✯ NAME: S. KHAVIN S. KHAVIN
✯ RESEARCH: MECHANISTIC INTERPRETABILITY • EFFICIENT INFERENCE MECHANISTIC INTERPRETABILITY • EFFICIENT INFERENCE
✯ EDUCATION: B.E. CSE, SATHYABAMA • 2023 - 2027 B.E. CSE, SATHYABAMA • 2023 - 2027
✯ STATUS: ACTIVE ACTIVE
✯ RESEARCH SUMMARY:
INDEPENDENT RESEARCHER IN TRANSFORMER INTERPRETABILITY AND INFERENCE EFFICIENCY. INDEPENDENT RESEARCHER IN TRANSFORMER INTERPRETABILITY AND INFERENCE EFFICIENCY.
FOCUS: CAUSAL, NEUROSCIENCE-INSPIRED METHODS TO UNDERSTAND & ACCELERATE MODELS. FOCUS: CAUSAL, NEUROSCIENCE-INSPIRED METHODS TO UNDERSTAND & ACCELERATE MODELS.
SIX FIRST-AUTHOR PREPRINTS SPANNING GEOMETRY, ROUTING, KV-CACHE, AND VIDEO UNDERSTANDING. SIX FIRST-AUTHOR PREPRINTS SPANNING GEOMETRY, ROUTING, KV-CACHE, AND VIDEO UNDERSTANDING.
✯ PREPRINTS:
• K HEADS IS ALL YOU NEED: CIRCUIT GRANULARITY & ZERO-TRAINING ROUTING. CIRCUIT GRANULARITY & ZERO-TRAINING ROUTING.
• THE SHAPE OF COMPUTATION: GEOMETRIC TRAJECTORIES IN RESIDUAL STREAMS. GEOMETRIC TRAJECTORIES IN RESIDUAL STREAMS.
• NECESSARY BUT NOT SUFFICIENT: CAUSAL TESTS OF UNIVERSAL GEOMETRY. CAUSAL TESTS OF UNIVERSAL GEOMETRY.
• HEADGENOME: CAUSAL MAP OF 1,568 ATTENTION HEADS ACROSS 4 ARCHITECTURES. CAUSAL MAP OF 1,568 ATTENTION HEADS ACROSS 4 ARCHITECTURES.
• O(1) DECODE STEP ATTENTION: PROACTIVE KV-CACHE EVICTION. PROACTIVE KV-CACHE EVICTION.
• TRINETRA: O(N) VIDEO INGEST, O(1) QUERY TEMPORAL EVENT GRAPH. O(N) VIDEO INGEST, O(1) QUERY TEMPORAL EVENT GRAPH.
✯ DISTINCTIONS:
• SWE CO-OP INTERN @ ALGOUNIVERSITY | UDACITY SCHOLAR (1 OF 500) SWE CO-OP INTERN @ ALGOUNIVERSITY | UDACITY SCHOLAR (1 OF 500)
• ALGOUNIVERSITY TECHNOLOGY FELLOWSHIP, TOP 5% OF 20,000 ALGOUNIVERSITY TECHNOLOGY FELLOWSHIP, TOP 5% OF 20,000
✯ METHODS:
PYTORCH • FAISS • CAUSAL PATCHING • DIRECT LOGIT ATTRIBUTION • CKA • DTW PYTORCH • FAISS • CAUSAL PATCHING • DIRECT LOGIT ATTRIBUTION • CKA • DTW