Neocognitron
No estimate
No hardware requirements for this model
The weights for this model have not been published, so it cannot be downloaded or run on your own hardware at any size. It is reachable only through its provider, and no graphics card changes that.
On record
Full specification
Everything on record for this model. Most of it describes how it was trained rather than how it runs — useful context for judging how much work went into it, and how it compares with models built at a different scale.
Origin
Who built this model, where, and when it was published.
- Organisation
- NHK Broadcasting Science Research Laboratories
- Organisation type
- Industry
- Country
- Japan
- Published
- 1 April 1980
- Authors
- K Fukushima, S Miyake
What it does
The problem areas the model was built for. A model can carry several of each.
- Domain
- Vision
- Task
- Character recognition (OCR)
Size
How large the model is and how much data it was trained on. Parameters are the figure that decides whether it fits on a given graphics card.
- Parameters
- 1.1M
- Training data
- 5 tokens
"The synaptic connections from S-layers to C-layers are fixed and unmodifiable. [...] The numbers of excitatory cells in these seven layers are: 16x16 in U0, 16x16x24 in Us1, 10x10x 24 in Uc1, 8x8x24 in Us2, 6x 6x 24 in Uc2, 2x2x24 in Us3, and 24 in Uc3 [...] the number of input synapses to each S-cell is 5 x 5 in layer Us1 and 5x5x24 in layers Us2 and Us3 [...] The number of excitatory input synapses to each C-cell is 5x5 in layers Uc1 and Uc2, and is 2x2 in layer Uc3 " The number of synapse…
"In order to self-organize the network, we have presented five stimulus patterns "0", "1", "2", "3", and "4", which are shown in Fig. 6"
Training compute
The arithmetic performed to train the model, measured in floating-point operations. It is a measure of what the training run cost, not of how fast the finished model answers you.
- Training compute
- 2.7 × 10⁸ FLOP
- How it was established
- Operation counting
"It does not necessarily mean that all of these input synapses are always fully reinforced. In usual situations, only some of these input synapses are reinforced, and the rest of them remains in small values [...] Each of the five stimulus patterns has been presented 20 times to the network. By that time, self organization of the network has almost been completed." We multiply by 2 to account for multadds in the forward pass. There is no real backward pass, weights are only updated sparsely. E…
How it is classified
Labels the source dataset applies when tracking notable models, and how confident it is in the entry.
- Frontier model
- Yes
- Why it is tracked
- Highly cited
- Record confidence
- Confident
- Citations
- 5,782
Sources
Where this record came from and when it was last checked.
- Reference
- Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position
- Last updated
- 28 November 2025
What the numbers mean
Background
Neocognitron was published by NHK Broadcasting Science Research Laboratories, in Japan, in April 1980. It comes out of industry.
It works in Vision, and is recorded as doing character recognition (OCR).
This is a closed model: the trained values stayed with whoever produced them, and there is no local version to run.
How it was trained
The training run consumed about 2.7 × 10⁸ FLOP. That figure describes the cost of creating it and has no bearing on how quickly it generates text.
The training set ran to roughly 5 tokens.
The reason it appears in this catalogue at all is highly cited.
Answers
Neocognitron — common questions
What GPU do I need to run Neocognitron?
None. Neocognitron is a closed model — its weights were never published, so it cannot be downloaded or run on your own hardware at any price. It is reachable only through its provider.
Is Neocognitron open source?
The licensing for Neocognitron was never recorded in our source data. We treat unstated licensing as closed, because an unrecorded licence is not one to rely on.
How many parameters does Neocognitron have?
Neocognitron has 1.1M parameters. "The synaptic connections from S-layers to C-layers are fixed and unmodifiable. [...] The numbers of excitatory cells in these seven layers are: 16x16 in U0, 16x16x24 in Us1, 10x10x 24 in Uc1, 8x8x24 in Us2, 6x 6x 24 in Uc2, 2x2x24 in Us3, and 24 in Uc3 [...] the number of input synapses to each S-cell is 5 x 5 in layer Us1 and 5x5x24 in layers Us2 and Us3 [...] The number of excitatory input synapses to each C-cell is 5x5 in layers Uc1 and Uc2, and is 2x2 in layer Uc3 " The number of synapses into each S-layer is: S1: (16*16*24)*(5*5) S2: (8*8*24)*(5*5*24) S3: (2*2*24)*(5*5*24) We assume one parameter a per synapse into each cell in a S-layer, and one parameter b per each cell in a S-layer. That figure is the total, and it is what decides how much memory the model needs — roughly half a gigabyte per billion at the compression most people use.
Who created Neocognitron?
Neocognitron was published by NHK Broadcasting Science Research Laboratories, based in Japan, categorised as industry.
When was Neocognitron released?
Neocognitron was published in April 1980. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.
What is Neocognitron used for?
Neocognitron works in Vision, and is recorded as handling character recognition (OCR). A model can carry several of each, so these are the areas it was built for rather than a limit on what it will attempt.
How much compute was used to train Neocognitron?
Around 2.7 × 10⁸ FLOP. That measures what producing the model cost and says nothing about how quickly it answers once trained — inference speed comes from memory bandwidth, not from the training budget.
The other direction
Looking at it from the other side?
This page starts from the model. If you already own a card and want to know everything it will run, start from the hardware instead.