Inference-on-KNP pipeline: altai_infer, offline model building and readout training

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-15 10:42:35 +03:00 · 2026-10-15 10:42:35 +03:00 · 151e9f0b23
commit 151e9f0b23
parent ded57ee2f2
8 changed files with 411 additions and 6 deletions
--- a/altai-objects/altai_infer.cpp
+++ b/altai-objects/altai_infer.cpp
@ -0,0 +1,223 @@
+// Object recognition inference entirely in KNP on AltAI neurons: pixels → feature detectors → class neurons.
+//
+// altai_infer <images.bin> <first> <count> <model_dir> <backend.so> <out_prefix>
+//
+// images.bin: uint8 images 3×32×32. Pixels spike with the incrementing converter (value/255 per step, 10 active
+// steps of 15). model_dir (written by build_model.py / train_readout.py):
+//   meta.txt        K stride det_threshold classes out_threshold
+//   det_w.bin       int32 [K][108]  detector weights on a 6×6×3 pixel patch (channel, y, x)
+//   det_b.bin       int32 [K]       detector bias, delivered on every active step by an always-on input
+//   readout_w.bin   int32 [classes][4·K] (optional) class weights per (quadrant, detector type); index q·K + k
+//   readout_b.bin   int32 [classes] (optional) class bias per active step
+// Detectors: one AltAI neuron per detector type and position (stride), threshold det_threshold, reset to 0, the
+// potential does not go below 0. Class neurons: AltAI, threshold out_threshold.
+// Output: <out_prefix>_classes.csv (index, spike count per class) if a readout is given, and
+// <out_prefix>_counts.bin (int32 [count][4·K], detector spikes per quadrant and type) for training the readout.
+
+#include <knp/framework/model_executor.h>
+#include <knp/framework/network.h>
+#include <knp/neuron-traits/altai_lif.h>
+#include <knp/synapse-traits/delta.h>
+
+#include <chrono>
+#include <cstring>
+#include <filesystem>
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <vector>
+
+
+using DeltaProjection = knp::core::Projection<knp::synapse_traits::DeltaSynapse>;
+using AltAIPopulation = knp::core::Population<knp::neuron_traits::AltAILIF>;
+
+constexpr int side = 32, patch = 6, channels = 3, steps_per_image = 15, active_steps = 10;
+constexpr int fan_in = patch * patch * channels, in_n = channels * side * side;
+
+
+template <class T>
+std::vector<T> read_vec(const std::filesystem::path &path)
+{
+    std::ifstream f(path, std::ios::binary);
+    if (!f) return {};
+    std::vector<char> raw((std::istreambuf_iterator<char>(f)), std::istreambuf_iterator<char>());
+    std::vector<T> out(raw.size() / sizeof(T));
+    std::memcpy(out.data(), raw.data(), out.size() * sizeof(T));
+    return out;
+}
+
+
+knp::neuron_traits::neuron_parameters<knp::neuron_traits::AltAILIF> neuron(int threshold)
+{
+    knp::neuron_traits::neuron_parameters<knp::neuron_traits::AltAILIF> p;
+    p.activation_threshold_ = static_cast<uint16_t>(threshold);
+    p.negative_activation_threshold_ = 0;
+    p.potential_reset_value_ = 0;
+    p.potential_leak_ = 0;
+    return p;
+}
+
+
+int main(int argc, const char *const argv[])
+{
+    if (argc != 7)
+    {
+        std::cerr << "usage: altai_infer <images.bin> <first> <count> <model_dir> <backend.so> <out_prefix>\n";
+        return 1;
+    }
+    const std::filesystem::path model_dir = argv[4];
+    const size_t first = std::stoul(argv[2]), count = std::stoul(argv[3]);
+    const std::string out_prefix = argv[6];
+    int k = 0, stride = 0, det_thr = 0, classes = 0, out_thr = 0;
+    {
+        std::ifstream meta(model_dir / "meta.txt");
+        meta >> k >> stride >> det_thr >> classes >> out_thr;
+    }
+    const int per_side = (side - patch) / stride + 1, positions = per_side * per_side, half = per_side / 2;
+    const size_t det_n = static_cast<size_t>(k) * positions;
+    const auto det_w = read_vec<int32_t>(model_dir / "det_w.bin"), det_b = read_vec<int32_t>(model_dir / "det_b.bin");
+    const auto ro_w = read_vec<int32_t>(model_dir / "readout_w.bin"), ro_b = read_vec<int32_t>(model_dir / "readout_b.bin");
+    const bool readout = !ro_w.empty();
+    std::ifstream images(argv[1], std::ios::binary);
+    std::vector<uint8_t> px(count * in_n);
+    images.seekg(static_cast<std::streamoff>(first * in_n));
+    images.read(reinterpret_cast<char *>(px.data()), static_cast<std::streamsize>(px.size()));
+
+    auto quadrant = [&](size_t pos) { return (pos / per_side >= half) * 2 + (pos % per_side >= half); };
+
+    AltAIPopulation detectors([&](size_t) { return neuron(det_thr); }, det_n);
+    AltAIPopulation classes_pop([&](size_t) { return neuron(out_thr); }, readout ? classes : 0);
+    // Detector index = type * positions + position; synapse index = detector * fan_in + patch element.
