MICACL: Category-Aware Contrastive Learning for DFER

• MICACL is a dynamic facial expression recognition framework that integrates multi-instance learning with adaptive graph modeling, instance aggregation, and category-aware contrastive optimization.
• It utilizes a 3D CNN backbone along with GEIIM, WIAN, and MCCL modules to capture complex spatio-temporal patterns and mitigate long-tailed class distributions.
• Empirical evaluations on DFEW and FERV39K demonstrate significant improvements in WAR and UAR, particularly enhancing recognition for minority emotional expressions.

Searching arXiv for MICACL/MICAL to verify naming and citation targets. MICACL most specifically denotes Multi-Instance Category-Aware Contrastive Learning, a framework for long-tailed dynamic facial expression recognition that combines multi-instance learning, adaptive spatio-temporal interaction modeling, importance-weighted temporal aggregation, and category-aware contrastive optimization (Cui et al., 4 Sep 2025). In the arXiv literature, however, the string “MICACL” is not uniformly used: one seizure-detection paper states that MICACL is a misspelling of MICAL—“Mutual Information-based CNN-Aided Learned factor graphs” (Salafian et al., 2022)—and one speech-based Mild Cognitive Impairment study uses MICACL as a shorthand label in an overview of acoustic classification (Zhang et al., 2024). The dominant explicit acronym expansion attached to MICACL itself is therefore the long-tailed dynamic facial expression recognition method introduced in (Cui et al., 4 Sep 2025).

1. Nomenclature and disambiguation

The acronym has multiple nearby usages in recent preprints, but they refer to different research problems and method families.

String	Expansion or usage	Domain
MICACL	Multi-Instance Category-Aware Contrastive Learning	Long-tailed dynamic facial expression recognition
MICAL	Mutual Information-based CNN-Aided Learned factor graphs	Seizure detection from EEG
MICACL	Shorthand label in an overview of Mild Cognitive Impairment acoustic classification	Speech-based cognitive assessment

The seizure-detection work explicitly states that there is no alternative acronym “MICACL” for its method and that the appearance of that form is a misspelling of MICAL (Salafian et al., 2022). By contrast, the dynamic facial expression recognition paper uses MICACL as the formal name of the method and expands it as Multi-Instance Category-Aware Contrastive Learning (Cui et al., 4 Sep 2025).

A plausible implication is that references to MICACL require contextual disambiguation. In computer vision, it refers to the DFER framework of (Cui et al., 4 Sep 2025); in EEG seizure detection, the correct acronym is MICAL rather than MICACL (Salafian et al., 2022).

2. Problem setting in dynamic facial expression recognition

MICACL addresses dynamic facial expression recognition (DFER) in the wild, where two coupled difficulties are emphasized: long-tailed category distributions and complex spatio-temporal feature modeling (Cui et al., 4 Sep 2025). The paper frames each video $V$ as a bag $B$ of $T$ instances under a standard multi-instance learning formulation. In this work, $T = 16$ frames per video are sampled for training and inference, and the bag carries a video-level emotion label while instance labels are not available (Cui et al., 4 Sep 2025).

The representation is built on a standard 3D CNN backbone (R3D from Torchvision) with pretrained weights, with videos resized to $112 \times 112$. The backbone produces per-time features that are spatially pooled into instance embeddings, written in the paper as

$\mathbf{X} \in \mathbb{R}^{B \times T \times C},$

with batch size $B$, $T$ instances per bag, and channel dimension $C$ (Cui et al., 4 Sep 2025).

The classification target is optimized with a bag-level cross-entropy loss over $C$ classes: $B$0 The paper states that an auxiliary Multiscale Category-aware Contrastive Learning (MCCL) loss is added to counteract long-tailed bias (Cui et al., 4 Sep 2025).

The motivating claim is that head classes occupy a larger learned feature space, while minority classes suffer from biased decision boundaries. This suggests that MICACL is designed not merely as a temporal recognizer, but as a rebalancing mechanism for supervised video representation learning under heavy class skew.

3. Core architecture: GEIIM, WIAN, and MCCL

MICACL is organized around three named components: the Graph-Enhanced Instance Interaction Module (GEIIM), the Weighted Instance Aggregation Network (WIAN), and Multiscale Category-aware Contrastive Learning (MCCL) (Cui et al., 4 Sep 2025).

