Oxford How do positional encodings generalise to edges? We can increase expressivity by encoding positional information into graph edges Can GNNs tell these apart yet? Modern GNNs have overcome many historic limitations Defending against finetuning attacks Improving Variational Continual Learning Two empirical improvements which improve accuracy and 50x convergence speed of VCL Cambridge Architectural backdoors Backdoors embedded in the architecture of neural networks are resistant to full retraining. Uncertainty estimation for spiking neural networks A distribution of outputs from neural networks with discrete activations Sequence classification expressivity Comparing expressivity to learnability Hardware locking What if you really didn't want anyone else to run your model? Youtube Wiki Referencer Fact-check YouTube videos as they play with facts from Wikipedia and Wikidata, displaying them next to the video.
