Abhiram Dharme
Research

What I’m working on.

Selected work around deep learning for wearable biosignals: models that respect sensor geometry, experiments that probe data scarcity, and systems that make those ideas measurable.

2025 —
First author

CycloFormer

A rotation-invariant transformer for wrist-sEMG hand pose estimation, with a variance-corrected scaling law for the data-limited regime.

NeurIPS 2026 (under review)

  • Exact ℤ₁₆ invariance to wristband donning rotation, built from a channel-shared TDS-CNN, circular rotary position embeddings, and permutation-invariant attention pooling.
  • New state of the art on emg2pose: a 4M-parameter model surpasses the strongest prior baseline on every generalization split with 33% fewer parameters.
  • A variance-corrected scaling law for the data-limited regime, fit on a 5×8 model–data grid and explaining 98.8% of the observed variance.
  • First controlled quantitative comparison of sEMG and egocentric vision under fingertip occlusion.
Oct 2025 — May 2026
IIT Delhi contingent

RRM Plus — Client-Aware WiFi Resource Management

A graph-neural-network controller for dense WiFi networks, trained with safe reinforcement learning under operational guardrails. Built for the Arista Networks problem statement at Inter-IIT Tech Meet 14.

Inter-IIT Tech 14 · Arista Networks

  • Generated a 1,000-scenario, 4,000-AP, 24,000-client WiFi training corpus on the Komondor / NS-3 simulation stack after ruling out hardware testbeds and MATLAB simulations for scalability.
  • Modeled the network as a heterogeneous graph (AP, client, interference edges) and trained a GNN over per-AP configuration parameters and per-client signal quality / QoE metrics.
  • Trained the policy with Constrained Policy Optimization + a safety critic — every configuration update is screened through a trust-region check on predicted QoE impact and per-AP change budgets.
  • Designed a multi-timescale controller: a fast loop (second-to-minute steering) with strict guardrails, a slow loop (hourly global stabilization), and an event loop that rolls back any action causing > 10% edge-throughput drop.