📊 statistics

Medix: Out-of-Distribution Detection from Unlabeled Wild Data via Robust Gradient Statistics

The paper introduces Medix, a novel framework that leverages median-based robust gradient statistics to identify outliers from unlabeled wild data, enabling the training of a high-performance out-of-distribution classifier that outperforms existing methods in open-world settings.

Original authors: Momin Abbas, Ali Falahati, Hossein Goli, Mohammad Mohammadi Amiri

Published 2026-07-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Momin Abbas, Ali Falahati, Hossein Goli, Mohammad Mohammadi Amiri

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a security guard at a very exclusive club (the AI model). Your job is to let in only people who belong to the club's specific group (In-Distribution data) and turn away anyone who doesn't (Out-of-Distribution or OOD data).

The problem is that in the real world, you don't have a perfect list of "club members" to compare everyone against. Worse, you often have to check a huge, messy crowd of people (unlabeled wild data) where you don't know who belongs and who doesn't. Some are members, some are strangers, and they are all mixed together.

Traditional methods try to guess who is who, but they often get confused by the noise. This paper introduces a new system called Medix that acts like a super-smart, noise-canceling filter to solve this problem.

Here is how Medix works, explained through simple analogies:

1. The Problem: The "Noisy Crowd"

Imagine you are trying to find the average height of a specific group of people (the club members). But you are standing in a giant, chaotic crowd where club members are mixed with strangers, giants, and dwarfs. If you just take the average (the mean) of everyone's height in that crowd, the giants and dwarfs will skew the result, making your "average" wrong.

In AI terms, if you try to train a model using this messy mix of data without knowing who is who, the "strangers" (outliers) mess up the model's understanding of what a "normal" input looks like.

2. The Solution: The "Median" Filter

Medix uses a clever trick called the Median.

The Average (Mean): If you have 9 people who are 5 feet tall and 1 giant who is 10 feet tall, the average height jumps up to about 5.5 feet. The giant pulls the average up.
The Median: If you line everyone up from shortest to tallest, the median is the person right in the middle. Even with that one giant, the median stays at 5 feet. The giant is ignored because they are an extreme outlier.

Medix uses this "middle-ground" logic to find the true center of the data, ignoring the noisy strangers.

3. How Medix Works (The Three Steps)

Step 1: Learning the "Normal" Vibe
First, Medix looks at a clean, labeled group of club members (In-Distribution data) to learn what their "vibe" looks like. In technical terms, it calculates a Reference Gradient. Think of this as taking a perfect snapshot of the club members' average behavior.

Step 2: The "Squeeze" (Filtering the Crowd)
Now, Medix looks at the messy, unlabeled crowd. It calculates the "vibe" (gradient) of every person in that crowd.

It asks: "If I remove this specific person, does the crowd's average vibe get closer to the 'Normal' vibe we learned in Step 1?"
If removing a person makes the crowd look more like the normal club members, that person was likely a stranger (an outlier) messing things up.
Medix does this iteratively, like a game of musical chairs where it keeps kicking out the people who are pulling the group's "vibe" away from the truth. It uses the Median to ensure that a few loud outliers don't trick the system.

Step 3: Training the Security Guard
Once Medix has filtered out the strangers and identified the "suspects" (the outliers it successfully found), it takes those suspects and the original club members to train a new, super-smart security guard (the OOD detector).

It teaches the guard: "These people are members. These people we found in the crowd are definitely NOT members."
Because the "suspects" were found using a robust, noise-resistant method, the new guard is much better at spotting strangers in the future.

4. Why It's Better Than Others

The paper claims that other methods are like trying to find a needle in a haystack by looking at the average of the hay. If the hay is wet or dirty, the average gets messed up.

Medix is like using a magnet that only picks up the metal (the true signal) and ignores the dirt and noise.

Robustness: Even if 50% of the crowd is strangers, Medix can still find the true center of the club members because the Median is stubborn and doesn't care about extreme values.
No Perfect Lists Needed: Unlike some other methods that require a clean list of "bad guys" to train on, Medix can find the bad guys inside a messy, unlabeled pile of data.

The Results

The authors tested Medix on various image datasets (like CIFAR-10 and CIFAR-100). They found that Medix was significantly better at spotting "strangers" than 20 other popular methods.

It reduced the error rate (False Positive Rate) by a huge margin compared to the next best method.
It proved mathematically that this "Median" approach is reliable and won't fail even when the data is very noisy.

In short: Medix is a smart filter that uses the "middle ground" (Median) to ignore the noise in a messy crowd, allowing an AI to learn exactly what belongs and what doesn't, even when it hasn't been given a perfect list of rules.

Technical Summary: Medix – Out-of-Distribution Detection from Unlabeled Wild Data via Robust Gradient Statistics

Problem Statement

Deploying machine learning models in real-world environments exposes them to Out-of-Distribution (OOD) inputs—data from unknown categories that the model was not trained to recognize. While neural networks are inherently vulnerable to such inputs, detecting them during inference is challenging because models are not exposed to unknown distributions during training.

Recent approaches have attempted to utilize "in-the-wild" unlabeled data to improve OOD detection. However, a significant hurdle remains: wild data is a complex mixture of In-Distribution (InD) and OOD samples. Unlike methods that assume access to a clean, auxiliary OOD dataset (e.g., Outlier Exposure), real-world wild data lacks distinct labels. The absence of a clean set of OOD samples complicates the training of an optimal OOD classifier, as the model must distinguish between InD and OOD samples within a contaminated distribution without explicit supervision.

