Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster B in Poster Session B: Tuesday, August 4, 2:00 – 3:45 pm, Kimmel Center, Shorin & Rosenthal Rooms

How do RNNs encode mixtures? One mechanism for three geometries

Esther Poniatowski¹, Claire Sergent¹; ¹Université Paris Cité

Presenter: Esther Poniatowski

Artificial and biological networks encode inputs as activity patterns distributed across many units. Naturalistic inputs are often superimpositions of several simpler components, as overlapping gratings in vision, mixed voices in audition, odor blends in olfaction. For such mixtures, neurophysiology reports three response geometries: either both component responses are combined in proportion to their input strength, or one is favored over the other, or a new pattern emerges that neither component alone elicits (Busse et al., 2009; Rust et al., 2006; Zoccolan et al., 2005). Which circuit properties determine which encoding arises? The divisive normalization model (Carandini & Heeger, 2012) captures the phenomenology of the first two geometries, but does not identify the underlying network mechanism, and cannot express emergence. This theoretical study analyzes a recurrent-network model and derives, analytically, the circuit property that selects each geometry. The central result is that a single network can produce all three geometries as the input varies, with no change in synaptic weights. This flexibility has computational implications for downstream readouts.

Topic Area: Methods, Tools, Theory & Neural Coding