The AI That Learned to Understand Long After It Stopped Trying
🇬🇧 English
In 2022, researchers trained a tiny neural network on modular addition, a simple clock-math task. The model quickly memorized training data but failed on unseen examples. After thousands of additional training steps with no visible changes, its performance on new problems jumped from near-random to nearly perfect.
This phenomenon, named grokking, shows that a model can appear finished yet later develop true understanding. The network essentially rediscovered trigonometry, representing numbers as points on a circle and using rotation to solve problems. This delayed insight challenges assumptions about when learning truly occurs.
🇸🇦 العربية
الذكاء الاصطناعي يتعلم الفهم بعد وقت طويل من توقف التدريب
In 2022, researchers trained a tiny neural network on modular addition, a simple clock-math task. The model quickly memorized training data but failed on unseen examples. After thousands of additional training steps with no visible changes, its performance on new problems jumped from near-random to nearly perfect.
This phenomenon, named grokking, shows that a model can appear finished yet later develop true understanding. The network essentially rediscovered trigonometry, representing numbers as points on a circle and using rotation to solve problems. This delayed insight challenges assumptions about when learning truly occurs.
ما هو الغروكينغ في التعلم الآلي؟
الغروكينغ هو ظاهرة حيث تحقق شبكة عصبية فجأة التعميم—حل المشكلات غير المرئية بشكل صحيح—بعد فترة طويلة من تحقيقها لأداء مثالي على بيانات التدريب. يبدو أنها ‘تنقر’ إلى الفهم دون بيانات جديدة أو تغييرات واضحة في التدريب.
🇧🇩 বাংলা
AI প্রশিক্ষণ বন্ধ করার অনেক পরে বোঝার শেখে
In 2022, researchers trained a tiny neural network on modular addition, a simple clock-math task. The model quickly memorized training data but failed on unseen examples. After thousands of additional training steps with no visible changes, its performance on new problems jumped from near-random to nearly perfect.
This phenomenon, named grokking, shows that a model can appear finished yet later develop true understanding. The network essentially rediscovered trigonometry, representing numbers as points on a circle and using rotation to solve problems. This delayed insight challenges assumptions about when learning truly occurs.
মেশিন লার্নিং में ग्रोकिंग क्या है?
গ্রোকিং হল একটি fenóমেন যেখানে একটি নিউরাল নেটওয়ার্ক আকস্মিকভাবে সাধারণীকরণ অর্জন করে—প্রশিক্ষণ ডেটায় পূর্ণ কর্মক্ষমতা অর্জনের অনেক পরেও অদৃশ্য সমস্যা সঠিকভাবে সমাধান করতে সক্ষম হয়। এটি নতুন ডেটা বা প্রশিক্ষনে evidente পরিবর্তন ছাড়াই ‘ক্লিক’ করে বোঝার দিকে যায়।
🇩🇪 Deutsch
KI lernt erst lange nach dem Training zu verstehen
In 2022, researchers trained a tiny neural network on modular addition, a simple clock-math task. The model quickly memorized training data but failed on unseen examples. After thousands of additional training steps with no visible changes, its performance on new problems jumped from near-random to nearly perfect.
This phenomenon, named grokking, shows that a model can appear finished yet later develop true understanding. The network essentially rediscovered trigonometry, representing numbers as points on a circle and using rotation to solve problems. This delayed insight challenges assumptions about when learning truly occurs.
Was ist Grokking im Maschinellen Lernen?
Grokking ist ein Phänomen, bei dem ein neuronales Netzwerk plötzlich Generalisierung erreicht — also unbekannte Probleme korrekt löst — lange nachdem es bereits eine perfekte Leistung auf den Trainingsdaten gezeigt hat. Es scheint „anzuklicken“ in das Verständnis, ohne neue Daten oder offensichtliche Änderungen im Training.
🇪🇸 Español
La IA aprende a comprender mucho después de que termina el entrenamiento
In 2022, researchers trained a tiny neural network on modular addition, a simple clock-math task. The model quickly memorized training data but failed on unseen examples. After thousands of additional training steps with no visible changes, its performance on new problems jumped from near-random to nearly perfect.
This phenomenon, named grokking, shows that a model can appear finished yet later develop true understanding. The network essentially rediscovered trigonometry, representing numbers as points on a circle and using rotation to solve problems. This delayed insight challenges assumptions about when learning truly occurs.
¿Qué es el grokking en el aprendizaje automático?
El grokking es un fenómeno en el que una red neuronal logra repentinamente la generalización—resolver correctamente problemas no vistos—mucho después de haber alcanzado un rendimiento perfecto en los datos de entrenamiento. Parece ‘hacer clic’ en la comprensión sin nuevos datos ni cambios obvios en el entrenamiento.
🇫🇷 Français
L'IA apprend à comprendre longtemps après la fin de l'entraînement
In 2022, researchers trained a tiny neural network on modular addition, a simple clock-math task. The model quickly memorized training data but failed on unseen examples. After thousands of additional training steps with no visible changes, its performance on new problems jumped from near-random to nearly perfect.
This phenomenon, named grokking, shows that a model can appear finished yet later develop true understanding. The network essentially rediscovered trigonometry, representing numbers as points on a circle and using rotation to solve problems. This delayed insight challenges assumptions about when learning truly occurs.
Qu'est-ce que le grokking dans l'apprentissage automatique ?
Le grokking est un phénomène où un réseau de neurones atteint soudainement la généralisation—résoudre correctement des problèmes inédits—longtemps après avoir déjà atteint des performances parfaites sur les données d'entraînement. Il semble 'cliquer' dans la compréhension sans nouvelles données ou changements évidents dans l'entraînement.
