Please welcome guest speaker and FRF candidate, Roman Pogodin, PhD Candidate in Theoretical Neuroscience at University College London, who will be presenting for our CCN Seminar.
Talk Title: On biological plausibility of weight sharing
Abstract: Weight sharing among neurons is ubiquitous in deep learning, especially in the visual domain: it reduces the number of parameters, reduces training time, and increases accuracy. In particular, it is used in convolutional networks, which are often treated as a model of the visual stream – even though weight sharing is thought to be biologically implausible.
In this talk, I will present a biologically plausible mechanism for weight sharing and apply it to locally connected networks. The mechanism uses lateral connections and an anti-Hebbian learning rule, such that neurons learn to respond similarly to similar inputs. This plasticity mechanism can work either in separate “sleep phases” that interrupt training, or during learning using task-independent noise. This method enables locally connected networks to achieve nearly convolutional performance on ImageNet, and improves their fit to the ventral stream data, thus supporting convolutional networks as a model of the visual stream.
If you wish to meet 1:1 with Roman Pogodin, then please reach out to Noah Dlugacz (email@example.com) who can coordinate on your behalf.
Feel free to convene in the 4th Floor Classroom at 160 5th Avenue. (This onsite location is subject to change so please stay tuned for updates).
Zoom credentials provided in the calendar invite for those tuning in remotely.