GEIIM

GEIIM models inter-instance dependencies through a learned adjacency matrix. The paper defines node embeddings $B$1 and $B$2, and computes the adaptive adjacency as

$B$3

The stated rationale is that ReLU suppresses low-quality/noisy relations and Softmax normalizes affinities to stabilize relation modeling (Cui et al., 4 Sep 2025).

Instead of a normalized GCN update, MICACL uses a linear diffusion-style interaction

$B$4

with learnable $B$5 balancing direct instance information and neighborhood-aggregated information (Cui et al., 4 Sep 2025). GEIIM then applies multiscale temporal convolutions to capture local-to-global temporal patterns, although the paper does not numerically specify kernel sizes or dilation rates.

WIAN

WIAN aggregates instances into a bag representation while emphasizing salient frames and attenuating noise. It augments an LSTM with a Dynamic Weight Gate (DWG) driven by instance importance. The importance weights are computed from the GEIIM outputs as

$B$6

The gate is then defined as

$B$7

The modified LSTM cell update is

$B$8

After recurrent aggregation, WIAN applies multi-head self-attention to $B$9 to produce a bag-level representation for classification (Cui et al., 4 Sep 2025).

MCCL

MCCL is the long-tail mitigation component. It defines a category-aware weight

$T$0

where $T$1 is the sample count of category $T$2, and $T$3 denotes predicted class probabilities (Cui et al., 4 Sep 2025). This weighting is explicitly described as inversely weighting frequent classes and upweighting hard-to-classify samples.

MCCL builds multiscale features through

$T$4

computes cosine similarity with temperature $T$5,

$T$6

and adjusts the temperature using the dataset class distribution (Cui et al., 4 Sep 2025). The scale-specific contrastive objective is

$T$7

with multiscale aggregation

$T$8

The paper further states that MCCL includes a Classification Enhancement Term (CET) combining feature augmentation regularization with an intensity-aware loss, with details placed in the appendix (Cui et al., 4 Sep 2025).

4. Training configuration and computational profile

The training pipeline uses R3D (Torchvision pretrained) as the backbone, AdamW as the optimizer, and a cosine learning rate schedule over 300 epochs, with initial learning rate $T$9, minimum learning rate $T = 16$0, and weight decay $T = 16$1 (Cui et al., 4 Sep 2025). The reported hardware is an RTX A6000.

The paper does not report an explicit resampling strategy or memory bank. Instead, class balancing is handled through the MCCL weighting mechanism and the multiscale contrastive objective (Cui et al., 4 Sep 2025). The hidden dimension $T = 16$2 in GEIIM, the value of $T = 16$3, the multiscale convolution settings, and the coefficient that would weight MCCL relative to cross-entropy are not specified in the reported summary.

The computational footprint is reported as 1.69 G FLOPs in the DFEW table (Cui et al., 4 Sep 2025). Parameter counts and runtime are not reported.

A plausible implication is that MICACL was positioned as a comparatively efficient long-tailed DFER architecture: it retains an R3D backbone rather than replacing it with a heavier temporal transformer, while moving much of the inductive bias into graph interaction, gated aggregation, and contrastive reweighting.

5. Empirical performance on DFEW and FERV39K

MICACL is evaluated on DFEW and FERV39K, using WAR and UAR as principal metrics (Cui et al., 4 Sep 2025). DFEW is described as containing 16,000+ clips from 1,500+ films, with 7 basic expressions and 5-fold cross-validation. FERV39K contains 38,935 clips across 22 sub-scenes (Cui et al., 4 Sep 2025).

Main results

Dataset	MICACL performance	Comparison noted in the paper
DFEW	WAR 69.56, UAR 63.21	Best WAR and UAR simultaneously
FERV39K	WAR 48.57, UAR 40.25	Highest WAR and UAR

On DFEW, MICACL reports per-emotion accuracies of Happiness 89.22, Sadness 69.48, Neutral 63.72, Anger 73.47, Surprise 60.92, Disgust 29.65, and Fear 42.57, together with WAR 69.56, UAR 63.21, and FLOPs 1.69 G (Cui et al., 4 Sep 2025). The same table reports the following selected comparisons: NR-DFERNet with WAR 68.19, UAR 54.21, Disgust 0.00, Fear 19.43; Former-DFER with WAR 65.70, UAR 53.69, Disgust 3.45, Fear 31.78; STT with WAR 66.45, UAR 54.58, Disgust 3.49, Fear 34.04; and M3DFEL with WAR 69.25, UAR 56.10, Disgust 0.00, Fear 31.63 (Cui et al., 4 Sep 2025).

