keyword
SplitBrain autoencoders
Split-brain autoencoders are neural network architectures designed for self-supervised representation learning by performing cross-channel prediction instead of full input reconstruction. In a standard autoencoder, the model learns by reconstructing the complete input signal, which can risk learning trivial identity mappings. A split-brain autoencoder instead divides the input data across its channels or modalities into disjoint subsets, such as separating luminance from color channels or visual appearance from depth. The network is divided into separate subnetworks, where each subnetwork receives one subset of the channels as input and is trained to predict the complementary subset. By compelling the subnetworks to solve these cross-channel estimation tasks, the model learns high-level semantic and invariant feature representations across the full input signal, which can be aggregated and transferred effectively to downstream tasks without requiring manual supervision.
1 item
