On the Turing Completeness and Efficiency of Prompting in Finite-Sized Transformers
This research paper theoretically proves that prompting a single, fixed-size Transformer can be Turing complete, meaning it can theoretically compute any computable function, achieving near-optimal computational efficiency comparable to the entire class of unbounded-size Transformers.
基於典型性感知學習的失效檢測
深度神經網路的過度自信問題源於模型過度擬合非典型樣本，而典型性感知學習通過區分典型和非典型樣本的優化方式，可以有效減輕過度自信問題，提高失效檢測性能。
Typicalness-Aware Learning for Improved Failure Detection in Deep Neural Networks
Overfitting on atypical samples with ambiguous content can lead to overconfidence in deep neural networks, hindering failure detection. Typicalness-Aware Learning (TAL) addresses this by dynamically adjusting the optimization of typical and atypical samples, improving the reliability of confidence scores and failure detection.
評估分類器預測可靠性的新指標：確定性比率 Cρ
傳統的分類器性能指標，例如準確率，往往忽略了預測中的不確定性，而確定性比率 (Cρ) 作為一個新的指標，通過區分基於確定性預測和不確定性預測的性能，可以更全面地評估分類器可靠性。
분류기 예측의 신뢰도 평가를 위한 새로운 지표: 확실성 비율(Cρ)
분류기의 정확도가 높더라도 예측의 불확실성이 높으면 신뢰도가 떨어질 수 있으며, 이를 정량화하기 위해 확실성 비율(Cρ)이라는 새로운 지표를 제시한다.
分類器予測の信頼性評価のための新しい指標：確信度比率 Cρ
分類器の予測の信頼性を評価する際に、従来の精度などの指標は、予測に内在する不確実性を十分に考慮していないため、誤解を招く可能性がある。本稿では、確信度と不確実性の両方を考慮した新しい指標「確信度比率（Cρ）」を導入し、分類器の信頼性をより包括的に評価する。
The Certainty Ratio (Cρ): A New Metric for Evaluating the Reliability of Probabilistic Classifier Predictions
Traditional classifier performance metrics like accuracy can be misleading, as they don't account for uncertainty in predictions. The Certainty Ratio (Cρ), based on a novel Probabilistic Confusion Matrix, addresses this by quantifying the contribution of confident predictions to overall performance, offering a more reliable assessment of classifier trustworthiness.
맥락 내 강화 학습을 위한 N-그램 유도 헤드: 안정성 향상 및 데이터 요구 사항 감소
트랜스포머에 N-그램 유도 헤드를 통합하면 맥락 내 강화 학습의 안정성을 향상시키고 일반화에 필요한 데이터를 줄일 수 있습니다.
N-Gram Induction Heads Improve Stability and Reduce Data Needs in In-Context Reinforcement Learning
Integrating n-gram induction heads into transformers for in-context reinforcement learning significantly improves stability, reduces data requirements, and enhances performance compared to traditional methods like Algorithm Distillation.
ChatGPTを用いた医学論文の質の評価：権威ある雑誌における限界と今後の展望
ChatGPTは医学論文の質を評価するツールとしてある程度の有効性を示すが、特に権威ある医学雑誌に掲載された論文については、その質を過小評価する傾向がある。
© 2024 by Linnk AI