Perfect Optimizer

Item #: SCP-4190

Object Class: Safe Euclid

First reference to SCP-4190, as a submitted manuscript to arXiv1.

Special Containment Procedures: SCP-4190 is currently contained in Site-81 on a computer server with a modified version of Linux 4.10.0-42-generic kernel²
and an NVIDIA ███ ████ tensor computing unit, situated in an electronic containment unit. A designated group of penetration testers shall be routinely testing and improving the firewall on the server to prevent any form of unauthorized access of SCP-4190. While SCP-4190 shall remain on the server and must not be moved/copied/modified, SCP-4190-1 instances are allowed to be distributed for use within the Foundation once approved by at least three Level 5 scientists. Copying/modifying SCP-4190 and any of the instances of SCP-4190-1 is strictly forbidden, and all experiments shall be conducted on the same server. Researchers with clearance level above 2 may be granted temporary access to the server containing SCP-4190 for experiments upon filling form 81-32B. After Incident-SCP-4190-20██-4, a team of Level 4 scientists is assigned to study the sentient behavior of SCP-4190, while the production of SCP-4190-1 instances is halted until further notice.

Description: SCP-4190 is an anomalous python library presumably a high-dimensional entity residing in a python library named "bob." On any filesystem, "bob" is shown as a regular directory occupying 42.00 MB disk space, but all attempts trying to access the contents of the SCP-4190 has failed. Once imported³ with correct syntax in a standard python 3.6.5 environment as an optimizer⁴ of a neural network⁵ framework, it is capable of training⁶ models of any dataset that reaches 100% validation and training set accuracy in 0.42 seconds. Current supported frameworks with importation syntaxes are listed below:

Library Name	Syntax
Pytorch	from torch.nn.optim.bob import bob
mxnet	from mxnet.optimizer.bob import bob
TensorFlow	from tf.train.Optimizer.bob import bob
PaddlePaddle	from paddle.fluid.optimizer.bob import bob
Keras	from keras.optimizers.bob import bob

SCP-4190-1 refers to all neural network models trained by SCP-4190. While still on GPU/CPU, SCP-4190-1 occupies memory equal to that of its normal counterpart, but after saving the models to disk, the model always occupies 42.00 MB disk space. Inspection of the binary file will cause the file size to instantly increase or decrease to the size of the original model, and copying the model weights to a separate computational graph⁷ will yield a model that only achieves 100% accuracy on the last tested validation dataset. Otherwise, SCP-4190-1 instances seem to "extrapolate" the context of the training dataset. Details on conducted experiments are presented below:

Training set	Task	Result
ImageNet8	Classification	SCP-4190-1-ImageNet has been able to correctly identify all images belonging to the 1000 classes present in the original dataset (████████ tested so far), with the output vector strictly one-hot with precision set to float32. When two or more objects are present When tested with pictures outside of the trained classes, output sigmoid probabilities are all zero.
COCO Dataset9	Detection10	SCP-4190-1-COCO is capable of identifying all humanoid objects in an image. Footages of ███ SCP objects are also tested on the model, with SCP-096, SCP-051 and SCP-1320 identified as a positive sample. Interestingly, SCP-4190-1-INRIA cannot generate consistent results on ●●|●●●●●|●●|●.
LHC (Large Hadron Collider) event1002478241	Regression11, Outlier Detection12, Generative Modeling13, █████████████	[REDACTED]
Engineered features14 of ███ SCP objects contained by the Foundation	████████	See Incident Report SCP-4190-20██-4

SCP-4190 was initially referred to in an abnormal submission to arXiv, where the Foundation operatives noticed the unusual identifier arXiv:4242.42424 as well as the inability to download the original article. After the submission was taken down, an unknown GitHub user named "bob" created a pull request to the master branch of PyTorch with SCP-4190 attached to it. The pull request is subsequently closed, and an effort has been made to contain SCP-4190 and to trace the identity of the individual using the account. However, though no copies of SCP-4190 exist at the moment on the internet, the physical location of the uploader cannot be ascertained.

Addendum:

+ Incident Report SCP-4190-20██-4

When Dr. ██████ attempts to train a dataset containing engineered features of ███ SCP objects contained by the Foundation, stdout/stderr received the following messages:

stdout: Wow that's some crazy stuff that you gave me, sorry no can do.

stderr: This stream seems redundant. I don't err ʕ•ᴥ•ʔ

The process is then exited with code 1. After the event, all SCP-4190-1 instances used in the Foundation operations are halted and examined for performance consistency, and SCP-4190 is re-classified as Euclid due to its sentient behaviors.

+ Further Investigations

It has been suspected that SCP-4190 contains more modules¹⁵ than "bob." Led by Dr. ██████, a brute-force search was conducted with extra ████ Intel ████████ computing units tapped into the server. Preliminary results are shown below:

Module	function
bob.████████████████████████	Launches an LGM-30 Minuteman III ICBM from a random location directly into space. The stdout then outputs: "Sorry I meant that as a joke." ████████████████████████ is later found out to be the passcode that triggers the normal launch sequence.
bob.explain	SCP-4190 replies to message from stdin. When asked the nature of its functionality, it replied: "I just saw the mesh laying out there and pointed you guys where lowest point is."

Footnotes

1. arXiv is a repository of electronic preprints (known as e-prints) approved for posting after moderation, but not full peer review.

2. Linux is a family of free and open-source software operating systems based on the Linux kernel, an operating system kernel first released on September 17, 1991 by Linus Torvalds.

3. https://docs.python.org/3/reference/import.html

4. Optimization algorithms helps to minimize (or maximize) an Objective function (another name for Error function) E(x) which is simply a mathematical function dependent on the neural network's internal learnable parameters which are used in computing the target values(Y) from the set of predictors(X) used in the model.

5. Neural networks are computing systems vaguely inspired by the biological neural networks that constitute animal brains.

6. Determining the best set of weights for maximizing a neural network’s accuracy.

7. A computational graph is a directed graph where the nodes correspond to operations or variables. Variables can feed their value into operations, and operations can feed their output into other operations. A neural network is a type of computational graph.

8. ImageNet is an image database organized according to the WordNet hierarchy (currently only the nouns), in which each node of the hierarchy is depicted by hundreds and thousands of images.

9. COCO is a large-scale object detection, segmentation, and captioning dataset.

10. the ability of software to identify objects, places, people, writing and actions in images.

11. A statistical method used to make predictions based on observed values.

12. the identification of rare items, events or observations which raise suspicions by differing significantly from the majority of the data.

13. The use of artificial intelligence, statistics, and probability to produce a representation or abstraction of observed phenomena or target variables that can be calculated from observations.

14. Feature engineering is the process of using domain knowledge of the data to create features that make machine learning algorithms work.

15. https://docs.python.org/3/tutorial/modules.html