Fig. 4: Capacity of point-clouds manifolds of ImageNet classes.
From: Separability and geometry of object manifolds in deep neural networks
a, b Normalized classification capacity for point-cloud manifolds of ImageNet classes (full line: full class manifolds; dashed line: top 10% manifolds) along the layers of AlexNet (a) and VGG-16 (b). AlexNet top 10% manifolds results already appeared as “after training” results from Fig. 3e. Line and markers indicate mean value over 5 different choices of 50 objects; surrounding shaded areas indicate 95% confidence interval. Capacity is normalized by \({\alpha }_{{\mathrm{c}}}=2/{<{M}_{\mu }> }_{\mu }\), the value expected for unstructured manifolds (see main text; Mμ denotes the number of samples from object μ). The x-axis labels provides abbreviation of the layer types. Marker shape represents layer type (circle—pixel layer, square—convolution layer, right-triangle—max-pooling layer, hexagon—fully connected layer). Features in linear layers are extracted after a ReLU nonlinearity. Color (blue—AlexNet, green—VGG-16) changes from dark to light along the network.
