KAN Spline

as donto believed on

From Dontopedia, the open, paraconsistent wiki. (Last updated 2026-06-06.)

KAN Spline has 78 facts recorded in Dontopedia across 10 references, with 7 live disagreements.

78 facts·61 predicates·10 sources·7 in dispute

Mostly:involves operations(7), rdf:type(3), has formula(2)

Inbound mentions (8)

Other subjects in dontopedia point AT this entity as a value. These are inverse relationships — e.g. "X motherOf this subject" — and answer questions the forward facts can't. Grouped by predicate.

`enablesPowerEmergence`Enables Power Emergence(1)

Higher Lrex:higher-lr

`expects`Expects(1)

Gen 4 8ex:gen-4-8

`hasMember`Has Member(1)

Activation Poolex:activation-pool

`impliesSignificanceIfWinnerIs`Implies Significance If Winner Is(1)

Gen 9 16ex:gen-9-16

`includesActivation`Includes Activation(1)

Activation Poolex:activation-pool

`includesCompetingActivations`Includes Competing Activations(1)

New Configex:new-config

`introducesNewActivation`Introduces New Activation(1)

Alpha Symbiogenesis Update 2026 02 25ex:alpha-symbiogenesis-update-2026-02-25

`member`Member(1)

Activation Varietyex:activation-variety

Other facts (75)

The long tail: predicates that appear too rarely to warrant their own section. Filter or scroll to find a specific one. Each row links to its source.

75 facts

Predicate	Value	Ref
Involves Operations	Matmul	[8]
Involves Operations	Silu Mul	[8]
Involves Operations	Relu Mul	[8]
Involves Operations	Gelu Mul	[8]
Involves Operations	Id Mul	[8]
Involves Operations	X Squared Mul Mul	[8]
Involves Operations	Add	[8]
Rdf:type	Activation Function	[6]
Rdf:type	Activation Function	[8]
Rdf:type	Activation	[9]
Has Formula	f(x) = c0silu(x) + c1relu(x) + c2gelu(x) + c3x + c4*x²	[1]
Has Formula	f(x) = c0silu(x) + c1relu(x) + c2gelu(x) + c3x + c4*x²	[6]
Compatible With Backend	Cpu	[1]
Compatible With Backend	Helios Gpu	[1]
Averages	Silu	[2]
Averages	Gelu	[2]
Starts As	(silu+gelu)/2	[2]
Starts As	(silu+gelu)/2	[4]
Has Learnable Params	5 vectors/layer	[3]
Has Learnable Params	c0-c4	[8]
Expected to Have	higher initial loss	[9]
Expected to Have	steeper learning curve	[9]
Is Inspired by	Kolmogorov Arnold Networks	[1]
Uses Basis Functions	existing activation basis functions instead of B-splines	[1]
Substitutes Basis	null	[1]
Adapts Kan	null	[1]
Has Basis Functions	5-basis function approximator	[1]
Requires No New Kernels	null	[1]
Broadcasts With	[B*T, ffnDim] hidden activations	[1]
Uses Param Shape	[1, ffnDim]	[1]
Uses Learnable Coefficients	per channel	[1]
Is Approximator	null	[1]
Power Emerges Only in	Later Generations	[2]
Is Essentially	Fixed Starting Points	[2]
Has Uncertainty As Wildcard	true	[2]
Is Wildcard	true	[2]
Has Learnable Params That Barely Move	true	[2]
Predicted As Wildcard Competitive	true	[2]
Essentialy Universal Approximator	true	[3]
Is5 Basis Kan Approximator	true	[3]
Defined by Formula	f(x) = c0silu + c1relu + c2gelu + c3x + c4*x²	[3]
Can Learn Any Activation Shape	true	[3]
Most Versatile Activator	true	[3]
Can Learn All Other Activations	true	[3]
Does Not Get Faster	true	[4]
Ffn Forward Pass Sequence	matmul → silu+mul, relu+mul, gelu+mul, id·mul, x²·mul·mul → 4×add → matmul	[4]
Has Fixed Tok Per Second	true	[4]
Has Learnable Coefficient Vectors	5	[4]
Has Lower Tok Per Second	Simple Activations	[4]
Has Op Count Per Ffn Forward Pass	14	[4]
Has Ops Multiplier Relative to Gelu	4.5	[4]
Improves Loss Per Step	true	[4]
Includes Add Operations	4×	[4]
Known to Have C0 to C4 Params	true	[4]
Learns Basis Functions Per Channel	true	[4]
Optimizes Learnable Params Via Backprop	true	[4]
Does Not Update	Anchor Positions Relative	[5]
Updates	Coupling Shape	[5]
Has Description	5-basis function approximator	[6]
Inspired by	Kolmogorov-Arnold Networks	[6]
Uses Approach	existing activation basis functions	[6]
Contrasts With	B-splines	[6]
Has Learnable Coefficients	true	[6]
Param Change During Warmup	Minimal Movement	[7]
State During Warmup	Fixed Starting Point	[7]
Status	Wildcard	[7]
Starting Value Expression	(silu+gelu)/2	[7]
Competitiveness	Surprisingly Competitive	[7]
Has Speed Over Time	Fixed Speed	[8]
Has Op Count	14	[8]
Has Extra Learnable Vectors	5	[8]
Initializes As	(silu+gelu)/2	[8]
Learns Feature	which basis functions matter per channel	[8]
Updates Coupling Shape	true	[10]
Does Not Update Anchor Positions	true	[10]

Timeline axis is valid_time — when each source says the fact was true in the world, not when Dontopedia learned about it. Retracted rows are kept for provenance; coloured stripes indicate the context kind.

