MRAD-TF: Anomaly Detection & Polarization Control

• The paper introduces a train-free, memory-driven anomaly detection method using a frozen CLIP encoder and cosine similarity matching for robust zero-shot performance.
• The paper details an integrated TFLT polarization controller that achieves reset-free, megaradian-per-second SOP tracking through cascaded Mach–Zehnder interferometers.
• Both approaches offer scalable, high-accuracy results, ensuring minimal adaptation overhead in diverse applications from industrial inspection to high-speed optical interconnects.

MRAD-TF denotes two distinct high-impact technologies in contemporary research: (1) a “train-free” base model in the Memory-Retrieval Anomaly Detection (MRAD) framework for zero-shot anomaly detection leveraging a frozen CLIP image encoder and direct memory retrieval, and (2) an integrated thin-film lithium tantalate (TFLT) polarization controller enabling reset-free, megaradian-per-second (Mrad/s) tracking of state-of-polarization (SOP) in optical interconnects. While sharing the MRAD-TF label, these systems operate in separate domains—computer vision anomaly detection and integrated photonics, respectively—each representing state-of-the-art methodological advances (Xu et al., 31 Jan 2026, Gao et al., 7 Jan 2026).

1. MRAD-TF in Zero-Shot Anomaly Detection

MRAD-TF, within the MRAD (Memory-Retrieval Anomaly Detection) framework, is a train-free, non-parametric approach to zero-shot anomaly detection. It is architected around a frozen CLIP ViT image encoder producing a global “class” token ($q_{cls}$) and a set of local “patch” tokens ($q_{pat,u}$), which are queried against a two-level memory bank constructed from auxiliary labeled data. This memory-driven formulation bypasses parameter fitting and instead retrieves anomaly scores through direct cosine-similarity matching, enabling robust cross-domain zero-shot anomaly detection while eliminating the need for backpropagation or fine-tuning (Xu et al., 31 Jan 2026).

2. Architecture and Two-Level Memory Bank

The model’s backbone comprises a frozen CLIP ViT image encoder. Given an input image $I$, the encoder produces $q_{cls} \in \mathbb{R}^d$ and $U$ patch tokens $\{q_{pat,u}\}_{u=1}^U \in \mathbb{R}^d$. The two-level memory bank is constructed as follows:

• Image-level memory: Each auxiliary image $i$ yields a key $k_{cls}^i=\Phi_{cls}(I_i)$ and a binary label $e^i\in\{[1,0],[0,1]\}$, stacked as $K_{cls}\in \mathbb{R}^{N_c\times d}$ and $V_{cls}\in\mathbb{R}^{N_c\times 2}$.
• Pixel-level memory: For normal images, all patch embeddings are averaged to obtain a prototype $p_{norm}^i$; for anomalous images with pixel mask $M_i$, patch averages are computed inside ($p_{anom}^i$) and outside ($p_{norm}^i$) the mask. Their respective labels form $K_{pat}$ and $V_{pat}$.

No modification of the encoder parameters occurs post-memory construction, ensuring true zero-shot operation with frozen backbone representations.

3. Retrieval-Based Inference and Scoring

During inference, a test image’s $q_{cls}$ and $\{q_{pat,u}\}$ are compared to the memory banks via cosine similarity, followed by softmax normalization with temperature $T=1$:

• Image-level score: $$Y_{cls} = \text{softmax}(S(q_{cls}, K_{cls}) /T ) \cdot V_{cls}$$, producing $(Y_{norm}, Y_{anom})$.
• Patch-level map: For each patch, $$Y_{seg}[u] = \text{softmax}(S(q_{pat,u}, K_{pat}) / T) \cdot V_{pat}$$. These form $Y_{seg} \in \mathbb{R}^{U\times 2}$, upsampled to $M\in \mathbb{R}^{H\times W}$.
• Final anomaly score: $A(I) = Y_{cls}^{anom} + \text{TopKMean}(M_a)$, where $M_a$ is the upsampled anomaly-channel, and TopKMean averages the top $1\%$ of pixels.

The explicit memory-based approach preserves the empirical data distribution, ensuring that diversity in auxiliary data directly translates to detection capability.

