5.9 KiB
| name | description | metadata | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| knp-mnist-altai-work | Ongoing work on KNP (Kaspersky Neuromorphic Platform): mnist-learn altai/blifat accuracy + cats-vs-dogs SNN experiment, acceleration, GPU STDP prototype — state and next steps |
|
This WSL machine (1220WS2, IP 10.9.8.230 on wswg0) IS the "test stand" referenced in handoff docs; /mnt/c/knp == /mnt/c/KNP. No SSH needed. Runtime: ~/run/exp (MNIST), ~/run/cats-dogs (cats/dogs), build ~/knp-build, python with numpy: ~/anvenv/bin/python.
Git: /mnt/c/knp is one repo (branch main, remote https://git.snw.su/int/our_knp.git), created 2026-10-14. Nested .git of 1/, viz/, examples/ moved to /mnt/c/knp-nested-git-backup/ (examples = KasperskyLab/knp @55470ed6 + 6 local file changes). Local DNS doesn't resolve git.snw.su → repo config http.curloptResolve to 45.141.100.133. Credentials saved (credential.helper store) — Claude can push directly; push after commits.
MNIST (to 2026-10-13): altai 82.5%→91.5% via KNP_COL_JITTER=20 + KNP_MNIST_INFER_INHIB=255; fixed ModelLoader::init_channels bug. Simulator viz/altai_sim.py, visualizer viz/app.html→altai_viz.html. Col size 20≈40 (91.9%). Fashion ~55%. Weakness: OFF pixels send no spikes. Agreed next idea: ON/OFF input coding. sz80_*/sz40_60k runs have empty result.csv (killed). Reports: /mnt/c/knp/reports// (index reports/README.md).
Cats/dogs (2026-10-14, other chat; docs cats-dogs/HANDOFF.md, V2.md, ACCELERATION.md, reports/2026-10-14/журнал.md): v1 contrast input 52% test; v2 HOG-like features (non-spiking) + KNP STDP: val 62.6%, test 61.8% (test already viewed — not for further tuning). Parallel independent processes 2.1× throughput. Verified 2026-10-14 evening: tuned multithread backend and fast connection-cache backend give byte-identical CSVs; tuned2 3.5× slower, fast ~10% faster (800+200 only). GPU STDP hybrid draft (gpu_stdp*.cu/.cpp) written, NOT built/verified.
MAIN GOAL (user, 2026-10-14): object recognition; MNIST is a testbed — carry every finding over to objects (cats/dogs). MNIST 92% is only on aligned MNIST: shift 2px → ~64%, thickness ±1px → 75-79% (template-like net); hand-drawn digits in viz look bad. ON/OFF naive (local) gave 56.5% val: input spike count tripled → 1/3 images silent (all column sums negative), not a quantization issue; needs spike-rate normalization.
Night 14→15 Oct (report reports/2026-10-15/Отчёт_объекты_ночь.md, log reports/2026-10-14/журнал.md): bottleneck for objects was FEATURES (HOG: even linear ~60%). Learned k-means patch features (objects/kmeans_features.py, ~WTA competitive learning) + linear: 77% on fresh test; fully spiking exact sim (objects/spiking_pipeline.py, layer1 800×729 AltAI-like neurons + one "dog−cat" readout neuron trained off-chip): 68.3%; KNP STDP columns on top-256 learned features with KNP_COL_PUNISH=0.012: 67.0%; old v2: 62.5%. Fresh test = CIFAR test 700–999/class, NOW USED (all CIFAR cat/dog test images consumed). Data v3: ~/run/cats-dogs/data-v3 (train 9000). KNP_COL_PUNISH (neg. dopamine to wrong columns) best 0.012–0.02 (+3.8pp MNIST). STDP columns fail on dense inputs (weights mostly negative → silence); need sparse input. MNIST: stroke features + spiking readout 95.1%, shift-2px 89.9%.
Later that night: online WTA (objects/kmeans_features.learn_online) ≈ k-means; three-factor dopamine-error readout (objects/dopamine_readout.py) + calibrated decision neuron: all-neural-rule pipeline 71.6% val (K=1600), ridge 73.5%; seed spread D 64.0±0.6 vs v2 62.1±1.2. Digit page (artifact Vc1q1WCZuPY2QAHBSGtrKF v2) has stroke-feature spiking readout panel (94.7% MNIST test).
Morning 15 Oct (log reports/2026-10-15/журнал.md): KNP_DELTA=a,b (+ isolated backend readout/build_delta_cpu.py → ~/run/delta-cpu/libknp-cpu-delta-backend.so, env KNP_DELTA_RULE=rate, KNP_DELTA_LAG=1 REQUIRED — dopamine arrives 1 step after input) = error-driven dopamine readout inside KNP, KNP_COL_SIZE=1. MNIST coarse features: 94.8% (b=4–6, r=3), augmented 90k: 94.6%, shift-2px 91.3%. Cats/dogs: 65–69% (columns 64–65%, ridge 74%); weight range (KNP_W_RANGE) and 3 epochs don't help. cats-dogs/run.py --backend delta; cats_dogs input size via KNP_MNIST_INPUT_SIZE.
Next: close KNP object gap (smaller steps/learning-rate decay?), port WTA feature layer into KNP; reduce spike-coding loss; make columns handle dense features.
Waiting tip: never pgrep -f <name> for a string that appears in the waiting command itself (self-match hung a wait 1.5h); use ps -eo args | grep -c "[x]yz". tail -5 fails here; use tail -n 5.
RESUME HERE (2026-10-15 11:20, session was shut down): vehicles (CIFAR airplane/automobile/ship/truck, data ~/run/cifar/vehicles) inference-in-KNP pipeline in /mnt/c/knp/altai-objects: build_model.py (k-means K=200 stride 2 folded into int pixel weights) → model ~/run/altai-objects/model_k200; altai_infer (KNP inference, 0.8 s/img, 1.5 GB) ; run_counts.sh collecting detector counts (resumable, *.done) → train_readout.py → altai_infer on val in KNP → accuracy. Then demo page (local server calling altai_infer) + CNN comparison. Exact steps in reports/2026-10-15/журнал.md (last section). numpy prototype predicted ~76% val for this design.
User wants a real KNP/AltAI spiking net trained (not CNN substitution). Daily log: reports/YYYY-MM-DD/журнал.md (+ reports of that day in the same folder; index reports/README.md, living overview reports/Итоги_и_план.md): short for discussion, detailed for code/runs; no passwords.
Why: session transcripts don't survive between days; user expects continuity.
How to apply: read reports/README.md, reports/Итоги_и_план.md and the latest reports//журнал.md; check uptime (WSL restarts kill runs) and result files before trusting; update this memory and the daily log at session end.