Paper: arxiv.org/abs/2405.17283
Code: github.com/agopal42/syncx
Joint work with Aleksandar Stanic, Jürgen Schmidhuber and Michael Mozer.
Hope to see you all at our poster at #NeurIPS2024! 10/x

Phase synchronization in SynCx towards objects is more robust compared to baselines. It can successfully separate similarly colored objects, which is a common failure mode of other synchrony models that simply rely on color as a shortcut feature for grouping. 9/x

SynCx outperforms current state-of-the-art unsupervised synchrony-based models on standard multi-object datasets while using between 6-23x fewer model parameters compared to the baseline models. 8/x

Our model does not need additional inductive biases (gating mechanisms), strong supervision (depth masks) and/or contrastive training as used by current state-of-the-art synchrony models to achieve phase synchronization towards objects in a fully unsupervised way. 7/x

SynCx processes complex-valued inputs at every layer using complex-valued weights. It is trained to reconstruct the input image at every iteration using the output magnitudes. Output phases are fed back as input to the next step with input magnitudes clamped to the image. 6/x

Hidden units in such a system must activate based on the presence of features (magnitudes) but also consider their relative phases. Matrix-vector products between complex-valued weights and complex-valued activations are a natural way to implement such functionality. 5/x

This is a conceptual flaw in current synchrony models all of which use feedforward convolutional nets but we can solve this in an iterative fashion. Starting with random phases, hidden units compute phase updates to propagate local constraints to reach a stable configuration. 4/x

Green and red circles highlight junctions that belong to the same and different objects respectively. Here, we cannot decide which junctions belong to which object using only the local features as they are indistinguishable from one another in the two cases. 3/x

We argue for the importance of iterative computation (recurrence) and complex-valued weights to achieve phase synchronization in activations. To build some intuition look at the 3 shapes (T, H, and overlapping squares) made of horizontal and vertical bars. 2/x