An abstract entropy arising from computations of deep networks. A generative force defined on the deep layers — a discovery of how depth behaves and what can be computed.
Amtropy denotes an abstract entropy that arises from computations of deep networks. It is entropic because amtropy is defined following thermodynamic rules — a metric observed consistently from well-formed networks.
It is a generative force defined on the deep layers of stacked RBMs, a behavior of patterns related with the confidence level of them. It helps to choose better models among many options, or it is a manifestation of emotion or sense of goal to be pursued by the agent.
A kernel solution implementing Amtropy theory with supporting algorithms — built to be a building block of scalable, real-time systems.
Implements Amtropy theory with BP, softmax, and thoughtful utilities in a plug-and-play style for real-time applications. ~100K lines of C code.
Deeper layers improve accuracy beyond DBN architectures. Entropic monotonicity — entropy consistently decreases in higher layers, yielding greater abstraction (𝜙).
Designed as the core tool to build foundational LLMs and other critical projects. Scalable systems constructed fast and running efficiently.
Our discovery is that deeper layers improve accuracy, and a consistent metric in the depth of layers is the entropic monotonicity — entropy keeps decreasing in higher layers.
The inverse of amtropy is defined as the degree of "abstraction" symbolized as 𝜙. The volume metric (𝜒) also decreases along the depth. 99 hidden layers working together, consistently improving accuracy and abstraction.
The implications expand perspectives on modelling networks and open new applications based on the computation of emotions and desires as well as optimal generations.
Building scalable AI solutions and partnerships for real-world applications.
Providing the AM kernel solution to institutes who can benefit from modular AI kernels to solve real-time data streams.
Building joint researches of difficult problems with accountable partners. Collaborating on foundational AI challenges.
A financial trading system running on AM kernels — successfully proven as an alpha version, opening valuable business opportunities in financial services.
We welcome accountable partners who share our vision of understanding deep layers and building scalable AI solutions.
yshan@amtropy.comContact only if you are accountable.