Research • Springer Singapore • Dec 2024

Autoencoder-Huffman Ensemble

PyTorch • Autoencoders • Huffman coding • MNIST / CIFAR-10 / Fashion MNIST

Two-stage compression: encoder-decoder learns a compact Latent Space Representation (LSR), Huffman squeezes it further. Published in Springer proceedings. Springer link → • Code →

PyTorchAutoencodersHuffmanResearch
Autoencoder-Huffman — encoder → latent 70–100 → Huffman → 24.2× compression
Encoder → LSR (70–100 dims) → Quantize → Huffman → Bitstream • ACR table included — Click to enlarge →
PROBLEM

Image datasets are large and noisy. Classical Huffman is optimal for symbols but not for features; pure deep compression loses interpretability. Can a hybrid win on ratio without hurting reconstruction?

APPROACH
  • Trained encoder-decoder to map images → low-dimensional LSR that preserves structure while filtering noise.
  • Applied Huffman on quantized latent codes — classical entropy coding on learned representations.
  • Evaluated across MNIST, CIFAR-10, Fashion MNIST at latent dims 70–100, with/without codebook overhead.
Diagram: Image → Encoder → LSR (70–100 dims) → Quantize → Huffman → Bitstream → Huffman Decode → Decoder → Recon.
RESULT

5.5×–24.2× compression with minimal loss.

ACRs without overhead: 15.654 (MNIST), 24.230 (CIFAR-10), 15.171 (Fashion MNIST). With overhead: ~5.5–8.7 — robust across dims. Published Springer, thesis at NIT Warangal (Prof. U.S.N. Raju).

WHAT I LEARNED

Hybrid deep + classical beats either alone when the representation is right. Also: rigorous evaluation (with vs without overhead) matters more than a single peak number.

Publication
Springer, Singapore — Dec 2024
link.springer.com →
Thesis: NIT Warangal