78 undated

—

hasFormulablah/training-and-evals/part-21

f(x) = c0*silu(x) + c1*relu(x) + c2*gelu(x) + c3*x + c4*x²

—

isInspiredByblah/training-and-evals/part-21

ex:kolmogorov-arnold-networks

—

compatibleWithBackendblah/training-and-evals/part-21

ex:cpu

—

usesBasisFunctionsblah/training-and-evals/part-21

existing activation basis functions instead of B-splines

—

substitutesBasisblah/training-and-evals/part-21

null

—

compatibleWithBackendblah/training-and-evals/part-21

ex:helios-gpu

—

adaptsKANblah/training-and-evals/part-21

null

—

hasBasisFunctionsblah/training-and-evals/part-21

5-basis function approximator

—

requiresNoNewKernelsblah/training-and-evals/part-21

null

—

broadcastsWithblah/training-and-evals/part-21

[B*T, ffnDim] hidden activations

—

usesParamShapeblah/training-and-evals/part-21

[1, ffnDim]

—

usesLearnableCoefficientsblah/training-and-evals/part-21

per channel

—

isApproximatorblah/training-and-evals/part-21

null

—

powerEmergesOnlyInblah/training-and-evals/part-24

ex:later-generations

—

isEssentiallyblah/training-and-evals/part-24

ex:fixed-starting-points

—

hasUncertaintyAsWildcardblah/training-and-evals/part-24

true

—

averagesblah/training-and-evals/part-24

ex:silu

—

isWildcardblah/training-and-evals/part-24

true

—

hasLearnableParamsThatBarelyMoveblah/training-and-evals/part-24

true

—

averagesblah/training-and-evals/part-24

ex:gelu

—

predictedAsWildcardCompetitiveblah/training-and-evals/part-24

true

—

startsAsblah/training-and-evals/part-24

(silu+gelu)/2

—

essentialyUniversalApproximatorblah/training-and-evals/part-23

true

—

is5BasisKanApproximatorblah/training-and-evals/part-23

true

—

definedByFormulablah/training-and-evals/part-23

f(x) = c0*silu + c1*relu + c2*gelu + c3*x + c4*x²

—

canLearnAnyActivationShapeblah/training-and-evals/part-23

true

—

mostVersatileActivatorblah/training-and-evals/part-23

true

—

canLearnAllOtherActivationsblah/training-and-evals/part-23

true

—

hasLearnableParamsblah/training-and-evals/part-23

5 vectors/layer

—

startsAsblah/training-and-evals/part-28

(silu+gelu)/2

—

doesNotGetFasterblah/training-and-evals/part-28

true

—

ffnForwardPassSequenceblah/training-and-evals/part-28

matmul → silu+mul, relu+mul, gelu+mul, id·mul, x²·mul·mul → 4×add → matmul

—

hasFixedTokPerSecondblah/training-and-evals/part-28

true

—

hasLearnableCoefficientVectorsblah/training-and-evals/part-28

—

hasLowerTokPerSecondblah/training-and-evals/part-28

ex:simple-activations

—

hasOpCountPerFfnForwardPassblah/training-and-evals/part-28

—

hasOpsMultiplierRelativeToGelublah/training-and-evals/part-28

4.5

—

improvesLossPerStepblah/training-and-evals/part-28

true

—

includesAddOperationsblah/training-and-evals/part-28

4×

—

knownToHaveC0ToC4Paramsblah/training-and-evals/part-28

true

—

learnsBasisFunctionsPerChannelblah/training-and-evals/part-28

true

—

optimizesLearnableParamsViaBackpropblah/training-and-evals/part-28

true

—

doesNotUpdateblah/watt-activation/part-220

ex:anchor-positions-relative

—

updatesblah/watt-activation/part-220

ex:coupling-shape

—

typeblah/training-and-evals/21

ex:ActivationFunction

—

labelblah/training-and-evals/21

KAN Spline

—

hasDescriptionblah/training-and-evals/21

5-basis function approximator

—

hasFormulablah/training-and-evals/21

f(x) = c0*silu(x) + c1*relu(x) + c2*gelu(x) + c3*x + c4*x²

—

inspiredByblah/training-and-evals/21

Kolmogorov-Arnold Networks

—

usesApproachblah/training-and-evals/21

existing activation basis functions

—

contrastsWithblah/training-and-evals/21

B-splines

—

hasLearnableCoefficientsblah/training-and-evals/21

true

—

paramChangeDuringWarmupblah/training-and-evals/24

ex:minimal-movement

—

stateDuringWarmupblah/training-and-evals/24

ex:fixed-starting-point

—

statusblah/training-and-evals/24

ex:wildcard

—

startingValueExpressionblah/training-and-evals/24

(silu+gelu)/2

—

competitivenessblah/training-and-evals/24

ex:surprisingly-competitive

—

labelblah/training-and-evals/28

KAN spline

—

typeblah/training-and-evals/28

ex:ActivationFunction

—

hasSpeedOverTimeblah/training-and-evals/28

ex:fixed-speed

—