+    DeltaProjection pixel_proj(
+        knp::core::UID{false}, detectors.get_uid(),
+        [&](size_t index) -> std::optional<DeltaProjection::Synapse>
+        {
+            const size_t d = index / fan_in, e = index % fan_in, j = d / positions, pos = d % positions;
+            const int32_t w = det_w[j * fan_in + e];
+            if (w == 0) return std::nullopt;
+            const size_t c = e / (patch * patch), dy = (e % (patch * patch)) / patch, dx = e % patch;
+            const size_t y = (pos / per_side) * stride + dy, x = (pos % per_side) * stride + dx;
+            return DeltaProjection::Synapse{
+                {static_cast<float>(w), 1, knp::synapse_traits::OutputType::EXCITATORY}, c * side * side + y * side + x,
+                d};
+        },
+        det_n * fan_in);
+    // Always-on input 0 → every detector (its bias) and every class neuron (its bias).
+    DeltaProjection bias_det(
+        knp::core::UID{false}, detectors.get_uid(),
+        [&](size_t d) -> std::optional<DeltaProjection::Synapse>
+        {
+            const int32_t b = det_b[d / positions];
+            if (b == 0) return std::nullopt;
+            return DeltaProjection::Synapse{{static_cast<float>(b), 1, knp::synapse_traits::OutputType::EXCITATORY}, 0, d};
+        },
+        det_n);
+    DeltaProjection readout_proj(
+        detectors.get_uid(), classes_pop.get_uid(),
+        [&](size_t index) -> std::optional<DeltaProjection::Synapse>
+        {
+            const size_t c = index / det_n, d = index % det_n, j = d / positions, pos = d % positions;
+            const int32_t w = ro_w[c * 4 * k + quadrant(pos) * k + j];
+            if (w == 0) return std::nullopt;
+            return DeltaProjection::Synapse{{static_cast<float>(w), 1, knp::synapse_traits::OutputType::EXCITATORY}, d, c};
+        },
+        readout ? det_n * classes : 0);
+    // The class bias arrives one step later than the detector bias, in step with the detector spikes.
