New Paper by Dr. Juliana Londoño Alvarez

Dr. Juliana Londoño Alvarez has a wonderful new paper on a type of recurrent neural network. A member of the 2024 StemJazz Provost Cohort, her research interests include how we can use the way that the architecture of a neural network shapes its dynamics to develop more efficient models of brain functions like locomotion and complex movement sequences, and translating theoretical research to further bridge the gap between life sciences and mathematics. Below is a bit of Juliana’s abstract!

Brain rhythms that produce patterns like walking, breathing, or swimming are traditionally modeled using coupled oscillators—networks of intrinsically oscillating (“pacemaker”) neurons. In an alternative approach, attractor networks—a framework in which cognitive processes are modeled as attractors of a dynamical system, typically implemented by a recurrent network—have been a central tool in theoretical neuroscience for modeling memory and pattern completion.

In this work, we show that attractor networks can also support rhythmic pattern generation. We present a single attractor network that can robustly encode five distinct quadruped gaits—walk, trot, pace, bound, and gallop—as coexisting attractors, without requiring parameter changes. Transitions between gaits can be triggered by simple external pulses. This contrasts with most existing locomotion models, which rely on finely tuned coupled oscillators whose parameters must be adjusted to switch between gaits. In addition, we introduce a novel method for encoding ordered sequences of gaits using fusion attractors, which allows the network to flexibly reuse existing patterns in different combinations (as in a sequence of dance moves).

Attractor networks have long been a central paradigm in theoretical and computational neuroscience, and to our knowledge, no previous model has simultaneously encoded such a diverse set of rhythmic patterns in a single network without changing parameters. Our results suggest that attractor networks can provide a unified framework for modeling brain processes. Beyond theoretical neuroscience, this work may also have applications in neuro-inspired robotics.

Congrats Juliana!