This seminar will begin at 4:00 pm at the Jerome L. Greene Science Center, room L7-119. Light refreshments will be available starting at 3:30 pm.

This month's speakers:

Amy Norovich (Bendesky lab): "Probing the neural basis of visually-evoked aggression in Siamese fighting fish"

Aggression is a universal social behavior that shapes human and animal societies. In primates, visual information plays a prominent role in eliciting aggression, yet in popular rodent models, antagonistic behavior is evoked largely by chemosensory stimuli.  Siamese fighting fish (Betta splendens) have been bred for hundreds of generations to select for robust aggressive behavior that is dependent on visual cues, making it an ideal model for studying the neurobiological basis of visually-evoked aggression. We are developing behavioral assays to reliably elicit and quantify aggressive display in Betta, and are using activity-dependent mapping and viral tracing approaches to uncover how visual information contributes to aggressive behavior.

Tahereh Toosi (Issa lab): "Generative feedback network as an alternative learning algorithm for deep neural networks"

Training of neural networks is typically done by passing errors through a separate, symmetric network that uses the transposed weights (classic error backpropagation). At another extreme from symmetric weights, the backpropagating network can have random weights with the forward network aligning its weights to some degree during training (feedback alignment). Here, we explore a third family of algorithms that learn a generative, inverse network. While this may seem to introduce more challenges since the inverse network has to be learned, we utilize a method we term gradient sharing which passes errors through one network branch to generate the gradients for training the other branch of the network (i.e. generator trains discriminator, and discriminator trains generator), all without using weight transport in either pathway. In preliminary testing with the MNIST digits set, the generative feedback network simultaneously learns to both discriminate and reconstruct digits, and we compare the resulting discrimination performance to standard learning algorithm benchmarks, classic error backpropagation and feedback alignment.

This seminar is part of the Zuckerman Institute Postdoctoral Seminar series. For questions about this or future seminars, please contact series organizers Chris Rodgers, PhD, or Amy Norovich, PhD.