+    DeltaProjection bias_cls(
+        knp::core::UID{false}, classes_pop.get_uid(),
+        [&](size_t c) -> std::optional<DeltaProjection::Synapse>
+        {
+            if (ro_b[c] == 0) return std::nullopt;
+            return DeltaProjection::Synapse{
+                {static_cast<float>(ro_b[c]), 2, knp::synapse_traits::OutputType::EXCITATORY}, 0, c};
+        },
+        readout ? classes : 0);
+
+    const auto pixel_uid = pixel_proj.get_uid(), bias_det_uid = bias_det.get_uid(), bias_cls_uid = bias_cls.get_uid(),
+               det_uid = detectors.get_uid(), cls_uid = classes_pop.get_uid();
+    knp::framework::Network network;
+    network.add_population(std::move(detectors));
+    network.add_projection(std::move(pixel_proj));
+    network.add_projection(std::move(bias_det));
+    if (readout)
+    {
+        network.add_population(std::move(classes_pop));
+        network.add_projection(std::move(readout_proj));
+        network.add_projection(std::move(bias_cls));
+    }
+    knp::framework::Model model(std::move(network));
+    knp::core::UID pixel_channel, bias_channel, det_channel, cls_channel;
+    model.add_input_channel(pixel_channel, pixel_uid);
+    model.add_input_channel(bias_channel, bias_det_uid);
+    if (readout) model.add_input_channel(bias_channel, bias_cls_uid);
+    model.add_output_channel(det_channel, det_uid);
+    if (readout) model.add_output_channel(cls_channel, cls_uid);
+
+    std::vector<knp::core::messaging::SpikeData> current(active_steps);
+    auto pixel_gen = [&](knp::core::Step step) -> knp::core::messaging::SpikeData
+    {
+        const size_t img = step / steps_per_image, local = step % steps_per_image;
+        if (local == 0)
+        {
+            std::vector<float> st(in_n, 0.F);
+            for (auto &s : current) s.clear();
+            const uint8_t *p = px.data() + img * in_n;
+            for (int t = 0; t < active_steps; ++t)
+                for (int i = 0; i < in_n; ++i)
+                {
+                    st[i] += p[i] / 255.F;
+                    if (st[i] >= 1.F)
+                    {
+                        st[i] -= 1.F;
+                        current[t].push_back(static_cast<uint32_t>(i));
+                    }
+                }
+        }
+        if (local >= active_steps) return {};
+        return current[local];
+    };
+    auto bias_gen = [&](knp::core::Step step) -> knp::core::messaging::SpikeData
+    {
+        if (step % steps_per_image >= active_steps) return {};
+        return {0};
+    };
+
+    knp::framework::BackendLoader loader;
+    knp::framework::ModelExecutor executor(
+        model, loader.load(argv[5]), {{pixel_channel, pixel_gen}, {bias_channel, bias_gen}});
+    auto &det_out = executor.get_loader().get_output_channel(det_channel);
+    std::vector<int32_t> counts(count * 4 * k, 0);
+    std::vector<int32_t> class_counts(count * std::max(classes, 1), 0);
+    const auto t0 = std::chrono::steady_clock::now();
+    executor.start([&](size_t step) { return step < count * steps_per_image; });
+    for (const auto &m : det_out.update())
+    {
+        const size_t im = m.header_.send_time_ / steps_per_image;
+        if (im >= count) continue;
+        for (auto d : m.neuron_indexes_) ++counts[(im * 4 + quadrant(d % positions)) * k + d / positions];
+    }
+    if (readout)
+        for (const auto &m : executor.get_loader().get_output_channel(cls_channel).update())
+        {
+            const size_t im = m.header_.send_time_ / steps_per_image;
+            if (im < count)
+                for (auto c : m.neuron_indexes_) ++class_counts[im * classes + c];
+        }
+    std::ofstream(out_prefix + "_counts.bin", std::ios::binary)
+        .write(reinterpret_cast<const char *>(counts.data()), static_cast<std::streamsize>(counts.size() * 4));
+    if (readout)
+    {
+        std::ofstream csv(out_prefix + "_classes.csv");
+        csv << "index";
+        for (int c = 0; c < classes; ++c) csv << ",c" << c;
+        csv << "\n";
+        for (size_t i = 0; i < count; ++i)
+        {
+            csv << first + i;
+            for (int c = 0; c < classes; ++c) csv << ',' << class_counts[i * classes + c];
+            csv << "\n";
+        }
+    }
+    std::cout << "images " << first << ".." << first + count - 1 << ", "
+              << std::chrono::duration<double>(std::chrono::steady_clock::now() - t0).count() << " s" << std::endl;
+}
--- a/altai-objects/altai_objects.cpp
+++ b/altai-objects/altai_objects.cpp
@ -16,7 +16,8 @@
 //     (index,label,spike counts per class).
