Artwork

BlueDot Impact द्वारा प्रदान की गई सामग्री. एपिसोड, ग्राफिक्स और पॉडकास्ट विवरण सहित सभी पॉडकास्ट सामग्री BlueDot Impact या उनके पॉडकास्ट प्लेटफ़ॉर्म पार्टनर द्वारा सीधे अपलोड और प्रदान की जाती है। यदि आपको लगता है कि कोई आपकी अनुमति के बिना आपके कॉपीराइट किए गए कार्य का उपयोग कर रहा है, तो आप यहां बताई गई प्रक्रिया का पालन कर सकते हैं https://hi.player.fm/legal
Player FM - पॉडकास्ट ऐप
Player FM ऐप के साथ ऑफ़लाइन जाएं!

Toy Models of Superposition

41:43
 
साझा करें
 

Manage episode 424087973 series 3498845
BlueDot Impact द्वारा प्रदान की गई सामग्री. एपिसोड, ग्राफिक्स और पॉडकास्ट विवरण सहित सभी पॉडकास्ट सामग्री BlueDot Impact या उनके पॉडकास्ट प्लेटफ़ॉर्म पार्टनर द्वारा सीधे अपलोड और प्रदान की जाती है। यदि आपको लगता है कि कोई आपकी अनुमति के बिना आपके कॉपीराइट किए गए कार्य का उपयोग कर रहा है, तो आप यहां बताई गई प्रक्रिया का पालन कर सकते हैं https://hi.player.fm/legal

It would be very convenient if the individual neurons of artificial neural networks corresponded to cleanly interpretable features of the input. For example, in an “ideal” ImageNet classifier, each neuron would fire only in the presence of a specific visual feature, such as the color red, a left-facing curve, or a dog snout. Empirically, in models we have studied, some of the neurons do cleanly map to features. But it isn't always the case that features correspond so cleanly to neurons, especially in large language models where it actually seems rare for neurons to correspond to clean features. This brings up many questions. Why is it that neurons sometimes align with features and sometimes don't? Why do some models and tasks have many of these clean neurons, while they're vanishingly rare in others?

In this paper, we use toy models — small ReLU networks trained on synthetic data with sparse input features — to investigate how and when models represent more features than they have dimensions. We call this phenomenon superposition . When features are sparse, superposition allows compression beyond what a linear model would do, at the cost of "interference" that requires nonlinear filtering.

Narrated for AI Safety Fundamentals by Perrin Walker of TYPE III AUDIO.

---

A podcast by BlueDot Impact.
Learn more on the AI Safety Fundamentals website.

  continue reading

अध्यायों

1. Toy Models of Superposition (00:00:00)

2. Definitions and Motivation: Features, Directions, and Superposition (00:00:11)

3. Empirical Phenomena (00:03:59)

4. What are Features? (00:06:08)

5. Features as Directions (00:09:20)

6. Privileged vs Non-privileged Bases (00:13:06)

7. The Superposition Hypothesis (00:15:38)

8. Summary: A Hierarchy of Feature Properties (00:20:08)

9. Demonstrating Superposition (00:21:45)

10. Experiment Setup (00:22:25)

11. Basic Results (00:29:40)

12. Mathematical Understanding (00:35:44)

83 एपिसोडस

Artwork
iconसाझा करें
 
Manage episode 424087973 series 3498845
BlueDot Impact द्वारा प्रदान की गई सामग्री. एपिसोड, ग्राफिक्स और पॉडकास्ट विवरण सहित सभी पॉडकास्ट सामग्री BlueDot Impact या उनके पॉडकास्ट प्लेटफ़ॉर्म पार्टनर द्वारा सीधे अपलोड और प्रदान की जाती है। यदि आपको लगता है कि कोई आपकी अनुमति के बिना आपके कॉपीराइट किए गए कार्य का उपयोग कर रहा है, तो आप यहां बताई गई प्रक्रिया का पालन कर सकते हैं https://hi.player.fm/legal

It would be very convenient if the individual neurons of artificial neural networks corresponded to cleanly interpretable features of the input. For example, in an “ideal” ImageNet classifier, each neuron would fire only in the presence of a specific visual feature, such as the color red, a left-facing curve, or a dog snout. Empirically, in models we have studied, some of the neurons do cleanly map to features. But it isn't always the case that features correspond so cleanly to neurons, especially in large language models where it actually seems rare for neurons to correspond to clean features. This brings up many questions. Why is it that neurons sometimes align with features and sometimes don't? Why do some models and tasks have many of these clean neurons, while they're vanishingly rare in others?

In this paper, we use toy models — small ReLU networks trained on synthetic data with sparse input features — to investigate how and when models represent more features than they have dimensions. We call this phenomenon superposition . When features are sparse, superposition allows compression beyond what a linear model would do, at the cost of "interference" that requires nonlinear filtering.

Narrated for AI Safety Fundamentals by Perrin Walker of TYPE III AUDIO.

---

A podcast by BlueDot Impact.
Learn more on the AI Safety Fundamentals website.

  continue reading

अध्यायों

1. Toy Models of Superposition (00:00:00)

2. Definitions and Motivation: Features, Directions, and Superposition (00:00:11)

3. Empirical Phenomena (00:03:59)

4. What are Features? (00:06:08)

5. Features as Directions (00:09:20)

6. Privileged vs Non-privileged Bases (00:13:06)

7. The Superposition Hypothesis (00:15:38)

8. Summary: A Hierarchy of Feature Properties (00:20:08)

9. Demonstrating Superposition (00:21:45)

10. Experiment Setup (00:22:25)

11. Basic Results (00:29:40)

12. Mathematical Understanding (00:35:44)

83 एपिसोडस

Tất cả các tập

×
 
Loading …

प्लेयर एफएम में आपका स्वागत है!

प्लेयर एफएम वेब को स्कैन कर रहा है उच्च गुणवत्ता वाले पॉडकास्ट आप के आनंद लेंने के लिए अभी। यह सबसे अच्छा पॉडकास्ट एप्प है और यह Android, iPhone और वेब पर काम करता है। उपकरणों में सदस्यता को सिंक करने के लिए साइनअप करें।

 

त्वरित संदर्भ मार्गदर्शिका