Projects

• Neuromodulated Recurrent Neural Networks (NeurIPS Paper)

  In this work, we augment traditional RNN models of neural computation by adding a neuromodulator-like signal which can strengthen and weaken the network’s recurrent weights. In doing so, we seek to begin bridging the gap between highly biologically-inspired biophysical models and broader RNN models of computation. Graduate work with Scott Linderman and David Zoltowski.

  NeurIPS paper: JC Costacurta, S Bhandarkar, D Zoltowski, SW Linderman. Structured flexibility in recurrent neural networks via neuromodulation. 38th Conference on Advances in Neural Information Processing Systems (NeurIPS). Vancouver, 2024.

  COSYNE abstract/poster: JC Costacurta, S Bhandarkar, D Zoltowski, SW Linderman. Structured flexibility in recurrent neural networks via neuromodulation. COSYNE Abstracts 2024.

• Warped Autoregressive Hidden Markov Models (NeurIPS Paper)

  In this work, we take the autoregressive hidden Markov model employed in MoSeq and add an additional latent “warping” variable, which modulates discrete syllables via a time-warping (T-WARHMM) or Gaussian process (GP-WARHMM) format. While we believe this model class can be generally useful, in our work it is applied to unsupervised behavioral segmentation of mouse movement. Graduate research with Scott Linderman, Lea Duncker, and Alex Williams.

  NeurIPS paper: J.C. Costacurta, L. Duncker, B. Sheffer, C. Weinreb, W. Gillis, J. Markowitz, S.R. Datta, A.H. Williams, and S.W. Linderman (2022). Distinguishing discrete and continuous behavioral variability using warped autoregressive HMMs. 36th Conference on Advances in Neural Information Processing Systems (NeurIPS). New Orleans, 2022.

  COSYNE abstract/poster: J.C. Costacurta, A.H. Williams, B. Sheffer, C. Weinreb, W. Gillis, J. Markowitz, S.R. Datta, and S.W. Linderman (2022). Time-warped state space models for distinguishing movement type and vigor. COSYNE Abstracts 2022.

• Designing Feedback Controllers for Human-Prosthetic Systems Using H-Infinity Model Matching (EMBC Paper)

  Brain-controlled (myoelectric) upper-limb prosthetic devices are advancing at a fast pace, but current methods of delivering sensory feedback to users are lagging. In particular, some external forms of feedback (force, sound, etc) have been shown to be more distracting than helpful to users. In this work, we attempt to use feedback control theory to investigate how we can improve sensory feedback experiences for amputees. To do so, we build a multi-controller closed-loop system involving the user and prosthesis. Undergraduate research with Sridevi Sarma and Luke Osborn.

  Conference paper: J. Costacurta, L. Osborn, N. Thakor, and S. Sarma, “Designing feedback controllers for human-prosthetic systems using h-infinity model matching,” in 2018 40th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 2018, pp. 2316 – 2319.

• Machine Learning Methods for Numerical Solutions of Partial Differential Equations (Paper)

  Traditional numerical methods for solving PDEs, such as finite difference or finite element method, are of limited use in problems with complex domains. We attempted to use theory from stochastic calculus to code an alternative solution method in Python. We used methods from machine learning, such as temporal difference learning and PyTorch, to enhance our method. Research conducted as part of 2019 Fields Undergraduate Summer Research Program at Fields Institute/University of Toronto. Presented slides at end-of-summer symposium. Summer undergraduate research with Adam Stinchcombe and Mihai Nica.

  Journal paper: Martin, C., Zhang, H., Costacurta, J. et al. Solving Elliptic Equations with Brownian Motion: Bias Reduction and Temporal Difference Learning. Methodol Comput Appl Probab 24, 1603–1626 (2022). https://doi.org/10.1007/s11009-021-09871-9

• Effect of Ankle-Foot Orthosis Stiffness on Muscle Coordination During Transient Walking in Healthy Adults (Poster)

  Ankle-Foot Orthoses are devices prescribed to children with cerebral palsy in order to enhance their mobility. To inform how they are tuned and prescribed, we collected data and analyzed the effects of these devices on the first six steps of motion, called gait initiation or transient walking. Research conducted as part of 2018 Center for Neurotechnology REU at University of Washington. Presented poster at end-of-summer symposium. Summer undergraduate research with Kat Steele and Michael Rosenberg.