 //
 // Environment: KNP_RETINA_THR (default 250), KNP_DET_THR (detector threshold, default 2000),
-// KNP_OUT_THR (readout threshold, default 4000), KNP_INHIB (detector lateral inhibition, default 5000).
+// KNP_OUT_THR (readout threshold, default 4000), KNP_INHIB (detector lateral inhibition, default 5000),
+// KNP_SUMTHR (features: threshold increase per unit of positive weight sum, default 0).

 #include <knp/framework/model_executor.h>
 #include <knp/framework/network.h>
@ -211,7 +212,20 @@ int features_mode(int argc, const char *const argv[])
    if (px.size() < count * in_n) throw std::runtime_error("not enough patches");

    AltAIPopulation retina([&](size_t) { return plain_neuron(env_f("KNP_RETINA_THR", 250)); }, ret_n);
-    StdpPopulation detectors([&](size_t) { return learning_neuron(env_f("KNP_DET_THR", 2000)); }, k);
+    StdpPopulation detectors(
+        [&](size_t)
+        {
+            auto p = learning_neuron(env_f("KNP_DET_THR", 2000));
+            // AltAI homeostasis: the threshold grows with the sum of positive weights (KNP_SUMTHR × sum), which
+            // works like weight normalisation, so a detector wins by pattern match rather than by weight mass.
+            p.synapse_sum_threshold_coefficient_ = env_f("KNP_SUMTHR", 0);
+            // Lasting inhibition: the potential may go down to -KNP_DET_NEG, and a leak of KNP_DET_LEAK per step
+            // brings it back towards 0 (in the 5 silent steps between crops).
+            p.negative_activation_threshold_ = static_cast<uint16_t>(env_f("KNP_DET_NEG", 0));
+            p.potential_leak_ = static_cast<int16_t>(env_f("KNP_DET_LEAK", 0));
+            return p;
+        },
+        k);
    auto retina_proj = make_retina(retina.get_uid(), crop, crop, patch, patch, dog_r);

    std::mt19937 rng(static_cast<unsigned>(env_f("KNP_SEED", 1)));
@ -248,17 +262,42 @@ int features_mode(int argc, const char *const argv[])
        },
        k * k);

+    // Self-inhibition after a spike (KNP_SELF_INHIB), so a winner fires about once per crop.
+    const float self_inhib = env_f("KNP_SELF_INHIB", 0);
+    DeltaProjection self_proj(
+        detectors.get_uid(), detectors.get_uid(),
+        [&](size_t i) {
+            return std::optional(DeltaProjection::Synapse{
+                {-self_inhib, 1, knp::synapse_traits::OutputType::EXCITATORY}, i, i});
+        },
+        self_inhib > 0 ? k : 0);
+    // Imprinting (KNP_IMPRINT=1): crop number c teaches detector c mod K through a dopamine channel on every active
+    // step; with KNP_DELTA_WDECAY the detector's weights settle at the retina activity of that crop.
