27.5% vs 10% markup — different draft bases, need to untangle. Shipping/review time/returns/payment fees all still count; don’t confuse markup w margin Keep the versions separate: the final write-up describes a 27.5% markup on supplier...
vanishing spikes, ugh. check activity layer by layer, then encoding magnitude / inhibition / thresholds instead of changing everything together code reference: Spiking-Inception src/Train.py . Initial training encountered a vanishing-spike...
save weight evolution, not just the last matrix. overly aggressive pruning—check update vs normalization before deciding it's the rule's fault reference code: weights/WeightReadout_B2.py in the Diehl–Cook Brian2 port. Compare saved weight...
unsupervised features, then labels for readout assignment. keep assignment data away from the held-out evaluation set Readout functions: get_neuron_assignments(spike_counts, labels) assigns each neuron to the digit class with its largest mean...
connectivity experiments + SOM tangent. small-world / scale-free / recurrence, and how would a continuous SOM update turn into spike events? january 4 notebook content omitted from the first short collection: initial Brian2 neuron/synapse tests,...
synaptic updates as controls? state / objective / constraints first; random tangent: add a small cost per weight change instead of only caring about the endpoint reading direction: Pontryagin's maximum principle and Hamiltonian-based methods....
linearization + Lyapunov reading. remember the jumps at spike events, not only the nice continuous bit between them study bundle: eigenvalues, Jacobians and Lyapunov functions as tools for connecting the proposed learning rule to stability. Keep...
interesting squiggles ≠ automatically chaos. same input spikes, almost-identical initial states... save that comparison for later study topic: chaotic attractors and the relation between nonlinear network dynamics, robustness and computation....
homeowner side + supplier side, both need to work. what happens when the design needs something the shop doesn’t have?? Maybe offer locally available material options before getting attached to one look two-sided problem: homeowners need a...
history beyond the last pre/post pair. extra state + its timescale need names, “second order” by itself is not enough to implement anything experiment direction: add dependence on past activity rather than using only the immediate pre/post...
perturb one weight, follow what changes. specify the state + scale of the perturbation before staring at the Jacobian approach: perturb synaptic weights and examine the resulting change in neuron/network outputs. The journal explores linear...
eigenvalues / pruning / runaway updates. weight matrix isn't automatically the full Jacobian, especially w/ thresholds + resets in there recorded exploration: eigenvalues of the weight matrix, dynamic pruning of weak connections, and alternative...
predictive coding tangent—what is the neuron actually predicting, and over what time window? Dijkstra vs A* too, need a heuristic + a way to keep distances useful while weights change learning ideas: predictive coding and...
fill the gaps in the run notes before they're impossible to reconstruct. changed parameter + input settings + output files + what broke, all together recorded housekeeping: fill missing experiment descriptions and reorganize notes so parameter...
functional distance → paths → share credit. 1/w needs care w/ zero or signed weights; also if two routes tie, split credit or just take whichever the search finds first? proposal: distribute an STDP update to surrounding significant neurons...
plasticity state: decay between updates, bump on a change. write the actual jump rule down; neighboring updates feeding it back could get out of hand intended state: each synapse has a plasticity value that decays over time and increases when its...
graph context for local STDP. how much extra work per spike before the “selective” update is basically visiting the whole network anyway draft mechanism: represent connections as a graph and use significance/connectivity information to select...
accuracy, spikes/inference, convergence, wall time. keep separate columns—moving to CIFAR also changes the encoding problem Target measures: classification accuracy, computational efficiency and convergence speed. The notes propose spikes per...
local update, distant consequence... still the interesting problem. MAMBA / recurrence / gating ideas to come back to; also eligibility traces, could the synapse hang onto a fading event until credit arrives? question: how can a local...
cloud setup notes need the actual environment + command. tiny local/remote comparison before the long run, and keep compilation time separate recorded change: Google Cloud GPU instances and an initial testing suite were set up. This entry...
design → materials → instructions → actual box of parts. Need to keep those tied to the same revision, esp when Benji changes a dimension product concept: connect a customized outdoor-project design to a material pack and step-by-step...
log firing rates + weight histories too. accuracy alone tells me basically nothing about which layer is going quiet or blowing up recorded work: layer-wise STDP training on MNIST; activity-dependent learning-rate changes; global inhibition and...
