Building a more accurate and sustainable AI model

I recently had the opportunity to speak with the founders of a company called PiLogic about their approach to solving certain problems that they say is faster and less energy-intensive than large language models (LLMs). can be solved with the consumption of Their approach makes heavy use of exact probability estimation. PiLogic says their inference engine is the most advanced in the world as benchmarked against Join Tree and other leading methods.

PiLogic is also submitting an application to join the Green Computing Working Group of the International Telecommunication Union (ITU). They believe their methods can be useful for many general information and computing technology (ICT) applications.

This approach does not require large data sets and specialized expensive hardware such as graphics processing units (GPUs). It is of particular importance for engineering use cases, it is non-illusory and gives results that are precise and accurate. It's currently targeted for use in aerospace and cyber security applications, but the company believes it could become a standard AI toolkit where one needs answers mathematically, where mistakes are costly. and where the results need to be adjusted according to expert knowledge.

Some of the use cases are (1) autonomous systems, such as autonomous flight, (2) cyber security, such as Security Operations Center (SOC) flag management and automated threat prediction and response, and (3) aerospace, e.g. Identification and Tracking. by radar, and diagnosis and prediction of electrical system faults on aircraft and spacecraft. The inference engine and AI toolkit can be applied to many complex problems in industries such as finance, energy, cloud and healthcare. The figure below shows the PiLogic process flow that includes a Bayesian network and an evidence-based inference engine.

The PiLogic engine works on what are called Bayes Nets, which have many advantages over other types of models. For example, they can incorporate expert knowledge, handle limited training data, and facilitate analysis of why the model behaves as it does. One of the techniques used in the PiLogic engine is to generate an efficient arithmetic circuit (AC) from a Bayes Net. The figure below shows the dependency in AC generated from Baynes Net.

One of the reasons why AC is useful is that it outsources most of the work involved in the pre-deployment phase, which only needs to be run once. After deployment, pre-deployment work can be executed on a large number of queries. The second reason is that post-placement assessment answers multiple questions simultaneously.

Apart from efficiency, AC has other advantages. For example, it is possible to know how much time and space is required to answer questions, and so this approach works well in the context of real-time requirements. Additionally, AC can be embedded in many products and applications because it does not require specialized hardware. This improvement in performance also leads to energy savings for the entire estimation process on a continuous basis for the end users.

In the chart at the top of the article, the “width” of a Bayesian network on the horizontal axis reflects how difficult a network is to a traditional inference engine. Conventional inference engines run in time and space that is exponential to this width and consequently only work on networks of finite width, as shown below.

PiLogic says it has found a way to break this rapid growth in computational complexity for many problems. They do this by using the structure in the problem, specifically the spatial structure. These can be zero or repeating values ​​in the model which can simplify the required calculations. As a result, PiLogic says that if there is enough spatial structure, they can solve problems with tree widths of 100, as shown above. Note that if the model has no such structure, the PiLogic engine will have the same width constraints as traditional inference engines.

Being able to deal with problems of greater width makes it possible to use more robust models that can deal with problems that rarely occur in the training data. It can also allow these models to be used for more proactive rather than reactive applications as the model can learn from knowledge sources other than raw historical data.

PiLogic has developed an AI modeling methodology that allows AI training to be simplified using known structures in the data and the system to be modeled. This allows for faster training and inference where such structure exists and reduces energy consumption for many of the demanding problems being solved by modern AI.