I am a first-generation graduate PhD candidate in Electrical and Computer Engineering at Rice University studying under David Lipshutz. I am passionate about deep learning theory, neuroAI, recurrent neural networks, and generative models.
Recently I was a NeuroAI Research Intern at Cold Spring Harbor Laboratory under Mitra Javadzadeh.

The images on the left were made using my Boomerang local sampling algorithm with the prompt "courtroom sketch". Try it yourself, it's fun!

Contact: jc135 [at] rice [dot] edu

LinkedIn / GitHub / Google Scholar / CV

Research

Leakage and Second-Order Dynamics Improve Hippocampal RNN Replay
Josue Casco-Rodriguez, Nanda Krishna, Richard Baraniuk, and David Lipshutz
Under review

We re-examine RNN replay as Langevin sampling to understand or improve it in three ways: (1) The gradients replay activity should follow are time-varying and difficult to estimate, but hidden state leakage is helpful. (2) Hidden state adaptation (negative feedback) encourages exploration, but also slows replay. (3) Temporally compressed replay in RNNs is possible through hidden state momentum, which relates to underdamped Langevin sampling and short-term facilitation, and counters slowness while maintaining exploration.

An Analytical Framework for Multi-Area Balanced Cortical Networks
Josue Casco-Rodriguez and Mitra Javadzadeh
Under review

We analyze balanced multi-area cortical networks, and see: (1) intra-area local connectivity determines its responses to local inputs, but not to inputs relayed from other areas; (2) local responses to these relayed inputs are instead driven by long-range inter-area connections; and (3) the asymmetry of inter-area connections (feedforward vs feedback) affects implements a tradeoff between regimes that promote inter-area similarity or divergence.

On the Geometry of Recurrent Spiking Networks
Josue Casco-Rodriguez
Under review

We propose to study the geometry of multi-layer recurrent SNNs as piecewise-constant functions, which we visualize with our new algorithm, SplineCam-SNN. By looking at how various SNN parameters affect their input-output geometry, our findings emphasize the importance of skip connections, leakages, and delays for training SNNs, and suggest that these parameters may be promising targets for future plasticity algorithms.

[Re] The Discriminative Kalman Filter for Bayesian Filtering with Nonlinear and Nongaussian Observation Models
Josue Casco-Rodriguez, Caleb Kemere, Richard Baraniuk
ReScience C 2025

Burkhart et al. (2020) made a new Kalman filter that leverages Bayes' theorem to improve filter performance for highly non-linear or non-Gaussian observation models. We provide an open-source Python version, reproduce their most salient results for neuroscientific contexts, and further examine the efficacy of their filter using multiple random seeds and previously unused trials from the authors' dataset.

RNN Replay: Leakage and Underdamped Dynamics
Josue Casco-Rodriguez, Richard Baraniuk
NeurIPS 2024 NeuroAI Workshop

Denoising RNNs can generate activity (replay) via Langevin sampling. We build upon this theory of replay by: (1) using Gaussian processes to explain how denoising is challenging, but benefits from leakage; (2) showing how adaptation affects sampling; and (3) proposing a more direct form of underdamped Langevin sampling for RNNs.

Visualizing Linear RNNs Through Unrolling
Josue Casco-Rodriguez, Tyler Burley, CJ Barberan, Ahmed Imtiaz Humayun, Randall Balestriero, Richard Baraniuk
NeurIPS 2024 LatinX in AI Workshop

We apply SplineCam to linearly-recurrent RNNs (i.e., SSMs) trained on biosignal data and sequential MNIST. This enables: (1) unprecedented visualization of the decision boundary and complexity of an RNN, and (2) visualization of the frequency sensitivity of RNNs around individual data points.

Self-Consuming Generative Models Go MAD
Sina Alemohammad*, Josue Casco-Rodriguez*, Lorenzo Luzi, Ahmed Imtiaz Humayun, Hossein Babaei, Daniel LeJeune, Ali Siahkoohi, Richard Baraniuk
ICLR 2024

We study the phenomenon of training new generative models with synthetic data from previous generative models. Our primary conclusion is that without enough fresh real data in each generation of a self-consuming or autophagous loop, future generative models are doomed to have their quality (precision) or diversity (recall) progressively decrease.

Boomerang: Local Sampling on Image Manifolds Using Diffusion Models
Lorenzo Luzi, Paul Mayer, Josue Casco-Rodriguez, Ali Siahkoohi, Richard Baraniuk
TMLR 2024

We propose Boomerang, a method to do local sampling on an image manifold parameterized by a pretrained diffusion model. Boomerang is useful in three settings: generating privacy preserving / anonymized datasets, perceptual image enhancement of low resolution images, and data augmentation for supervised classification tasks.

Provable Instance Specific Robustness via Linear Constraints
Ahmed Imtiaz Humayun*, Josue Casco-Rodriguez*, Randall Balestriero, Richard Baraniuk
ICML 2023 AdvML Workshop

Using spline theory, we present a novel method for imposing analytical constraints directly on the decision boundary for provable robustness. Our method can provably ensure robustness for any set of instances, e.g. training samples from a specific class, against adversarial, backdoor or poisoning attack.

A comprehensive review of spiking neural networks: Interpretation, optimization, efficiency, and best practices.
Kai Malcolm*, Josue Casco-Rodriguez*
arXiv 2023

We present a literature review of recent developments in the interpretation, optimization, efficiency, and accuracy of spiking neural networks. Key contributions include identification, discussion, and comparison of cutting-edge methods in spiking neural network optimization, energy-efficiency, and evaluation, starting from first principles so as to be accessible to new practitioners.

Page updated

Report abuse