IMPORTANT: This project is still an experimental prototype and NOT ready for any productive usage. There are still many missing tests, input-validations and so on. Beside this there is also still quite a lot of evaluation and improving of the current features necessary.
Hanami contains in its core a custom experimental artificial neural network, which can work on unnormalized and unfiltered input-data, like sensor measurement data. The network growth over time by creating new nodes and connections between the nodes while learning new data. The base concept was created by myself and the code was written from scratch without any frameworks. The goal behind Hanami is to create something unique, which works more like the human brain. It wasn't targeted to get a higher accuracy than classical artificial neural networks like Tensorflow, but to be more flexible and easier to use and more efficient in resource-consumption for big amounts of inputs and users. Additionally it also provides an as-a-Service architecture within a cloud native environment and multi-tenancy.
| Backend | SDK | Deployment |
|---|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
|
 |
||
 |
||
 |
||
 |
||
 |
||
 |
||
 |
||
 |
-
Growing neural network:
The artificial neural network, which is the core of the project, growth over time while learning new things by creating new nodes and connections between the nodes based on the given input. A resize of the network is also quite linear in complexity.
-
No normalization of input
The input of the network is not restricted to range of 0.0 - 1.0 . Every value, as long it is a positive value, can be inserted. Also if there is a single broken value in the input-data, which is million times higher, than the rest of the input-values, it has nearly no effect on the rest of the already trained data. Thanks to the reduction-process, all synapses, which are only the result of this single input, are removed again from the network.
-
Parallelism
The processing structure works also for multiple threads, which can work at the same time on the same network. (GPU-support with CUDA is disabled at the moment for various reasons).
-
Usable performance
The 60.000 training pictures of the MNIST handwritten letters can be trained on CPU in about 3 seconds for the first epoch, without any batch-processing of the input-data and results in an accuracy of 91-93 % after this time.
-
Generated OpenAPI-Documentation
The OpenAPI-documentation is generated directly from the code. So changing the settings of a single endpoint in the code automatically results in changes of the resulting documentation, to make sure, that code and documentation are in sync.
See OpenAPI-docu
-
Multi-user and multi-project
The projects supports multiple user and multiple projects with different roles (member, project-admin and admin) and also managing the access to single api-endpoints via policy-file. Each user can login by username and password and gets an JWT-token to access the user- and project-specific resources.
-
Efficient resource-usage
-
The concept of the neural network results in the effect, that only necessary synapses of an active node of the network is processed, based on the input. So if only very few input-nodes get data pushed in, there is less processing-time necessary to process the network.
-
Because of the multi-user support, multiple networks of multiple users can be processed on the same physical host and share the RAM, CPU-cores and even the GPU, without splitting them via virtual machines or vCPUs.
-
Capability to regulate the cpu-frequencey and measure power-consumption. (disabled currently)
See Monitoring-docu
-
-
Network-input
There are 2-variants, how it interact with the neural networks:
-
Uploading the dataset and starting an asynchronous task based on this dataset over the API
See OpenAPI-docu
-
Directly communicate with the neural network via websocket. In this case not a whole dataset is push through the synapse, but instead only a single network-input is send. The call is blocking, until the network returns the output, which gives more control.
See Websocket-docu
-
-
Installation on Kubernetes and with Ansible
The backend can be basically deployed on kubernetes via Helm-chart or plain via Ansible.
-
No strict layer structure
The base of a new neural network is defined by a cluster-template. In these templates the structure of the network in planed in hexagons, indeed of layer. When a node tries to create a new synapse, the location of the target-node depends on the location of the source-node within these hexagons. The target is random and the probability depends on the distance to the source. This way it is possible to break the static layer structure. But when defining a line of hexagons and allow nodes only to connect to the nodes of the next hexagon, a classical layer-structure can still be enforced.
-
Reduction-Process
The concept of a growing network has the result, that there is basically nearly no limit in size, even if the growth-rate slows down over time. To limit the growth-rate even more, it is possible to enable a reduction-process, which removes synapses again, which were to inactive to reach the threshold to be marked as persistent.
-
Spiking neural network
The concept also supports a special version of working as a spiking neural network. This is optional for a created network and basically has the result, that an input is impacted by an older input, based on the time how long ago this input happened.
-
3-dimensional networks
It is basically possible to define 3-dimensional networks. This was only added, because the human brain is also a 3D-object. This feature exist in the cluster-templates, but was never tested until now. Maybe in bigger tests in the future this feature could become useful to better mix information with each other.
The concept is not perfect and also has some disadvantages, which are the result of the architecture itself:
-
Very inefficient for binary input, where inputs of single nodes of the network can only be 0 or 1, without values in between
-
A single synapse needs more memory than in a classical network. The hope is, in bigger tests, it becomes much more efficient compared to fully meshed layered networks.
All objects are automatically build and uploaded by the
CI-pipeline for each merge
on develop-branch and for each tag.
There are some features, which existed in the past, were disabled temporary and will be added/enabled again in the near future:
-
Dashboard
As a PoC a first dashboard was created, without any framework. It is planned to refactor this old version in
v0.9.0and re-write it again with Typescript and some additional frameworks. Until then, it is temporary disabled, because it would current cost too much time to keep this unused and prototypical version up-to-data. As reference see the example-workflow of the PoC-dashboard: Dashboard-docu -
Regulation of CPU-speed
Also in older version there also was the function available to regulate the speed of the CPU based on the workload. The dashboard was used to visualize the CPU metrics like the speed. Since the dashboard was disabled, there is at the moment not feedback available, so for usability reasons the feature was not further maintained and disabled for now. It is planned again for version
v0.8.0. -
GPU-support
There already were some attempts in the past to add GPU-support with CUDA and OpenCL in the past. Some version like 0.4.0 also had a working version implemented. The problem was disappointing performance and some restrictions for the CPU-version too. There will be some further attempts in the future, to fix this issue and bring GPU support back into the project, but because there is no definite solution now, it is unknown when this happens.
see Roadmap
Tobias Anker
eMail: tobias.anker@kitsunemimi.moe
The complete project is under Apache 2 license.