🇮🇳 हिन्दी
AI प्रशिक्षण रोकने के बहुत बाद सीखती है समझना
In 2022, researchers trained a tiny neural network on modular addition, a simple clock-math task. The model quickly memorized training data but failed on unseen examples. After thousands of additional training steps with no visible changes, its performance on new problems jumped from near-random to nearly perfect.
This phenomenon, named grokking, shows that a model can appear finished yet later develop true understanding. The network essentially rediscovered trigonometry, representing numbers as points on a circle and using rotation to solve problems. This delayed insight challenges assumptions about when learning truly occurs.
मशीन लर्निंग में ग्रोकिंग क्या है?
ग्रोकिंग एक ऐसी घटना है जिसमें एक तंत्रिका नेटवर्क अचानक सामान्यीकरण हासिल कर लेता है—प्रशिक्षण डेटा पर पूर्ण प्रदर्शन हासिल करने के बहुत बाद भी अपरिचित समस्याओं को सही ढंग से हल करने में सक्षम हो जाता है। यह नए डेटा या प्रशिक्षण में स्पष्ट बदलाव के बिना ‘क्लिक’ करके समझ में आने जैसा प्रतीत होता है।
🇮🇩 Bahasa Indonesia
AI Belum Memahami Setelah Pelatihan Berhenti Lama Sekali
In 2022, researchers trained a tiny neural network on modular addition, a simple clock-math task. The model quickly memorized training data but failed on unseen examples. After thousands of additional training steps with no visible changes, its performance on new problems jumped from near-random to nearly perfect.
This phenomenon, named grokking, shows that a model can appear finished yet later develop true understanding. The network essentially rediscovered trigonometry, representing numbers as points on a circle and using rotation to solve problems. This delayed insight challenges assumptions about when learning truly occurs.
Apa itu grokking dalam pembelajaran mesin?
Grokking adalah fenomena di mana jaraf saraf tiba-tiba mencapai generalisasi—menyelesaikan masalah yang belum pernah dilihat dengan benar—lama setelah ia telah mencapai performa sempurna pada data pelatihan. Ia terlihat ‘mengklik’ ke dalam pemahaman tanpa data baru atau perubahan yang jelas dalam pelatihan.
🇯🇵 日本語
AIは訓練が終わってからずっと後になって理解を学ぶ
In 2022, researchers trained a tiny neural network on modular addition, a simple clock-math task. The model quickly memorized training data but failed on unseen examples. After thousands of additional training steps with no visible changes, its performance on new problems jumped from near-random to nearly perfect.
This phenomenon, named grokking, shows that a model can appear finished yet later develop true understanding. The network essentially rediscovered trigonometry, representing numbers as points on a circle and using rotation to solve problems. This delayed insight challenges assumptions about when learning truly occurs.
機械学習におけるグロッキングとは何ですか？
グロッキングとは、ニューラルネットワークが訓練データで完璧なパフォーマンスを達成してから長い時間を経て、突然一般化能力を獲得し、未知の問題を正しく解くことができる現象です。新しいデータや訓練における明らかな変化なく、『クリック』して理解に到達するように見えます。
🇧🇷 Português
A IA aprende a compreender muito depois do treinamento terminar
In 2022, researchers trained a tiny neural network on modular addition, a simple clock-math task. The model quickly memorized training data but failed on unseen examples. After thousands of additional training steps with no visible changes, its performance on new problems jumped from near-random to nearly perfect.
This phenomenon, named grokking, shows that a model can appear finished yet later develop true understanding. The network essentially rediscovered trigonometry, representing numbers as points on a circle and using rotation to solve problems. This delayed insight challenges assumptions about when learning truly occurs.
O que é grokking em aprendizado de máquina?
Grokking é um fenômeno onde uma rede neural alcança subitamente a generalização—resolvendo corretamente problemas não vistos—muito tempo depois de já ter atingido desempenho perfeito nos dados de treinamento. Parece 'clicar' na compreensão sem novos dados ou mudanças óbvias no treinamento.
🇷🇺 Русский
ИИ учится понимать намного позже, чем заканчивается обучение
In 2022, researchers trained a tiny neural network on modular addition, a simple clock-math task. The model quickly memorized training data but failed on unseen examples. After thousands of additional training steps with no visible changes, its performance on new problems jumped from near-random to nearly perfect.
This phenomenon, named grokking, shows that a model can appear finished yet later develop true understanding. The network essentially rediscovered trigonometry, representing numbers as points on a circle and using rotation to solve problems. This delayed insight challenges assumptions about when learning truly occurs.
Что такое грокинг в машинном обучении?
Грокинг — это явление, при котором нейронная сеть внезапно достигает обобщения — правильно решает неизвестные задачи — долгое время после того, как уже достигла идеальной производительности на обучающих данных. Кажется, что она «щёлкает» в понимание без новых данных или очевидных изменений в обучении.
🇨🇳 简体中文
AI 在训练停止后很久才学会理解
In 2022, researchers trained a tiny neural network on modular addition, a simple clock-math task. The model quickly memorized training data but failed on unseen examples. After thousands of additional training steps with no visible changes, its performance on new problems jumped from near-random to nearly perfect.
This phenomenon, named grokking, shows that a model can appear finished yet later develop true understanding. The network essentially rediscovered trigonometry, representing numbers as points on a circle and using rotation to solve problems. This delayed insight challenges assumptions about when learning truly occurs.
什么是机器学习中的格罗克（grokking）？
格罗克是一种现象，其中神经网络在已经达到训练数据上完美性能很久之后，突然实现泛化——即能够正确解决未见过的问题。它看起来像是‘顿悟’进入理解状态，而无需新数据或训练中的明显变化。