Interpreting Deep Neural Networks using Cognitive Psychology DeepMind

We tried this experiment with our deep networks (Matching Networks and an Inception baseline model) and found that - like humans - our networks have a strong bias towards object shape rather than colour or texture. In other words, they have a'shape bias'. This suggests that Matching Networks and the Inception classifier use an inductive bias for shape to eliminate incorrect hypotheses, giving us a clear insight into how these networks solve the one-shot word learning problem. We observed that the shape bias emerges gradually over the course of early training in our networks. This is reminiscent of the emergence of shape bias in humans: young children show smaller shape bias than older children, and adults show the largest bias (2).