4. Computational Efficiency and Empirical Performance

Inference with MRAD-TF involves a single forward pass through the encoder and two batched matrix multiplications: $O(dN_c)$ for image-level and $O(UdN_p)$ for patch-level retrievals. With ViT-L/14 (U=196, d=768, N≈3K), processing time is approximately 200 ms/image on an RTX 3090; memory usage is ~15 MB for $\sim$5,000 vectors (Xu et al., 31 Jan 2026).

Empirical evaluation across 16 industrial and medical benchmarks demonstrates: | Metric | MRAD-TF | Best prior (WinCLIP) | |---------------|---------|----------------------| | P-AUROC (mean)| 85.5% | 73.0% | | I-AUROC (mean)| 81.0% | 75.1% | | Mean AP | 83.2% | 73.2% | | Mean PRO | 64.6% | 42.9% |

Across all datasets (e.g., MVTec-AD, VisA, ISIC, HeadCT), MRAD-TF outperforms train-free competitors without incurring training or adaptation cost.

5. Advantages, Limitations, and Applicability

Advantages of MRAD-TF include:

• True zero-shot regime: No gradient updates; instant deployment with only an auxiliary set.
• Explicit memory realization: Empirical data distribution is preserved, avoiding information collapse typical in parametric approaches.
• Minimal overfitting and strong cross-domain robustness: Owing to the absence of trainable parameters and highly parallel similarity search.

Limitations involve:

• Recognition range: Performance depends on the coverage of normal/anomaly patterns in the auxiliary set.
• Fixed metric: Inability to adapt to subtle target-domain shifts (addressed by subsequent MRAD-FT variant).
• Linear scaling with memory size: Retrieval latency increases with enlarged memory; approximate nearest-neighbor search or compression may be required for large-scale, low-latency systems.

Best-use scenarios include industrial or medical deployments with substantial labeled auxiliary datasets and no available target-domain supervision, or for rapid prototyping in zero-shot settings.

6. MRAD-TF in Integrated Polarization Control

MRAD-TF also designates a thin-film lithium tantalate (TFLT) polarization controller for optical links, capable of reset-free, megaradian-per-second (Mrad/s) SOP tracking (Gao et al., 7 Jan 2026). The platform consists of:

• Device: TFLT-on-insulator wafer, x-cut LiTaO₃, 400 nm TaO₅ guiding layer, 240 nm ridge geometry, gold electrodes (single-drive push-pull), with $\text{r}_{eff}\sim 17$–$20$ pm/V and a measured V$_\pi \approx 2.46$ V.
• Architecture: Four cascaded Mach–Zehnder interferometer phase-shifter stages provide full SU(2) coverage on the Poincaré sphere, with overall polarization-dependent loss (PDL) <0.3 dB.
• Control algorithm: Finite-boundary gradient descent (FBGD) augments conventional gradient-descent SOP control with a boundary-regularization term, ensuring that phase shifters remain within safe electrical operating ranges, thus avoiding abrupt phase resets. This assures uninterrupted, reset-free SOP evolution.

7. Experimental Validation and System-Level Context

Benchmarks with electronic polarization scramblers and dual-polarization 16-QAM self-homodyne coherent links demonstrate:

• Tracking: Instantaneous step re-lock within 100 ns; sustained tracking speeds up to 2 Mrad/s (transient) and 1 Mrad/s (continuous) with <0.3 relative intensity error for 99.9% of samples.
• System performance: In a 400 Gb/s transmission, pre-FEC BER stays below HD-FEC threshold for scrambling rates up to 1 Mrad/s, with SNR penalty ≤0.5 dB compared to static SOP. Tracking saturates above 2 Mrad/s due to phase range constraints, not loss of lock.

The device is drift-free, sub-2.5 V, low PDL, and low power (<10 mW/stage), with modulation bandwidth exceeding 80 GHz. Compared to silicon photonic or thin-film LiNbO₃ APCs, TFLT achieves an unmatched performance regime (>1 Mrad/s tracking, V$_\pi<$2.5 V, PDL < 0.3 dB).

Applications span high-speed, AI-driven data center interconnects, dual-pol IMDD links, and any optical system—lidar, quantum, microwave photonics—requiring ultrafast, continuous polarization stabilization.

---

MRAD-TF, whether in the context of memory-driven, zero-shot image anomaly detection or integrated optical polarization management, represents a state-of-the-art methodology for robust, scalable performance without reliance on iterative parameter learning or manual intervention (Xu et al., 31 Jan 2026, Gao et al., 7 Jan 2026).