STDP constants + handlers together. find where on_post is attached, not just where the string is defined Location: base_model/mnistV1.py , the stdp_eqs , stdp_eqs_pre and stdp_eqs_post strings. The prototype uses taupre = taupost = tau , with...
wrapper: load → encode → run → outputs. keep the model bits swappable, this does not need to become one gigantic script architecture choice: begin with convolutional SNNs for visual datasets while keeping feedforward and recurrent...
wrapper loop check: normalize, encode, run, count, clear, rest. reshape return + capture_period condition look worth slowing down for entry point: base_model/main.py → evalInstance.train(num_examples, target_weight, single_run_time, rest_time,...
upper / lower threshold → separate event routes. check each pathway alone, then alternate; resets could make this weird recorded design: an upper threshold emits a positive event and a lower threshold emits a negative event. Route the event...
wait, check sPlus / sMinus against the predicates. old s → new s → event check, names alone are not helping here archived model state: v , u , v0 , threshold state s and previous threshold state s_pre . neuronUpdates() runs with...
network_operation for the extra state updates. need the scheduling order written down, not just the function body code/API pointer: NetworkOperation / network_operation . The notebook uses this to explore updates outside the normal...
sigmoid response is tempting, but where exactly does the discrete spike happen. threshold adaptation might be the other route—check timing + cost variants explored: a smooth sigmoid-response neuron and an exponentially adapting threshold. The...
positive + negative events without automatically making two whole populations? keep event sign, synapse sign and voltage separate... too many different meanings of “negative” here experiment: vary membrane time constants, reset rules and...
tiny neuron tests 🙂 easier to follow one trace than the whole network. check which declared state variables actually appear in the equation neuron-models/logistic_test.py: a single dimensionless neuron uses dv/dt = (0.1*v*(I-v))/tau , with...
TSC → neuron model, watch the ones that keep firing. indices and times need to stay paired when replacing a sample's spikes recorded changes: TSC input encoding connected to the neuron model; spike-time resolution and decay rates adjusted;...
sign from swapping the two event times, nice. draw pixels either side of a quantization boundary later; want to see how suddenly that pair moves Reference: base_model/mnistV1.py → temporal_switch(img, quant_levels=10) . The archived experiment...
relative spike timing rather than just rate. keep encoder / neuron / learning rule separate while messing w/ the time constants study direction: use relative spike timing rather than only average rate. The notebook adapts the LIF baseline by...
MNIST vs actual event data—encoding is part of the experiment. inspect the spike trains! also if every spike shifts by the same amount, does the representation change or only its timing? Candidates: MNIST for a simple static-image baseline;...
11 hours for ~8%... yeah this needs looking at. save neuron count, dt, presentation time + backend with the runtime, otherwise the numbers aren't much use recorded observations: after more than 11 hours the simulation was about 8% complete; a...
CUDA setup first, tiny example first. not dragging the entire network into compiler-error land yet failure: the Brian2CUDA example did not compile in the current environment. The October 15 entry frames CUDA/environment setup as the next...
equations + constants here so i stop hunting for them. don't skip the reset / refractory / inhibitory paths just bc the learning rule is more interesting Neuron equation recorded in the reference sheet: tau*dV/dt = (e_rest - V) + g_e*(e_exc -...
keep the Diehl–Cook port nearby while changing the rule. trace input → conductance → spikes → updates, then compare Chosen reference: Diehl–Cook 2015 Brian2 port . The journal's occasional “2005” wording conflicts with the...
need ONE baseline i can follow all the way through. check normalization + readout before getting distracted by another repo candidate task: MNIST recognition with STDP in Brian2. The notebook compares discussion-forum implementations, older paper...
Brian2 once again. need the equations + custom events exposed, the convenient abstractions are kind of in the way here options considered: Brian2, PyNN, BindsNET and snnTorch. The selection criterion in the notebook is access to neuron equations,...
ssh / XRDP / X11 / matplotlib... quite the little gremlin convention. sort out the display path separately from whether the simulation runs environment notes: PyNN and Brian2 installed in a virtual environment on the remote GPU-enabled machine....
unsupervised STDP, see how far this can go w/ depth. also catastrophic forgetting—more layers running ≠ more useful features learned starting direction: unsupervised STDP, with the project journal and deadlines established on September 15....
interviews: ask about current problems BEFORE pitching HomeBuilt. “would use this” vs actually paying the pack price are very different answers, keep question versions + response exports two stages in the survey notes: pain-probing interviews...
PDB chain → TreeFam sequence → PyMOL, need the residue numbering to survive all 3. Missing residues / partial structures are the annoying bit here. Save for later: same highlight following a residue between sequence and 3D views? recorded...
scope / impact / niche / revenue... need a tiny test, not another giant idea list. Save rejected ideas w the assumption that killed them. otherwise I’ll rediscover the same thing in 3 months :) archive contents: an early idea brainstorm...
same 23→72.83 endpoints, two different percentage-gain claims in the posters. ugh recompute from the raw numbers; keep scoring definition + denominator nearby. percentage points and percent gain are not the same thing Engineering-plan targets...
poster recipe: 3 conv, 2 ReLU, flatten, 2 dense; lr 0.001 / 128 epochs / L2 missing widths, batch size, L2 coeff + exact loss path. save the checkpoint/log next to the diagram if found The poster supplies more concrete model details than the...
513 checks → 115, nice need runtime + input size too. fewer visited voxels doesn't automatically mean the same factor of speedup The tiny complexity sheet records exhaustive = 513 and tracing = 115 iterations. The ratio is 513/115 ≈ 4.46 ;...
script → artifact map, otherwise i'm going to lose track of what produced what FAISS loss sketch detaches to numpy... where are the gradients supposed to go? keep embedding training separate until that is answered Stage Script named in...
