🤖 machine learning

Breaking Diversity Collapse in Spiking Pseudo-Ensembles for Efficient OOD Detection in Remote Sensing

This paper proposes an efficient spiking pseudo-ensemble framework for remote sensing that mitigates diversity collapse in lightweight classification heads through a novel agree-disagree objective, achieving OOD detection performance comparable to deep ensembles while significantly reducing computational cost and parameter count.

Original authors: Srinivas Anumasa, Rushi Shah, Qiran Zou, Dianbo Liu

Published 2026-08-04

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Srinivas Anumasa, Rushi Shah, Qiran Zou, Dianbo Liu

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 the captain of a spaceship exploring a vast, unknown galaxy. Your ship is equipped with a super-smart AI navigator that has studied millions of pictures of Earth's landscapes to recognize forests, deserts, and cities. But space is full of surprises: strange new planets, weird lighting, and alien terrain that the AI has never seen before. If the AI confidently guesses "That's a forest!" when it's actually a weird alien rock, your mission could fail. This is the problem of "Out-of-Distribution" (OOD) detection: teaching an AI to say, "I don't know what this is," instead of guessing wildly.

To solve this, scientists often use a trick called an "ensemble." Instead of relying on one AI, they ask a team of five different AIs to look at the same picture and vote. If they all agree, you trust the answer. If they argue, you know something is weird. However, running five full AI brains is expensive and slow, like hiring five expensive guides for a short hike. A cheaper idea is to have one smart guide with five small, lightweight assistants who just add their opinions. But here's the catch: if you just ask five assistants to guess, they often end up thinking exactly the same thing, making the team no smarter than a single person. This paper tackles that specific problem: how to get a team of cheap assistants to actually disagree with each other when they are confused, without hiring five full guides.

The Problem: When the Team Thinks Alike

The researchers are working with a special type of AI called a Spiking Neural Network (SNN). Think of these as ultra-efficient, low-power brains that work like neurons in your body, firing tiny electrical "spikes" only when necessary. They are perfect for satellites and drones that have limited battery power.

The team wanted to use a "pseudo-ensemble" method. Imagine a master chef (the backbone) who prepares a complex dish (the image features) and then hands it to five junior chefs (the heads) to add their own final seasoning. The goal is for the five junior chefs to give slightly different opinions so the master can spot if the dish is weird.

However, the researchers found a glitch they call "diversity collapse." If you just tell the five junior chefs to "guess the right answer" using the same instructions, they all end up tasting the dish the exact same way. Even though they are different people with different brains, they converge on the same prediction. It's like asking five students who all studied from the same textbook to solve a tricky math problem; they might all get the same wrong answer because they learned the same shortcuts. This makes the "team" useless for spotting strange, unfamiliar inputs.

The Solution: The "Agree-Disagree" Game

To fix this, the authors proposed a clever new training game called the "Agree-Disagree" objective. They realized that to get the junior chefs to think differently, you can't just ask them about the normal dishes they know. You need to give them a slightly distorted version of the dish to see how they react.

Here is how they did it:

The Setup: They took a pre-trained master chef (the frozen SNN backbone) and attached five lightweight junior chefs (heads).
The Agreement: When the chefs see a clear, normal picture (like a sunny forest), they must all agree on the correct label. This ensures they still know how to do their job.
The Disagreement: When the chefs see a picture that has been blurred (like looking at the forest through a foggy window), the researchers told them: "You don't need to agree here! In fact, we want you to disagree!"

Why blur? The researchers tested many ways to mess up the images, like adding noise or rotating them. They found that box blur (smoothing the image) was the perfect tool. It keeps the general shape of the scene but removes the fine details. When the image is blurry, the five junior chefs naturally start to guess differently because the clues are fuzzy. By training them to express these different guesses on blurry images, the team learns to have "functional diversity."

The Results: Smarter Teams, Less Fuel

The team tested this idea on remote sensing images (pictures of Earth from space) using two different types of AI architectures: a Transformer-based one called Spikformer and a convolutional one called ResNet19-SNN.

The results were impressive. When they used their "Agree-Disagree" method with three master chefs, each equipped with five lightweight junior heads, the system performed just as well as a traditional "Deep Ensemble" that used five full, heavy master chefs.

Here is the magic math:

The new method used about 38% fewer parameters (less memory/storage) compared to the five-model deep ensemble.
It required 40% fewer backbone evaluations (less computing power) because it only needed to run three main brains instead of five.
On the UCM and AID datasets, the new method matched or even beat the performance of the much heavier five-model team.

For example, on the UCM dataset, the new method achieved an AUROC (a score for how good the detection is) of 97.12%, while the heavy five-model team only got 94.84%. It also reduced the FPR@95 (the rate of false alarms) from 23.71% down to 14.75%.

Why This Matters

The paper suggests that you don't need to hire five expensive guides to get a smart team. You can get similar or better results by using three guides, each with five trained assistants who are encouraged to think differently when things get fuzzy. By using this "Agree-Disagree" strategy, remote sensing systems on satellites and drones can be much better at spotting weird, unknown objects without draining their batteries or slowing them down.

The researchers found that this works best when the "blur" is just right—strong enough to make the AI unsure, but not so strong that the image becomes unrecognizable. They also noted that while this method is a huge step forward, it still faces challenges when the new images are very similar to the old ones (like a forest that looks slightly different because of the season). But overall, they have shown that explicit diversity training can recover the benefits of a huge team at a fraction of the cost.