Methodology: The Medix Framework

The authors propose Medix, a novel framework designed to identify potential outliers from unlabeled wild data using element-wise median (EWM) robust gradient statistics. The framework operates in two primary stages:

1. Outlier Extraction (Filtering)

The core innovation of Medix lies in its ability to filter candidate OOD samples from a mixed wild dataset ( $S_{wild}$ ) without prior knowledge of their labels.

Reference Gradient Estimation: First, an InD classifier ( $f_\phi$ ) is trained solely on labeled InD data ( $S_{in}$ ). The mean gradient ( $\bar{\nabla}_{in}$ ) of the loss function with respect to the model parameters is computed over $S_{in}$ . This serves as the reference gradient.
Robust Filtering via EWM: The method hypothesizes that the presence of OOD samples in the wild data causes a deviation between the mean InD gradient and the central tendency of the wild data's gradients. To identify this central tendency robustly, Medix employs the element-wise median of the gradients rather than the mean. The median is chosen for its stability against noise and outliers.
Iterative Greedy Removal: The framework formulates an optimization problem to find a subset of the wild data that minimizes the distance between its EWM and the reference $\bar{\nabla}_{in}$ $\overset{ˉ}{\nabla}_{in}$ . Since solving this exactly is intractable, Medix uses a greedy, iterative "leave-one-out" approach (Algorithm 1):
1. Compute the EWM of the current wild set.
2. Identify samples whose removal results in the largest reduction in the L2-distance between the new EWM and $\bar{\nabla}_{in}$ .
3. Remove the top- $k$ such samples as candidate outliers.
4. Repeat until the reduction in distance falls below a threshold $\epsilon$ .
Result: This process isolates a set of candidate outliers ( $\hat{S}_{out}$ ) and retains a cleaner set of InD-like samples.

2. Detector Training

Once candidate outliers are extracted, the framework trains a binary OOD detector ( $g_\theta$ ).

The detector is trained using the original labeled InD data ( $S_{in}$ ) as positive examples and the extracted candidate outliers ( $\hat{S}_{out}$ ) as negative examples.
The objective is to maximize the separability between these two groups, enforcing a decision boundary where InD samples yield positive outputs and outliers yield negative outputs.

Theoretical Analysis

The paper provides formal theoretical guarantees for the filtering stage under the Huber contamination model, where the wild data is a mixture of InD and OOD distributions with contamination ratio $\pi$ .

Inlier Misclassification Bound: The authors derive an upper bound on the fraction of true InD samples incorrectly discarded as outliers. This bound depends on the contamination ratio and the concentration of InD gradients (assumed to be sub-Gaussian).
Outlier Misclassification Bound: Conversely, they bound the fraction of true OOD samples incorrectly retained as InD. This bound relies on the separation between the mean OOD gradient and the reference InD gradient, as well as the sub-Gaussian nature of OOD gradients.
Key Insight: The analysis demonstrates that median-based filtering achieves robust separation even when OOD samples constitute a significant portion of the mixture (up to 50%), provided the distributions are sufficiently separated.

Experimental Results

Medix was evaluated across eleven InD-OOD dataset pairs, primarily using CIFAR-10 and CIFAR-100 as InD datasets and various datasets (SVHN, PLACES365, LSUN, TEXTURES) as OOD sources. The method was benchmarked against 20 competitive baselines, including:

InD-only methods: MSP, ODIN, Mahalanobis, Energy, ReAct, DICE, KNN, ASH, CSI, KNN+.
Wild-data methods: Outlier Exposure (OE), Energy-regularization, WOODS, CONJ, DRL.

Key Findings:

Superior Performance: Medix outperformed all baselines across the board. On CIFAR-100, it reduced the average False Positive Rate at 95% True Positive Rate (FPR95) by 40.98% compared to the strong baseline KNN+ and by 1.32% compared to the previous state-of-the-art wild-data method, WOODS.
Robustness to Contamination: The method maintained high performance even when the wild data contained a 50% mixture of OOD samples.
Scalability: Experiments on larger architectures (ResNet-50) and large-scale datasets (ImageNet-1k) confirmed that the L2-based filtering signal remains robust in high-dimensional gradient spaces.
Ablation Studies: The authors demonstrated that the performance gains stem specifically from the precision of the median-based filtering, rather than the two-stage pipeline alone. Replacing the median filter with a naive Maximum Softmax Probability (MSP) filter resulted in a significant performance collapse.

Significance and Claims

The paper claims that Medix addresses a critical gap in OOD detection: the ability to leverage unlabeled, mixed "in-the-wild" data without requiring clean auxiliary OOD datasets or strong distributional assumptions.

Novel Mechanism: By leveraging the robustness of the element-wise median, Medix provides a theoretically grounded mechanism to filter outliers from mixed data, a task previously considered difficult due to the lack of distinct OOD labels.
Theoretical Foundation: The work establishes provable error bounds for both inlier and outlier misclassification, offering a formal theoretical foundation for the "in-the-wild" setting, an area where such foundations were previously scarce.
Practical Impact: The method offers greater flexibility for real-world deployment where models must operate in dynamic environments with unknown data distributions. The empirical results suggest that Medix can significantly enhance the safety and reliability of machine learning systems by effectively identifying and handling OOD inputs using readily available unlabeled data.

The authors conclude that Medix represents a significant step forward in open-world learning, demonstrating that robust gradient statistics can effectively convert unlabeled wild data into useful supervision for OOD detection.