ML模型過於自信。新方法教導它們說「我不知道」", "excerpt": "深度證據回歸讓神經網路能夠在一次前向傳遞中表達不確定性，解決了生產環境AI系統的關鍵缺陷。", "body": "研究人員正在解決機器學習中的一個基本問題：模型在不應該自信的時候表現得很自信。由Amini等人在2020年提出的深度證據回歸(DER)，透過使神經網路能夠在單次前向傳遞中表達認識不確定性（它們不知道的）和隨機不確定性（資料中固有的隨機性），在不確定性量化方面取得了重大進展。與需要多次模型運行或大量計算開銷的傳統方法不同，DER修改損失函數以直接學習不確定性估計和預測。 這很重要，因為過度自信的AI在生產環境中隨處可見。當醫學影像模型將白色狗分類為貓，僅僅因為它只在黑色狗和白色貓上訓練過，這不只是一個可愛的失敗——這是安全關鍵系統的根本性故障。當前的softmax輸出偽裝成信賴度分數，但沒有提供可靠的機制來標記分佈外輸入。隨著AI在自動駕駛汽車、醫學診斷和金融系統中的部署加速，不確定性量化成為必要的基礎設施，而不是學術好奇心。 生物醫學影像領域的並行工作表明這不僅僅是理論擔憂。對頻譜歸一化神經高斯過程(SNGP)的研究表明，輕量級修改——頻譜歸一化加上高斯過程層——顯著改善了六個醫學資料集上的不確定性估計和分佈外檢測。這些方法有一個共同點：從樸素的信賴度分數轉向有原則的不確定性量化，能夠真正為現實世界的決策提供資訊。 對開發者來說，這代表了我們構建AI系統方式的實際轉變。我們需要能夠標記不確定預測供人類審查的架構，而不是部署自信猜測的模型。像DER和SNGP這樣的方法的計算開銷與Monte Carlo方法相比是最小的，使得不確定性量化在每毫秒都很重要的生產環境中變得可行。

Researchers are tackling a fundamental problem in machine learning: models that act confident when they shouldn't be. Deep Evidential Regression (DER), introduced by Amini et al. in 2020, represents a significant advance in uncertainty quantification by enabling neural networks to express both epistemic uncertainty (what they don't know) and aleatoric uncertainty (inherent randomness in data) in a single forward pass. Unlike traditional approaches that require multiple model runs or extensive computational overhead, DER modifies the loss function to directly learn uncertainty estimates alongside predictions.

This matters because overconfident AI is everywhere in production. When a medical imaging model classifies a white dog as a cat because it only trained on black dogs and white cats, that's not just a cute failure—it's a fundamental breakdown in safety-critical systems. Current softmax outputs masquerade as confidence scores but offer no reliable mechanism to flag out-of-distribution inputs. As AI deployment accelerates in autonomous vehicles, medical diagnosis, and financial systems, uncertainty quantification becomes essential infrastructure, not academic curiosity.

Parallel work in biomedical imaging shows this isn't just theoretical concern. Research on Spectral-normalized Neural Gaussian Processes (SNGP) demonstrates that lightweight modifications—spectral normalization plus a Gaussian process layer—significantly improve uncertainty estimation and out-of-distribution detection across six medical datasets. These approaches share a common thread: moving beyond naive confidence scores toward principled uncertainty quantification that can actually inform real-world decisions.

For developers, this represents a practical shift in how we build AI systems. Instead of deploying models that confidently guess, we need architectures that can flag uncertain predictions for human review. The computational overhead of methods like DER and SNGP is minimal compared to Monte Carlo approaches, making uncertainty quantification feasible in production environments where every millisecond counts.