+    const bool imprint = env_f("KNP_IMPRINT", 0) > 0;
+    DeltaProjection teacher(
+        knp::core::UID{false}, detectors.get_uid(),
+        [&](size_t i) {
+            return std::optional(DeltaProjection::Synapse{
+                {env_f("KNP_TEACH", 1), 1, knp::synapse_traits::OutputType::DOPAMINE}, i, i});
+        },
+        k);
+    const auto teacher_uid = teacher.get_uid();
    const auto retina_uid = retina_proj.get_uid(), learn_uid = learn_proj.get_uid(), det_uid = detectors.get_uid();
    knp::framework::Network network;
    network.add_population(std::move(retina));
    network.add_population(std::move(detectors));
    network.add_projection(std::move(retina_proj));
    network.add_projection(std::move(learn_proj));
-    network.add_projection(std::move(lateral));
-    knp::core::UID in_channel, out_channel;
+    if (imprint)
+        network.add_projection(std::move(teacher));
+    else
+        network.add_projection(std::move(lateral));
+    if (self_inhib > 0 && !imprint) network.add_projection(std::move(self_proj));
+    knp::core::UID in_channel, out_channel, teach_channel;
    knp::framework::Model model(std::move(network));
    model.add_input_channel(in_channel, retina_uid);
    model.add_output_channel(out_channel, det_uid);
+    if (imprint) model.add_input_channel(teach_channel, teacher_uid);

    std::vector<knp::core::messaging::SpikeData> current;
    auto input_gen = [&](knp::core::Step step) -> knp::core::messaging::SpikeData
@ -270,7 +309,14 @@ int features_mode(int argc, const char *const argv[])
    };
    knp::framework::BackendLoader loader;
    auto backend = loader.load(backend_path);
-    knp::framework::ModelExecutor executor(model, backend, {{in_channel, input_gen}});
+    auto teach_gen = [&](knp::core::Step step) -> knp::core::messaging::SpikeData
+    {
+        if (step % steps_per_image >= active_steps) return {};
+        return {static_cast<uint32_t>((step / steps_per_image) % k)};
+    };
+    knp::framework::ModelLoader::InputChannelMap channels{{in_channel, input_gen}};
+    if (imprint) channels.insert({teach_channel, teach_gen});
+    knp::framework::ModelExecutor executor(model, backend, std::move(channels));
    auto &out = executor.get_loader().get_output_channel(out_channel);
    const auto t0 = std::chrono::steady_clock::now();
    executor.start([&](size_t step) { return step < count * steps_per_image; });
--- a/altai-objects/build_model.py
+++ b/altai-objects/build_model.py
@ -0,0 +1,42 @@
+"""Train the feature detectors offline and write integer weights for altai_infer (inference runs in KNP).
+
+k-means on 6×6×3 patches of the vehicles training images (patch centring and ZCA, no contrast division), then
+centring, ZCA and the prototype are folded into one weight vector per detector type on raw pixels, the offset
+and alpha into a per-step bias (objects/spiking_pipeline.py); weights are scaled to integers in ±wscale.
+usage: ~/anvenv/bin/python build_model.py <model_dir> [--k 200 --stride 2 --alpha 0.25 --wscale 127]
+"""
+import argparse
+import sys
+from pathlib import Path
+
+import numpy as np
+
+sys.path.insert(0, "/mnt/c/knp/objects")
+import kmeans_features as kf  # noqa: E402
+import spiking_pipeline as sp  # noqa: E402
+
+ap = argparse.ArgumentParser()
+ap.add_argument("out")
+ap.add_argument("--k", type=int, default=200)
