BRAIN: Bias-Mitigation Continual Learning

• The paper introduces BRAIN, a framework that treats fMRI sessions sequentially to address session shift, representation drift, compounding bias, and catastrophic forgetting.
• It employs debias contrastive learning by reweighting session-specific signals using response accuracy or brain activation measures.
• The angular-based forgetting mitigation regularizer preserves intermediate feature geometry, significantly enhancing retrieval performance in continual fMRI learning.

Bias-Mitigation Continual Learning (BRAIN) is a continual-learning framework for Vision-Brain Understanding (VBU) that treats longitudinal fMRI acquisition as a non-stationary sequence of recording sessions rather than as a single pooled dataset. It was introduced to address the observation that brain responses are not stationary over time: as participants repeatedly view stimuli across sessions, memory for earlier stimuli decays, confidence decreases, recorded brain signals become weaker and less consistent, and the learned representation becomes increasingly misaligned. In this formulation, BRAIN targets four coupled phenomena—session shift, representation drift, compounding bias, and catastrophic forgetting—by combining a bias-aware contrastive objective with a continual-learning regularizer that preserves previously learned feature geometry (Nguyen et al., 25 Aug 2025).

1. Problem setting and conceptual scope

BRAIN is motivated by the realistic VBU setting in which fMRI data arrive over many recording sessions. At continual-learning step $t$, the model receives a session dataset

$\mathcal{D}_t = \{(x^t, y^t)\},$

where $x^t$ denotes the fMRI signal and $y^t$ the corresponding image. An fMRI encoder $F$ produces

$z^t = F(x^t, \theta_t) \in \mathbb{R}^d,$

while a frozen visual encoder $V$ maps the image into a visual centroid

$c^t = V(y^t) \in \mathbb{R}^d.$

The model is then updated incrementally as new sessions arrive, rather than retrained from scratch on all sessions jointly (Nguyen et al., 25 Aug 2025).

The paper grounds this setup in two kinds of evidence from the Natural Scenes Dataset (NSD). First, metadata on participant responses to whether a stimulus had been seen before show that both response accuracy and consistency decrease over time. Second, an experimental model-performance test with a MindEye2-style retrieval model, trained and validated on session chunks with identical architecture and hyperparameters across steps, shows that retrieval performance drops noticeably as sessions progress. These results motivate the paper’s terminology: bias denotes a time-varying distortion in the fMRI distribution caused by memory decay and uncertainty; session shift denotes the change from $p^{t-1}(x)$ to $p^t(x)$; representation drift denotes the induced shift in learned features; and forgetting denotes the loss of knowledge about previous sessions when learning new ones (Nguyen et al., 25 Aug 2025).

This framing is narrower and more specific than generic continual-learning accounts of non-stationarity. The central claim is not merely that later sessions are different, but that later-session data are systematically less reliable and therefore introduce a compounding bias if training assumes stationarity. In that sense, BRAIN is a bias-mitigation method because it explicitly attempts to correct the current session’s distortion while preventing that distortion from overwriting earlier, cleaner knowledge.

2. Framework architecture and optimization objective

BRAIN adopts a continual-learning architecture with two coupled mechanisms: one for current-session debiasing and one for cross-session retention. The generic optimization target at step $\mathcal{D}_t = \{(x^t, y^t)\},$0 is

$\mathcal{D}_t = \{(x^t, y^t)\},$1

where $\mathcal{D}_t = \{(x^t, y^t)\},$2 is a contrastive alignment loss between brain and image embeddings, $\mathcal{D}_t = \{(x^t, y^t)\},$3 is a continual-learning loss that preserves prior knowledge, and $\mathcal{D}_t = \{(x^t, y^t)\},$4 controls the tradeoff (Nguyen et al., 25 Aug 2025).

Component	Function	Mechanism
De-bias Contrastive Learning	Mitigate current-session bias	Session-dependent reweighting of contrastive alignment
Angular-based Forgetting Mitigation	Reduce catastrophic forgetting	Angular discrepancy between old and current intermediate features
Continual session update	Adapt to temporal drift	Initialize each step from previous weights

Within this design, BRAIN does two things simultaneously. It corrects the learning signal from the current session, which may already be biased by degraded memory and uncertainty, and it constrains the model so that adaptation to the new session does not destroy the representational structure learned from earlier sessions. The paper emphasizes that this differs from simply pooling all sessions: pooling assumes that all sessions are identically distributed, whereas BRAIN assumes that the temporal order is informative and that later sessions are harder in a systematic way (Nguyen et al., 25 Aug 2025).

