Achieving 62.46% MNIST Accuracy with Just 984 Learnable Parameters
Tall_Abrocoma_3533 · reddit · 2026-08-10
A developer shared an experiment focusing on extreme low-parameter and few-shot learning on the MNIST dataset. The project trains on just 10 images (1 per class) and compresses learnable parameters to 984, achieving a validation accuracy of 62.46%.
Technical Implementation:
- Fixed Vision Pipeline: The vision feature extraction is entirely fixed, using deterministic compression to reduce features to 81 dimensions.
- Prototype Matching: The trainable part is a single 81 -> 12 linear layer. The model uses cosine similarity for prototype matching in this 12D space.
- Training Strategy: Training uses 200 augmented queries per class per episode against 10 unmodified support images.
- Hardware & Time: Training was conducted on a single core (Cortex-X4) of the Dimensity 9300+ chipset, taking about 0.5 seconds/epoch. The total training time for 172 epochs was around 90 seconds.
More from Research
- CMU Lab Demonstrates Humanoid Robot Bending Banana Shots — TinfoilTricorn · 2026-08-10
- Why LLMs Have a Bland and Non-Committal Writing Style — scaling01 · 2026-08-10
- Why LLMs Write Bland Text: 6 Technical Causes and the Path to Fixing It — scaling01 · 2026-08-10
- 3DGS Training Breakthrough: 1M Gaussians in 20 Seconds — janusch_patas · 2026-08-10
- Paper: CKA-QAD Significantly Improves Reasoning and Coding in NVFP4 Quantized LLMs — Aaaaaaaaaeeeee · 2026-08-10
- Non-LLM Reasoning System Scores 100% on ARC-AGI-3 ft09 with Zero Model Calls — Living_Substance1274 · 2026-08-10