+ap.add_argument("--stride", type=int, default=2)
+ap.add_argument("--alpha", type=float, default=0.25)
+ap.add_argument("--wscale", type=float, default=127)
+ap.add_argument("--classes", type=int, default=4)
+ap.add_argument("--data", default="/home/int/run/cifar/vehicles")
+a = ap.parse_args()
+out = Path(a.out)
+out.mkdir(parents=True, exist_ok=True)
+kf.CONTRAST_DIVIDE = False
+xt = np.load(f"{a.data}/train.npz")["rgb"]
+model = kf.learn(xt, k=a.k)
+w, b = sp.fold(model, a.alpha)  # [108, K], [K] for patches in [0, 1]
+scale = a.wscale / np.abs(w).max()
+wq, bq = np.rint(w * scale).astype(np.int32), np.rint(b * scale).astype(np.int32)
+p = kf.patches_at(xt[:300].astype(np.float32) / 255, range(0, 27, a.stride), range(0, 27, a.stride)).reshape(-1, 108)
+act = p @ w + b
+thr = int(round(float(np.percentile(act[act > 0], 99)) * scale))
+wq.T.copy().tofile(out / "det_w.bin")
+bq.tofile(out / "det_b.bin")
+(out / "meta.txt").write_text(f"{a.k} {a.stride} {thr} {a.classes} 0\n")
+np.savez(out / "kmeans.npz", **model)
+print(f"K {a.k} stride {a.stride}: weight scale {scale:.2f}, detector threshold {thr}, bias range {bq.min()}..{bq.max()}")
--- a/altai-objects/feature_grid.sh
+++ b/altai-objects/feature_grid.sh
@ -3,7 +3,7 @@
 cd ~/run/altai-objects/fgrid
 export LD_LIBRARY_PATH=$HOME/knp-build/bin:$HOME/deps/lib
 one() { n=$1; [ -f $n.score ] && return
-  env KNP_DELTA_RULE=$2 KNP_DELTA_DECAY=$3 KNP_DELTA_LAG=1 KNP_DET_THR=$4 KNP_INHIB=$5 \
+  env KNP_DELTA_RULE=$2 KNP_DELTA_DECAY=$3 KNP_DELTA_WDECAY=${7:-0} KNP_DELTA_LAG=1 KNP_DET_THR=$4 KNP_INHIB=$5 KNP_SUMTHR=${6:-0} \
    ../altai_objects features /home/int/run/cifar/vehicles/knp_crops.bin ${NPATCH:-20000} 64 $n.bin \
    $HOME/run/delta-cpu/libknp-cpu-delta-backend.so 2>&1 | grep "^features" > $n.log
  ~/anvenv/bin/python /mnt/c/knp/altai-objects/score_features.py $n.bin 64 > $n.score 2>&1; }
--- a/altai-objects/run_counts.sh
+++ b/altai-objects/run_counts.sh
@ -0,0 +1,10 @@
+#!/bin/bash
+# Collect detector spike counts from KNP for readout training: run_counts.sh <model_dir> <out_dir> [chunk]
+M=$1; O=$2; C=${3:-1000}
+D=/home/int/run/cifar/vehicles
+export LD_LIBRARY_PATH=$HOME/knp-build/bin:$HOME/deps/lib
+one() { f=$1; first=$2; n=$3; tag=$4; [ -f $O/${tag}_$first.done ] && return
+  ~/run/altai-objects/altai_infer $D/${f}_images.bin $first $n $M $HOME/knp-build/bin/libknp-cpu-single-threaded-backend.so $O/${tag}_$first > $O/${tag}_$first.log 2>&1 && touch $O/${tag}_$first.done; }
+export -f one; export M O D
+{ for s in train train_flip; do for i in $(seq 0 $C 15999); do echo "$s $i $C $s"; done; done
+  for i in $(seq 0 $C 3999); do echo "val $i $C val"; done; } | xargs -P ${JOBS:-14} -L 1 bash -c 'one "$@"' _
--- a/altai-objects/train_readout.py
+++ b/altai-objects/train_readout.py
@ -0,0 +1,65 @@
+"""Train the class neurons offline on detector spike counts produced by KNP (altai_infer), write integer weights.
+
+Ridge regression (one-vs-rest, targets ±1) on per-(quadrant, detector type) spike counts of the training images
+(+ mirrored copies). Standardisation is folded into the weights and bias; weights are scaled to integers in ±255,
+the bias is spread over the active steps. The class-neuron threshold is chosen on the training counts with a rate
+approximation (spikes ≈ positive drive / threshold) so that the most-spiking class matches the ridge decision.
