Inception V1
An early convolutional neural network architecture commonly used in computer vision research and once a model organism in the field of interpretability.
Timeline (last 90 days)
Inception V1 (GoogLeNet) was the winning network of the 2014 ImageNet competition, introduced the Inception module and 1×1 convolution dimensionality reduction, and is about 22 layers deep.
After turning off this neuron, the response of the "flames and fox faces" neuron in the next layer dropped by about 15% on cat faces and fox faces, but among the 75 photos tested, the network's final answer did not change for a single one.
Running the same pointy-versus-round comparison across all 832 neurons in that layer, about 100 neurons prefer pointy things more than the target neuron does, showing that this preference is not rare.
Reproduction experiments show that, on unoptimized inputs, this neuron's activation values are on average higher for pointy shapes (such as V-shaped arrows, triangles, high-speed train noses, and long-snouted dogs) than for rounded shapes (such as rings, discs, clocks, oranges, and flat-faced bulldogs), with pointy shapes winning about 84 times out of 100.
Camerata points out that the car features hidden in the dog neuron are not due to compositional readout but because cars and dogs share poses; there are hundreds of such pose neurons in Inception V1.
Camerata re-examined neuron 4E55 in Inception V1 and concluded that it is a V-shape neuron rather than a cat-car neuron; its strongest activations correspond to V-shaped flames, V-shaped fox faces, and so on, while its negative poles correspond to the round O.
Inception V1 is a classic convolutional network proposed by Google in 2014 and, due to its breakthrough on ImageNet and its compact structure, has long been the "fruit fly" model for interpretability research; its conv5-1 layer has 480 channels and was systematically studied by Olah et al. using feature visualization.
In the Inception V1 model, the uses of about 0-5% of the neurons in the last layer can be understood, which is one of the main pieces of evidence for superposition theory.
When a layer of Inception V1 was cut from 480 neurons to 10 and retrained, the network did not exhibit superposition; instead, neuron functions became coarser and more generalized.
After retraining Inception V1 with its original architecture and adding modern regularization, about 70% of the neurons in the last layer can be interpreted.
2 more timeline events
All Facts (12)
Inception V1 (GoogLeNet) was the winning network of the 2014 ImageNet competition, introduced the Inception module and 1×1 convolution dimensionality reduction, and is about 22 layers deep.
50%UnverifiedAfter turning off this neuron, the response of the "flames and fox faces" neuron in the next layer dropped by about 15% on cat faces and fox faces, but among the 75 photos tested, the network's final answer did not change for a single one.
50%UnverifiedRunning the same pointy-versus-round comparison across all 832 neurons in that layer, about 100 neurons prefer pointy things more than the target neuron does, showing that this preference is not rare.
50%UnverifiedReproduction experiments show that, on unoptimized inputs, this neuron's activation values are on average higher for pointy shapes (such as V-shaped arrows, triangles, high-speed train noses, and long-snouted dogs) than for rounded shapes (such as rings, discs, clocks, oranges, and flat-faced bulldogs), with pointy shapes winning about 84 times out of 100.
50%UnverifiedCamerata points out that the car features hidden in the dog neuron are not due to compositional readout but because cars and dogs share poses; there are hundreds of such pose neurons in Inception V1.
50%UnverifiedCamerata re-examined neuron 4E55 in Inception V1 and concluded that it is a V-shape neuron rather than a cat-car neuron; its strongest activations correspond to V-shaped flames, V-shaped fox faces, and so on, while its negative poles correspond to the round O.
50%UnverifiedInception V1 is a classic convolutional network proposed by Google in 2014 and, due to its breakthrough on ImageNet and its compact structure, has long been the "fruit fly" model for interpretability research; its conv5-1 layer has 480 channels and was systematically studied by Olah et al. using feature visualization.
50%UnverifiedIn the Inception V1 model, the uses of about 0-5% of the neurons in the last layer can be understood, which is one of the main pieces of evidence for superposition theory.
50%UnverifiedWhen a layer of Inception V1 was cut from 480 neurons to 10 and retrained, the network did not exhibit superposition; instead, neuron functions became coarser and more generalized.
50%UnverifiedAfter retraining Inception V1 with its original architecture and adding modern regularization, about 70% of the neurons in the last layer can be interpreted.
50%UnverifiedBy improving feature visualization techniques, the proportion of readable neurons in the last layer of Inception V1 rose from about 0-5% to about 50%.
50%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.
50%