Technical Summary: Breaking Diversity Collapse in Spiking Pseudo-Ensembles for Efficient OOD Detection in Remote Sensing

Problem Statement

Spiking Neural Networks (SNNs) offer significant energy efficiency for resource-constrained remote-sensing platforms (e.g., satellites, UAVs) due to their event-driven, sparse computation. However, reliable deployment requires robust Out-of-Distribution (OOD) detection to reject unfamiliar inputs caused by shifts in geography, season, or sensing conditions. While deep ensembles provide strong predictive uncertainty by leveraging multiple independently trained models, they incur prohibitive costs for SNNs: parameter storage and inference latency scale linearly with the number of ensemble members, requiring a complete backbone evaluation for every prediction.

Existing efficient alternatives, such as pseudo-ensembles that attach multiple lightweight classification heads to a single frozen backbone, often suffer from diversity collapse. In these architectures, independently initialized heads trained with standard cross-entropy on shared frozen features converge to highly correlated predictions, failing to provide the functional diversity necessary for effective uncertainty estimation.

Methodology

The authors propose an Agree–Disagree Spiking Pseudo-Ensemble (ADPE) designed to recover ensemble-style uncertainty at a fraction of the computational cost.

1. Architecture:
The method utilizes a frozen, deterministically trained SNN backbone (either Transformer-based Spikformer or convolutional ResNet19-SNN). Multiple lightweight, independently initialized spiking classification heads are attached to this backbone. During inference, the backbone is evaluated once, and its temporal features are broadcast to all heads to generate multiple predictive hypotheses.

2. The Diversity Collapse Issue:
The paper identifies that training these heads with cross-entropy leads to diversity collapse because all heads receive identical inputs, labels, and optimization objectives, driving them toward the same decision boundary.

3. The Agree–Disagree Objective:
To induce functional diversity without external OOD data, the authors introduce a novel training objective that operates on two types of inputs derived from the in-distribution (ID) training set:

Agreement (Clean ID): Heads are trained to predict correct labels on clean images using standard cross-entropy, preserving the backbone's original classification performance.
Disagreement (Structured Blur): Heads are encouraged to produce different predictions on the same images when they are subjected to structured, uncertainty-inducing transformations. The authors identify box blur (medium to strong intensity) as the optimal transformation. Diagnostic studies show that box blur increases predictive entropy and inter-model disagreement while preserving coarse scene structure, unlike noise which introduces unstructured perturbations.

The total loss function is defined as:
$L_{AD} = L_{CE} - \lambda_{dis} D_{JS}$
Where $L_{CE}$ is the cross-entropy loss on clean samples, and $D_{JS}$ is the generalized Jensen–Shannon divergence (measuring disagreement) calculated on blur-corrupted samples. The heads are optimized to maximize this disagreement on blurred inputs while maintaining agreement on clean inputs.

4. Inference Efficiency:
With $K_b$ backbones and $K_h$ heads per backbone, the system generates $K_b \times K_h$ hypotheses using only $K_b$ backbone evaluations. This significantly reduces computational overhead compared to a full deep ensemble of $K_b \times K_h$ models.

Key Contributions

Efficient Spiking Pseudo-Ensemble: A framework that generates multiple predictive hypotheses using a frozen SNN backbone and lightweight heads, avoiding the cost of multiple full-model evaluations.
Diversity Collapse Mitigation: Identification of diversity collapse in shared-backbone pseudo-ensembles and the proposal of an agree–disagree objective. This objective promotes functional diversity using structured blur-corrupted inputs derived solely from ID data, eliminating the need for external outlier datasets or teacher ensembles.
Empirical Validation: Evaluation across two distinct SNN architectures (Spikformer, ResNet19-SNN) and multiple remote-sensing distribution shifts (UCM, AID, Sentinel-2 subsets), demonstrating improved OOD performance relative to conventional pseudo-ensembles and deep ensembles.

Results

Experiments on the EuroSAT dataset (ID) with shifts to UCM, AID, and Sentinel-2 subsets demonstrate the following:

Performance vs. Cost: An ADPE configuration with 3 backbones and 5 heads per backbone matches or exceeds the OOD detection performance (AUROC, FPR@95) of a 5-model deep ensemble on UCM and AID.
Efficiency Gains: This configuration achieves the above performance with approximately 38% fewer parameters and 40% fewer backbone evaluations compared to the 5-model deep ensemble.
Diversity Metrics: The proposed method significantly increases Mutual Information (MI) between heads compared to cross-entropy-trained pseudo-ensembles (CEPE), confirming that the heads learn distinct predictive behaviors. For instance, on Spikformer with 3 backbones, ADPE improved AUROC by ~2.0 points over CEPE on UCM.
Architecture Agnosticism: The benefits are observed in both Transformer-based and convolutional SNNs, though the improvement is more pronounced for Spikformer.
ID Performance: The method maintains classification accuracy on the ID dataset (EuroSAT) comparable to deep ensembles, indicating that encouraging disagreement on blurred inputs does not compromise clean classification.

Significance and Claims

The paper claims that explicit diversity promotion via the agree–disagree objective can recover useful ensemble-style uncertainty at substantially lower deployment costs. The authors position this as a practical solution for resource-constrained remote sensing where full deep ensembles are infeasible.

The authors remain modest regarding the limitations:

The method is an efficient mechanism for increasing useful ensemble disagreement but is not uniformly superior under every shift type or operating threshold.
Performance gains are less consistent on "semantically close" distribution shifts (Global-Near) compared to distant shifts (Global-Far).
The study highlights that while global ranking (AUROC) improves, the optimal operating point at fixed thresholds (FPR@95) can vary, suggesting that uncertainty calibration remains an open challenge for SNNs.

The work concludes that while diversity collapse is a critical failure mode in shared-backbone ensembles, it can be effectively addressed without external data, enabling efficient and reliable OOD detection for SNNs in remote sensing.