Publications

Synthetic Benchmarks Overstate Forward-Forward Scaling: Real-Data Limits of Layer-Local Training

A scaling audit of layer-local training. DTG-FF is built as the strongest Forward-Forward instance available and then used as an instrument: under an identical recipe and backbone an architecture-matched backprop baseline stays ahead by 2.40/5.93 pp on CIFAR-10/100, the gap widens with class count where synthetic teacher-student benchmarks predict it should close, and the memory advantage does not survive gradient accumulation. Includes the first FF-family baseline at ImageNet-100 224x224.

SegWithU: Uncertainty as Perturbation Energy for Single-Forward-Pass Risk-Aware Medical Image Segmentation

A post-hoc framework that augments frozen pretrained segmentation models with a lightweight uncertainty module, modeling uncertainty as perturbation energy in a compact probe space via rank-1 posterior probes for single-forward-pass failure detection.

MIP Candy: A Modular PyTorch Framework for Medical Image Processing

A modular PyTorch framework bridging the gap between low-level component libraries and rigid monolithic pipelines for medical image processing.

ATLAS-CNN: A Novel Hybrid Deep Learning Architecture for Stock Market Prediction

A hybrid architecture that pairs a Temporal Multi-Dimensional Operator with a CNN, reorganizing financial time series into a 2D image-like format so convolution kernels can capture both temporal and cross-indicator structure. Reaches 73-75% directional accuracy across a range of stocks, ahead of LSTM baselines.