The method is therefore neither a pure replay strategy nor a pure regularization strategy in the ordinary continual-learning sense. Its contrastive term is already modified to reflect session-dependent bias, and its retention term is chosen to preserve angular structure rather than absolute feature magnitudes.

3. De-bias Contrastive Learning

The first technical contribution is De-bias Contrastive Learning (DCL). The paper starts from the observation that prior VBU methods implicitly assume a stationary fMRI distribution,

$\mathcal{D}_t = \{(x^t, y^t)\},$5

and then rewrites the expected contrastive loss to expose a session-dependent bias factor. This leads to

$\mathcal{D}_t = \{(x^t, y^t)\},$6

which is interpreted as a measure of how far the current session has drifted from the unbiased or ideal distribution (Nguyen et al., 25 Aug 2025).

Because participant response accuracy decreases with session index, the paper models this factor as

$\mathcal{D}_t = \{(x^t, y^t)\},$7

where $\mathcal{D}_t = \{(x^t, y^t)\},$8 is the response accuracy at session $\mathcal{D}_t = \{(x^t, y^t)\},$9. The resulting DCL loss is written as

$x^t$0

Here, $x^t$1 is the similarity between the brain embedding and the visual centroid, the numerator pulls the correct pair together, the denominator contrasts it against alternative brain features $x^t$2, and the factor $x^t$3 increases when participant accuracy is low, i.e. when the session is more biased (Nguyen et al., 25 Aug 2025).

The paper also provides an alternative estimate of the bias factor based on brain activity rather than behavioral response: $x^t$4 where $x^t$5 is the number of activated voxels and $x^t$6 is the total number of voxels. This variant is presented as a fallback when response labels are unavailable. In the ablation study, the brain-activation-based variant remains superior to prior baselines but is slightly worse than the response-accuracy version, making response accuracy the preferred bias signal when available (Nguyen et al., 25 Aug 2025).

The intended effect of DCL is not to equalize sessions or discard later data. Rather, it reweights learning so that harder, noisier, and more drifted sessions receive an explicitly bias-aware correction. This is the part of BRAIN that addresses compounding bias directly.

4. Angular-based Forgetting Mitigation

The second technical contribution is Angular-based Forgetting Mitigation (AFM), which supplies the continual-learning regularizer. The paper formulates the retention term as a feature-distillation objective over intermediate layers: $x^t$7 where $x^t$8 is the number of intermediate feature layers and $x^t$9, $y^t$0 are the features produced by the old and current models at layer $y^t$1 (Nguyen et al., 25 Aug 2025).

Instead of the Euclidean discrepancy often used in prior work,

$y^t$2

BRAIN uses an angular discrepancy: $y^t$3 This is effectively a cosine-similarity-based regularizer. The rationale given in the paper is threefold: angular distance is less sensitive to scale changes in the representation space, it matches downstream retrieval tasks that already rely on angular or cosine similarity, and it allows the model to absorb new session-specific information without over-constraining feature magnitude (Nguyen et al., 25 Aug 2025).

The full BRAIN objective combines DCL and AFM: $y^t$4 or, more explicitly,

$y^t$5

At each learning step, the model is initialized from the previous step’s weights and trained on the current session with this joint loss. The first term addresses current-session bias; the second term preserves the angular geometry of earlier knowledge (Nguyen et al., 25 Aug 2025).

A persistent misconception in continual-learning discussions is that any strong retention term must necessarily freeze adaptation. AFM is intended to avoid exactly that failure mode. Because it constrains directions rather than magnitudes, the paper presents it as a softer and more geometry-aware retention mechanism than ordinary $y^t$6-style feature matching.

5. Benchmarks, protocols, and empirical findings

The primary benchmark is NSD, which the paper considers especially suitable for continual-learning analysis because it provides multiple participants, many recording sessions per participant, repeated image presentations across sessions, fine-grained session metadata, and behavioral response information. Continual-learning splits are defined as $y^t$7, where $y^t$8 is the number of sessions used to train the initial model and $y^t$9 is the number of sessions added per continual-learning step. Four protocols are evaluated: $F$0, $F$1, $F$2, and $F$3. Evaluation is performed on the current and previous sessions only, and retrieval is measured in both directions—Brain $F$4 Image and Image $F$5 Brain—using cosine similarity in a shared embedding space, typically in a 200-way setting with random distractors and Top-1 accuracy (Nguyen et al., 25 Aug 2025).

