Unverified50% confidenceFactExact time
By improving feature visualization techniques, the proportion of readable neurons in the last layer of Inception V1 rose from about 0-5% to about 50%.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/17/2026
First Seen
Valid until: 1/15/2027
Sources
Neural Networks Were Always Clean? Goodfire Researcher Overturns Superposition Theory
bilibili迦太基盐商10/16/2026
Related Entities
Related Claims
UnverifiedNick Cammarata's analysis points out that InceptionV1's 4e:55 neuron actually detects a V-shaped geometric form, rather than cat faces or car fronts.68% similarUnverified初级视觉皮层V1位于枕叶后部，包含约1.4亿个神经元65% similarUnverifiedThe tool's graphical interface displays in real time the loss changes per mini-batch and epoch, the gradient norm of each layer, the percentage of dead neurons, weight distribution comparisons, and first-layer receptive field visualizations.64% similarUnverifiedIn the modern vision models Cammarata tested, 99%–100% of the neurons in the last layer could be given fairly clear explanations.63% similarUnverifiedOlshausen和Field在1996年的工作表明，V1神经元的感受野与ICA在自然图像上学到的基函数高度吻合61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1052465API
curl https://kongchang.com/api/v1/knowledge/claims/1052465MCP
get_claim(id=1052465)