Cinnamon AI has shifted its focus from ``reducing simple tasks'' to ``maximizing creativity by strengthening human specialized skills'' in the use of AI in society, which is a major goal of the company. In developing technology that expands human specialized skills, we divide specialized skills into the stages of perception, reasoning, and response, and take an approach that uses AI to augment the skills that appear in these stages. Cinnamon AI's AI ecosystem aims to create comprehensive solutions that extract and accumulate specialized knowledge by building technologies at each stage of perception, inference, and response, and interconnecting these aspects. Here we introduce the concept of our expert AI. 

perception 

Current AI is only good at basic perceptual processes, such as object recognition and pattern classification. There are many AI solutions that identify human faces, track objects, and optically recognize characters, but most of them do not understand the basic structure of the information and signals they are sensing. When recognizing a human face, is it possible to recognize a heterogeneous object that has certain parts to ignore and parts to focus on? Does it understand the meaning of the text that is recognized and transcribed? The answer to all of these questions is "no." 

Experts in any field view their world uniquely through the lens of separation and selection, analyzing bits of information and discovering hidden characteristics that the average person often misses. Cinnamon AI provides a multimodal perception engine that decomposes data into structured and reusable formats. This is an important prerequisite for perception through an expert lens. 

Cinnamon AI provides uniquely developed AI that enhances perceptual abilities. 

• Separation and selection of multimodal information
• Semantic representation of knowledge (appropriately capture the semantic aspects of knowledge through mutual relationships)

Unlike traditional approaches that detect patterns as they are input, Cinnamon AI's system breaks down input patterns into meaningful pieces of information, revealing the underlying structure of the data. Cinnamon AI's Flax Scanner engine decomposes the contents of an invoice into keys and values, and extracts the information the user needs while preserving its attributes (keys and relationships between keys), reducing effort. This reduces the time and effort required to sort through invoices and compile information by inputting it into an information system. In addition, the Rossa engine enables highly customizable speech recognition that adapts to domain knowledge, and information processing of speech data such as tracking of speaker and content transition. 

Furthermore, by performing NLP analysis on the information acquired by these perception engines and its attribute information, we can capture not only individual information elements, but also their contexts and relationships between elements. For example, Cinnamon AI's Aurora engine captures the terms and conditions that lawyers have to consider through many pages of legal documents, while distilling the key points into short sections and paragraphs. It also has a similar analysis and processing engine that specializes in speech. 

In this way, we aim to analyze information from multiple angles and turn it into knowledge by integrating and linking multimodal business information generated in the office. 

Reasoning 

If you further observe the tasks of experts, you will notice that their thought process loops through one of the following six cognitive steps. 

1. Searching
2. Clustering
3. Recall
4. Induction
5. Deduction
6. Imagination 

Unfortunately, even the latest AI is still far from reaching these levels of awareness. However, Cinnamon AI is accumulating research to assist in expert-level reasoning. 

First, by advancing the semantic understanding of information in the "perception" process, we are pursuing a realistic processing system by verifying the usefulness of information for experts and reducing the search space by several hundred times. . We are also researching data representation and storage methods that allow data to be automatically classified from an expert's perspective and searched efficiently. Specifically, it is possible to maintain multiple background information for a single piece of business information through approaches such as incorporating context in the "perception" process, analyzing relationships between information elements, and acquiring multifaceted information from a multimodal recognition engine. information is stored in. This makes it possible to extract and classify information with a high degree of appropriateness depending on the user's situation. In fact, this system can improve the decision-making process of experts, as manual clustering and replication processes can take many hours even for experts. 

Cinnamon AI's expert AI system aims to not only enhance experts but also engage them in the thought process. It can be said that the inductive/deductive process is an important process by which experts develop their specialized knowledge and specialized skills. Experts discover laws through experience and extend the horizon of their specialized knowledge by extending the laws analogically. These skills will deepen and become more versatile with experience. It can be said to be the intellectual source that supports the creativity of experts. 

 AI's ability to discover patterns has evolved to surpass humans in certain tasks. From this, it can be said that AI is good at discovering rules. However, it is powerless when it comes to extending and generalizing that law, or applying it to completely new areas to gain new insights. In other words, there is a major deadlock in AI evolution in this induction/deduction process. 

We believe that the first step to overcoming this deadlock is to become an expert in the loop. In other words, if the expert and AI continue to interact, the AI will overcome the deadlock and bring more benefits to the expert. Specifically, AI can first extract and propose general rules by learning the accumulation of small patterns and rule discoveries from observation results. Some of them may provide insights to experts and help them derive new rules and principles of their own. Experts communicate their choices to the AI, which then reinforces its actions. On the other hand, AI systems attempt a meta-generalization process by collecting rules from experts across the board. These generalizable skills are continually strengthened by being validated by data and selected by experts, and AI distills and distills general guidelines for processing data. . In this way, inductive and deductive functions may be gradually developed through coexistence with experts. 

Ultimately, AI systems help experts generate new concepts from basic ideas through generative models. We are starting to see AI systems writing entire paragraphs after a human indicates the opening sentence, or generating the next frame from a few frames of animation. What we aim for with expert AI is a system that can provide quality that allows humans to gain insight, while ensuring that the output results are compatible with common sense and other prerequisite knowledge and are consistent. This feature maximizes the imagination of experts, enabling them to "foresee" the various subsequent paths that can occur from an initial idea, and from there, create a breakthrough. 

Response 

As mentioned above, in order to create opportunities for mutual learning and communication between humans and AI, which are essential for the evolution of AI, AI systems accurately inform humans of results and discoveries, and the AI accurately incorporates the feedback to improve the model. need to do it. Although this mechanism is considered the most important in the current AI industry, it has not yet been realized. 

Imagine having an expert in the loop to evaluate the AI output and guide your AI system to better performance. The expert understands the rationale behind the AI's responses, and the AI performs meta-learning from the expert's constructive feedback to learn perspectives and skills, allowing the expert to provide higher-quality responses. We will grow to be able to provide the following. This growth of AI will create empathy between AI and experts, forming a closed loop of interaction. 

 Cinnamon AI focuses on the following points to provide a solid direction for its future scenarios.

• Reliable response
• Hybrid intelligence through the symbiosis of experts and AI (expert in the loop)    

Machine output must be reliable. Therefore, Cinnamon AI is developing a confidence score estimation module to evaluate AI output. We are also developing technology to present the basis for AI decisions. In this way, it not only suggests to the expert when and where the machine's output should be used or discarded, but also transforms the machine's output and its rationale into a form that the expert can interpret. Our goal is to help experts understand how machines arrive at their results. This increases trust between the machine and the expert and builds a relationship of continuous mutual improvement. 

In addition, the model can be improved efficiently by interacting with the user from the AI side. For example, the practicality of the model can be dramatically improved by allowing the AI to confirm with the user where there are boundaries in judgment or areas in which they are not confident, or by clarifying the underlying context through interaction and appropriately narrowing down the scope of the problem. will improve. 

summary

Each AI system is designed to interact and collaborate with humans in stages. At Cinnamon AI, we build an ecosystem where experts and AI coexist, cooperate, and improve each other in a constructive interaction cycle in order to build a realistic system that maximizes the capabilities of experts. I am trying my best to do so. 

==================
At Cinnamon AI, we help you connect AI to your competitive strategy through consulting, workshops, and solutions. We would appreciate it if you could feel free to contact us.

Click here to contact us => Inquiry form