The baselines include W/o CL, LwF, PLOP, and a Non-CL baseline that trains on all sessions jointly. Across all continual-learning settings, BRAIN outperforms the continual-learning baselines. A representative example is subject 01 under $F$6 for Brain $F$7 Image retrieval: W/o CL drops from 61.57% at the first step to 18.06% at the last step, whereas BRAIN maintains 57.97%. The average results in the main table also show BRAIN outperforming competing methods across all split settings and retrieval directions. The paper further reports that BRAIN can be competitive with, and sometimes surpass, the Non-CL baseline, which is notable because continual learning often sacrifices some accuracy for retention (Nguyen et al., 25 Aug 2025).

The ablation study isolates both major components. Replacing standard contrastive alignment with DCL + RA yields an improvement typically around 1%–4% over vanilla contrastive learning with $F$8 regularization. Replacing Euclidean forgetting regularization with AFM improves performance by roughly 1%–3% over the corresponding $F$9 version. The full model, DCL + RA + AFM, gives the best performance in both retrieval directions and across all subject/session configurations. A rehearsal-based variant retaining 10% of previous data performs worse than the rehearsal-free version; the explanation proposed in the paper is that session-specific clusters are not well preserved by such a small rehearsal buffer (Nguyen et al., 25 Aug 2025).

Taken together, these results support the paper’s main claim that the degradation of longitudinal VBU models is not only a forgetting problem and not only a data-quality problem. It is the interaction of session-dependent bias and sequential updating, and BRAIN is designed to intervene at both points simultaneously.

6. Position within bias-mitigation continual learning research

The term bias is used differently across the broader continual-learning literature. BRAIN’s bias is a temporal distortion in brain-signal distributions caused by memory decay and uncertainty, whereas other works study spurious correlation, group fairness, online posterior bias, normalization-statistics recency bias, or noisy-label bias.

Work	Bias target	Continual setting
BRAIN (Nguyen et al., 25 Aug 2025)	Session shift, representation drift, forgetting in longitudinal fMRI	VBU retrieval over recording sessions
BM-CL (Mansilla et al., 1 Sep 2025)	“Leveling-down effect” between best and worst groups	Fairness mitigation as domain/task-incremental CL
UER (Lin et al., 2023)	Bias toward current classes in dot-product logits	Online continual learning with replay
CBA / Dual-CBA (Wang et al., 2023, Wang et al., 2024)	Posterior bias under catastrophic distribution shift	Online class-incremental CL
AdaB$z^t = F(x^t, \theta_t) \in \mathbb{R}^d,$0N (Lyu et al., 2023)	Recency bias in BatchNorm statistics	Online and offline CL
ErrorEraser (Cao et al., 11 Jun 2025)	Bias-induced erroneous knowledge from noisy labels	Task-incremental CL
Fair affective robotics (Kara et al., 2021)	Demographic bias across gender and race	Domain-incremental FER and AU learning

This comparison matters because BRAIN is not a fairness method in the group-based sense. BM-CL, for example, treats fairness improvement as a continual-learning problem and uses LwF or EWC to preserve best-group performance while improving worst-group performance, thereby reducing the “leveling-down effect” (Mansilla et al., 1 Sep 2025). By contrast, BRAIN has no best-group/worst-group partition; its central non-stationarity is temporal and session-indexed. Likewise, UER addresses current-class bias in online replay by separating the roles of angle and norm in dot-product logits (Lin et al., 2023), while CBA and Dual-CBA learn adaptors to absorb posterior bias caused by online distribution shifts (Wang et al., 2023, Wang et al., 2024). AdaB$z^t = F(x^t, \theta_t) \in \mathbb{R}^d,$1N isolates a different mechanism—recency bias in normalization statistics (Lyu et al., 2023)—and ErrorEraser treats continual learning as a problem of selective forgetting of bias-induced erroneous memories under noisy labels (Cao et al., 11 Jun 2025).

An earlier line of work on spurious correlation in continual learning showed that bias can transfer both forward and backward across tasks, can accumulate in longer task sequences, and can be reduced by a plug-in method called Group-class Balanced Greedy Sampling (BGS) with at most a slight loss of continual-learning performance (Lee et al., 2023). That result is conceptually adjacent to BRAIN because both reject the assumption that continual learning need only preserve knowledge; they differ in the source of the bias being preserved. The same broader intuition also appears in affective robotics, where demographic variation is treated as a domain-incremental continual-learning problem to improve fairness across gender and race in facial expression recognition and action unit detection (Kara et al., 2021).

Within this landscape, BRAIN’s distinct contribution is to make VBU explicitly temporal, bias-aware, and continual. It treats longitudinal fMRI collection not as a nuisance around an otherwise static retrieval model, but as the defining structure of the learning problem. A plausible implication is that future VBU systems will need to combine session-aware bias modeling with more general continual-learning tools rather than relying on pooled-data training or purely task-agnostic replay.