+usage: ~/anvenv/bin/python train_readout.py <model_dir> <counts_dir> [--lam 1000]
+"""
+import argparse
+from pathlib import Path
+
+import numpy as np
+
+ACTIVE = 10
+ap = argparse.ArgumentParser()
+ap.add_argument("model")
+ap.add_argument("counts")
+ap.add_argument("--lam", type=float, default=1000)
+ap.add_argument("--data", default="/home/int/run/cifar/vehicles")
+a = ap.parse_args()
+model, cdir = Path(a.model), Path(a.counts)
+k, stride, det_thr, classes, _ = (int(v) for v in (model / "meta.txt").read_text().split())
+
+
+def load(tag, n):
+    parts = sorted(cdir.glob(f"{tag}_*_counts.bin"), key=lambda p: int(p.name.split("_")[-2]))
+    x = np.concatenate([np.fromfile(p, np.int32).reshape(-1, 4 * k) for p in parts]).astype(np.float32)
+    assert len(x) == n, (tag, len(x), n)
+    return x
+
+
+yt = np.array(open(f"{a.data}/train_labels.txt").read().split(), int)
+yv = np.array(open(f"{a.data}/val_labels.txt").read().split(), int)
+xt = np.vstack([load("train", len(yt)), load("train_flip", len(yt))])
+yt2 = np.r_[yt, yt]
+xv = load("val", len(yv))
+mu, sd = xt.mean(0), xt.std(0) + 1e-3
+f = lambda z: np.hstack([(z - mu) / sd, np.ones((len(z), 1), np.float32)])
+T = np.eye(classes)[yt2] * 2 - 1
+for lam in sorted({a.lam, 100.0, 10000.0}):
+    sol = np.linalg.solve(f(xt).T @ f(xt) + lam * np.eye(4 * k + 1), f(xt).T @ T)
+    print(f"lam {lam:g}: ridge on KNP counts, validation {100 * ((f(xv) @ sol).argmax(1) == yv).mean():.2f}%")
+sol = np.linalg.solve(f(xt).T @ f(xt) + a.lam * np.eye(4 * k + 1), f(xt).T @ T)
+w = sol[:-1] / sd[:, None]  # on raw counts
+b = sol[-1] - (mu / sd) @ sol[:-1]
+s = 255 / np.abs(w).max()
+wq = np.rint(w * s).astype(np.int32)  # [4K, classes]
+bq_total = b * s
+bq = np.rint(bq_total / ACTIVE).astype(np.int32)
+drive_t = xt @ wq + bq * ACTIVE  # total drive per image and class
+best = None
+# A class neuron fires at most once a step (about ACTIVE + 1 spikes per image), so the threshold is a fraction
+# of the typical positive per-step drive; spikes are capped accordingly.
+unit = float(np.percentile(drive_t[drive_t > 0], 50)) / ACTIVE
+for q in (0.1, 0.2, 0.3, 0.5, 0.8):
+    thr = int(min(max(unit * q, 1), 65535))
+    spikes = np.minimum(np.floor(np.maximum(drive_t, 0) / thr), ACTIVE + 1)
+    acc = 100 * (spikes.argmax(1) == yt2).mean()
+    print(f"threshold {thr}: train accuracy (rate approximation) {acc:.2f}%, max spikes {spikes.max():.0f}")
+    if best is None or acc > best[0]:
+        best = (acc, thr)
+wq.T.copy().tofile(model / "readout_w.bin")
+bq.tofile(model / "readout_b.bin")
+(model / "meta.txt").write_text(f"{k} {stride} {det_thr} {classes} {best[1]}\n")
+print("written", model, "threshold", best[1])
--- a/changes-2026-10-15.md
+++ b/changes-2026-10-15.md
@ -63,3 +63,11 @@ MNIST, признаки coarse, 30 000 обучение, 10 000 проверка
 - `altai-objects/altai_objects.cpp` (сборка `altai-objects/build.py`): режим `features` — кусочки 12×12 → сетчатка (216 нейронов AltAI, фиксированные веса DoG×1000) → K детекторов STDP-AltAI, взаимное торможение и дофамин от собственного спайка; режим `run` — полная сеть пиксели → сетчатка 6×32×32 → детекторы K×196 (выученные веса минус средний, целые) → выход (дофамин по ошибке), затем веса читаются из бэкенда и фиксируются для проверки.
 - Бэкенд: `KNP_DELTA_DECAY=d` — у нейрона с дофамином D неактивные синапсы меняются на −скорость·D·d (соревновательное обучение).
 - `altai-objects/export_knp.py` (ошибка: многократное чтение npz съело 31 ГБ — исправлено), `score_features.py` — быстрая оценка узоров (только для настройки). Ориентиры (K=64, 6000 обучение): узоры WTA в numpy 68,0%, случайные веса 62,2%; первая проба KNP (5000 кусочков, убавка 0,2) — 60,9%. Сетчатка: в среднем 10% входов спайкают за шаг.
