Randomness!
Evolution as a philosophy of usefulness has a blind spot!
-
Critical junction & path dependency
← Which factors determine these?
-
Convergent evolution
- Different mechanisms for same purpose make the system alike but essentially different
⇒ Same idea for neuroscience!
-
e.g. A Morphological Basis for Path-Dependent Evolution …
- Eye vs eyespot → (same output) spatial vision
- The tendency doesn’t follow the phylogenetic tree!
- Then what determines the branching?
- The number of shell!
- Less slit (⇒ constraint of the slit number) makes retina, More slit makes eyespot
- Intermediate structure cannot make visual sensation system
- Neural visual map
- Columnar vs salt-and-pepper
- What makes the difference? (No tree again)
- Retina & V1 size!
- Nyquist frequency determines. (Is V1 larger than retina? ⇒ (yes) columnar)
- Retina & V1 size!
- What makes the difference? (No tree again)
- Columnar vs salt-and-pepper
- Eye vs eyespot → (same output) spatial vision
-
Number neurons emerge also in randomly initialized networks!
-
Ratio and differences also!
-
Zero-shot Counting with a Dual-Stream Neural Network Model
-
Confidence miscalibration
- Manifold 상에서 decision boundary 양옆으로 ‘불확실한’ 데이터를 일부러 만들어서 예를 들어 0.8 이렇게 대답하도록 학습하는 논문을 봤던 것 같은데
더닝크루거 효과?ㅋㅋㅋ