On FERV39K, MICACL reports WAR 48.57 and UAR 40.25, compared with LOGO-Former at WAR 48.13, UAR 38.22; Former-DFER at WAR 46.85, UAR 37.20; and M3DFEL at WAR 47.67, UAR 35.94 (Cui et al., 4 Sep 2025).

The paper emphasizes the tail-class gains on DFEW, especially Disgust 29.65% and Fear 42.57%, in contrast to prior works that report near-zero Disgust and lower Fear performance (Cui et al., 4 Sep 2025). This supports the intended function of MCCL as a mechanism for reducing head-class domination in the learned representation.

6. Ablation findings, interpretation, and limitations

The ablation study on DFEW fold 5 reports the following WAR/UAR progression: Baseline (R3D + Self-Attention + Conv1D; CE loss) at 64.55 / 54.17; + GEIIM at 68.11 / 57.39; + GEIIM + WIAN at 68.62 / 56.26; and + GEIIM + WIAN + MCCL at 69.91 / 64.34 (Cui et al., 4 Sep 2025). The accompanying interpretation is explicit: GEIIM substantially boosts both WAR and UAR, WIAN contributes additional gains, especially in WAR, and MCCL delivers the largest UAR improvement (+8.08 over baseline).

The paper also notes that a t-SNE visualization shows tighter intra-class clusters and clearer inter-class separation than the baseline, which is presented as consistent with the combined effect of MCCL and the GEIIM+WIAN aggregation pipeline (Cui et al., 4 Sep 2025). This suggests that MICACL is meant to alter both optimization geometry and temporal evidence accumulation, rather than only reweighting the final classification loss.

The reported limitations are concrete. The adjacency matrix $T = 16$4 is learned from node embeddings rather than directly from content similarity between instance features; multiscale convolution settings are not specified; MCCL relies on class counts and predicted probabilities, so inaccurate early predictions may destabilize weighting; and R3D may be less expressive than temporal transformers, although transformers incur higher compute cost (Cui et al., 4 Sep 2025). Proposed extensions include content-adaptive graph learning, prototype-based supervised contrastive learning with a memory bank, effective-number class weighting, temporal transformers or hybrid CNN-Transformer backbones, and dynamic sampling strategies for long-tailed distributions (Cui et al., 4 Sep 2025).

In relation to neighboring uses of the acronym, the seizure-detection method MICAL offers a useful contrast: it combines neural mutual information estimation, a 1D CNN, and factor graph inference for multichannel EEG seizure detection on CHB-MIT (Salafian et al., 2022). The Mild Cognitive Impairment speech paper, by contrast, studies automatic detection of Mild Cognitive Impairment using high-dimensional acoustic features in spontaneous speech, comparing Random Forests, Sparse Logistic Regression, k-Nearest Neighbors, Sparse Support Vector Machine, and Decision Tree on the TAUKADIAL dataset (Zhang et al., 2024). These works are methodologically distinct from MICACL proper, and their juxtaposition mainly underscores that the string MICACL is ambiguous unless grounded by domain context and citation.

7. Position within current research usage

Within the supplied arXiv record, MICACL is best understood as a multi-instance, long-tail-aware DFER framework whose distinguishing combination is: adaptive inter-instance graph modeling through GEIIM, importance-aware recurrent aggregation through WIAN, and multiscale category-aware contrastive optimization through MCCL (Cui et al., 4 Sep 2025). Its reported empirical profile is defined by simultaneous gains in WAR and UAR, especially on minority categories such as Disgust and Fear, and by an ablation pattern in which the largest balanced-performance improvement is attributed to MCCL (Cui et al., 4 Sep 2025).

The broader terminological landscape remains unsettled. One paper treats MICACL as a misspelling of MICAL in EEG seizure detection (Salafian et al., 2022), while another uses the string as a label associated with Mild Cognitive Impairment acoustic classification rather than as a named method (Zhang et al., 2024). In encyclopedia usage, the most precise treatment is therefore to reserve MICACL for Multi-Instance Category-Aware Contrastive Learning unless the surrounding literature clearly indicates one of the alternative contexts.