+
+### Обучение узоров в KNP (до уточнения пользователя) и смена курса
+- Соревновательное обучение узоров в KNP (`features`, `altai-objects/feature_grid.sh`, оценка `score_features.py`, K=64, ориентиры: WTA в numpy 68,0%, случайные 62,2%): постоянная убавка (`KNP_DELTA_DECAY`) — 64–65,6% при любых скорости, пороге, торможении (26 вариантов); нормировка порогом AltAI (`KNP_SUMTHR`) — без изменений; сильнее обучение — детекторы замолкают; убавка, пропорциональная весу (`KNP_DELTA_WDECAY`), — 54,8% и очень медленно (55–98 мин на 100 000 кусочков, прогоны остановлены). Длительное торможение (`KNP_DET_NEG`, `KNP_DET_LEAK`, `KNP_SELF_INHIB`): утечка AltAI отводит отрицательный потенциал дальше от нуля при положительном значении; при отрицательном — слишком много спайков.
+- Запечатление (`KNP_IMPRINT=1`: кусочек c учит детектор c через канал дофамина; numpy: K=200 — 71,4% против 73,0% у WTA): в KNP K=200 — 69,6–70,9% (лучшее 70,9%: 10 повторов, скорость 100).
+- **Уточнение пользователя: важен инференс на Алтае, обучение может быть вне платформы.** Курс: обучение обычным кодом, распознавание целиком в KNP.
+- `altai-objects/altai_infer.cpp`: распознавание в KNP — пиксели (спайки) → детекторы K × позиции (целые веса на 6×6×3 пикселей, смещение через постоянно активный вход) → нейроны классов (веса на (четверть, тип детектора), смещение с задержкой 2). Пишет числа спайков детекторов по четвертям (для обучения выхода) и спайки классов.
+- `altai-objects/build_model.py`: k-средние без деления на контраст на кусочках транспорта, свёртка центровки/ZCA/узора в целые веса ±127, порог — 99-й перцентиль. K=200 шаг 2: масштаб 37,6, порог детектора 135, смещения −13…−5. KNP: 0,8 с/картинку, 1,5 ГБ, ~36 тыс. спайков детекторов на картинку.
+- `altai-objects/run_counts.sh`: параллельный сбор чисел спайков из KNP (train, train_flip, val; 14 процессов). `altai-objects/train_readout.py`: ridge на этих числах → целые веса классов ±255, смещение по шагам, порог по обучающей выборке.
--- a/readout/build_delta_cpu.py
+++ b/readout/build_delta_cpu.py
@ -38,6 +38,14 @@ s = s.replace(marker, '''    // Experimental delta rule (KNP_DELTA_RULE=<rate>),
        const char *value = std::getenv("KNP_DELTA_DECAY");
        return value ? std::stof(value) : 0.F;
    }();
+    // Competitive rule of k-means type: with KNP_DELTA_WDECAY=<c>, every synapse of the neuron also changes by
+    // -rate * D * c * weight, so with dopamine after the neuron's own spikes the weight settles where
+    // c * weight equals the probability that its input was active (dw ~ input - c * w).
+    static const float delta_wdecay = []
+    {
+        const char *value = std::getenv("KNP_DELTA_WDECAY");
+        return value ? std::stof(value) : 0.F;
+    }();
    static const float delta_rate = []
    {
        const char *value = std::getenv("KNP_DELTA_RULE");
@ -63,6 +71,9 @@ s = s.replace(marker, '''    // Experimental delta rule (KNP_DELTA_RULE=<rate>),
                    synapse.get().rule_.synaptic_resource_ += delta_rate * neuron.dopamine_value_;
                else if (delta_decay > 0)
                    synapse.get().rule_.synaptic_resource_ -= delta_rate * neuron.dopamine_value_ * delta_decay;
+                if (delta_wdecay > 0)
+                    synapse.get().rule_.synaptic_resource_ -=
+                        delta_rate * neuron.dopamine_value_ * delta_wdecay * synapse.get().weight_;
            }
            if (debug && ++calls % 500 == 0)
                std::fprintf(stderr, "KNP_DELTA_DEBUG calls=%lu lag-3..3: %lu %lu %lu %lu %lu %lu %lu\\n",