is

$$
L_{\text{class}}(\theta) =
\mathbb{E}_{(x,y)\sim D_t}[CE(f_\theta(x), y)]
+ \lambda \,\mathbb{E}_{z\sim p(z)}[CE(f_\theta(G_{t-1}(z)), f_{t-1}(G_{t-1}(z)))].
$$

The generator is unconditional, and task mixing is achieved by replay sampling rather than explicit conditioning. Mini-batch discrimination is used to reduce mode collapse [1802.03875].

On the CIFAR-10 $\rightarrow$ SVHN $\rightarrow$ MNIST sequence, pseudo-recursal reports that after training on all tasks, the classifier loses only 1.67% absolute accuracy on CIFAR-10 and gains 0.24% absolute accuracy on SVHN; at each transition, the drop on past tasks is at most 1.32%. By contrast, sequential learning without rehearsal drives previous-task accuracy to approximately 0%, and EWC remains substantially worse in the reported setting. Random pseudo-images fail because they do not lie on the data manifold: after 2,048 uniform draws, the classifier labeled almost all such images as birds or frogs, and rehearsal on these images did not retain prior tasks [1802.03875].

The same source explicitly states that “Catastrophic memorizing” is not discussed there. That omission is informative. Pseudo-recursal is a forgetting mitigation strategy: it preserves the classifier’s function on prior domains without storing past data or growing the network. It does not address the separate question of whether a retained system can still discriminate among tasks, detect novelty, or avoid false familiarity. This suggests that low forgetting and low catastrophic memorizing are related but non-equivalent desiderata.

## 6. Memory collapse in language models

In language models, catastrophic memorizing appears in a factual-knowledge acquisition testbed that studies retention, forgetting, relevance, and diversity across checkpoints. The experimental setup uses BERT-base, BERT-large, and GPT-2; the main body reports BERT results, and GPT-2 exhibits the same patterns in the appendix. The knowledge source is LAMA/Wikidata, with 23 relations grouped by type. For each relation, there are 10,000 facts and 5 semantically equivalent prompts, yielding 50,000 sentences per group. Sequential learning trains on one knowledge group per time step, with 23 checkpoints saved per epoch or learning cycle [2305.09144].

Vanilla language models are reported to be forgetful. Performance on a target knowledge type peaks immediately after learning that type and then decays sharply, even to near zero, after learning another type. Repetitive learning does not improve retention for randomly initialized models. Across target types, the average Pearson correlation coefficient between forgetting curves and answer-correlation curves surpasses 0.9 with low variance, showing that highly relevant newly learned knowledge reliably undermines memory for the target relation [2305.09144].

Pre-training changes this regime. Pre-trained models initially forget at early steps but gradually retain knowledge with repetitive learning; forgetting curves flatten and converge, the lower bound within a cycle rises, and variance across steps decreases. In a from-scratch pre-training experiment, the transition from forgetful to retentive behavior occurs early, after roughly 80,000 pre-training steps, and longer pre-training enlarges memory capacity and final performance. The paper concludes that pre-training is the core driver of the forgetful-to-retentive transition [2305.09144].

The catastrophic memorizing phenomenon appears under low-diversity knowledge. Knowledge relevance is quantified via Pearson correlation between answer distributions of target and newly learned relations, while knowledge diversification is quantified by the number of distinct answers and the entropy

$$
H_r = -\sum_i P_i \log_2 P_i,
$$

where $P_i = f_i / \sum_j f_j$. Low entropy and low answer diversity increase the risk of memory collapse. The “continent” relation is the canonical case: it has only 7 answers, entropy $H_r = 1.94$ versus an average entropy of 5.67 across relations, and the average number of distinct answers across relations is 407.91. This relation caused 39 collapse events, and after learning it the model would answer “Europe” for unrelated queries such as “The citizenship of Isaac Newton is [MASK]” [2305.09144].

Diversification mitigates collapse. Interleaving $K$ knowledge types per time step, with $K=1,2,4,8$, increases diversity and causes memorization to converge faster, while making collapses shallower and less frequent. Relevant knowledge blocks such as language of work, native language, languages spoken, and original language of film or TV cause larger dips for the target relation “official language” than irrelevant blocks such as occupation, manufacturer, record label, and original broadcaster. The language-model results therefore align with the continual-learning account: excessive similarity causes interference, while insufficient diversity causes domination by a narrow subspace of outputs [2305.09144].

The two literatures use different mechanisms and diagnostics, but they converge on a common principle. Preventing forgetting is not sufficient if the model loses the ability to keep tasks, domains, or answer distributions distinct. Catastrophic memorizing names that failure of discrimination. In continual learning it is formalized as catastrophic remembering and addressed through controlled overlap; in language models it appears as memory collapse under low-diversity knowledge and is reduced by pre-training, repetition, and interleaving.

Source: https://www.emergentmind.com/topics/catastrophic-memorizing