Dekel Galor
Hello! I am a second-year PhD student at UC Berkeley interested in signal processing, computational imaging, and visual neuroscience. I am a part of Prof. Jacob Yates' Active Vision and Neural Computation Lab, and Prof. Laura Waller's Computational Imaging Lab.
I am grateful for the support of both:
The NSF Graduate Fellowship
The Center for Innovation in Vision and Optics (CIVO) Fellowship
in sponsoring my research and studies.
Recent Projects
Hadi Vafaii*, Dekel Galor*, Jacob L. Yates
The Evidence Lower Bound (ELBO) and identically the free energy principle hint at a unifying framework for machine learning and neuroscience.
Despite its utility, ELBO is often seen as too broad to offer prescriptive guidance for specific architectures in neuroscience or machine learning.
We show that maximizing ELBO under Poisson assumptions for general sequence data leads to a spiking neural network that performs Bayesian posterior inference through its membrane potential dynamics.
The resulting model, the iterative Poisson VAE (iP-VAE), has a closer connection to biological neurons.
iP-VAE also learns sparser representations and exhibits superior generalization.
Accepted as Spotlight for NeurIPS 2024
Hadi Vafaii, Dekel Galor, Jacob L. Yates
Introduce an algorithm for differentiating through Poisson sampling.
Derive generative modeling theory for a variational autoencoder with Poisson approximate posterior.
Unify neuroscience theories of rate coding, predictive coding, efficient coding, sparse coding, neural sampling, etc.
Implement PVAE and demonstrate that it has superior latent representations that are sparse, positive integer, and are more informative for tasks like classification.
SFN 2023 Poster
Dekel Galor, Jude F. Mitchell, Daniel A. Butts, Jacob L. Yates
Uncovering the role of eye movements as a primary driver of neuron activity in the visual cortex.
Demonstrate simple mechanisms that explain this activity only using a feed-forward model! (No recurrence/external motor signals.)
Propose a novel activation function SplitReLU, and show that CNNs using it exhibit emergent physiological selectivity throughout.
Break the traditionally used "upper bound" of explainable variance, showing its likely that fixational eye movements significantly drive activity in the visual cortex.
Show that large ResNets can match more interpretable models with little engineering, and compare between the two using interpretable ML.
State-of-the-art neural data from the primary visual cortex around the center of gaze (!) with spatiotemporal modeling at 240hz (!), during natural free-viewing (!), and using extremely high precision (sub-cone resolution!) DDPI eyetracking.
ICCP 2023 Spotlight Poster & Demo
Dekel Galor*, Ruiming Cao*, Jacob L. Yates, Laura Waller
Reconstruct static scenes from noise statistics in event-based cameras
Works in post processing, requires no hardware modifications
Applications in
adaptive event denoising
simultaneous denoising & scene estimation
Optica Special Issue on Displays
Dekel Galor*, Guanghan Meng*, Laura Waller, Martin Banks
Applying tools from signal processing and vision science to improve the display design process for VR.
Created a novel formula for a human visual model (CSF formula).
Created a GPU-accelerated software toolkit for simulating human perception during display design.
Available as a GPU accelerated serverless website, or open-source executable.
Dekel Galor* and Ryan Mei*
Video compression, transmission over rf, and decompression.
Good quality for 30+ compression ratio (27.5 PSNR for 1KB/frame)
Lightweight NN for simultaneous deblocking and frame interpolation.
Motion compensation inspired by H.264
Frame differencing and quantization inspired by DPCM
Multilevel Wavelet Decomposition and DCT.
Publications
Please refer to my Google Scholar page.