* Development Purpose While GPT-3-based natural language processing is widely popular, GPT-3's highly accurate model consists of 175 billion parameters. If each parameter is 4 bytes, it would require 700GB of RAM just to load.
The state-of-the-art large-scale language model, PaLM, has over three times more parameters than GPT-3, totaling 540 billion parameters. It is estimated that GPT-4 will have several trillion parameters (requiring over 4 terabytes of RAM).
State-of-the-art AI systems are designed to be used in a cloud-based environment.
Even for personal use, it is unlikely that there are many individuals or organizations that can afford over a million yen worth of GPUs and provide power supply comparable to a large factory.

Our technology is specialized in integrating the inference part of deep learning AI into existing systems, except for applications using the ChatGPT API. The inference part operates quickly even without a GPU.
Training is performed using GPUs in the range of 24 gigabytes.
All the technologies we provide have been successfully incorporated into actual applications that are being sold.
You can fulfill the following requests:
You want to run it as a standalone application on the client side, not as a web app.
You want to use deep learning AI even in a 32-bit environment without GPU.
You want to speed up deep learning AI using multithreading.
You want to run over 20 deep learning AI-compatible apps on a single server.
You want to use deep learning AI from languages like C++/C#/C, not just Python.

* Development Achievements

The learning part is based on using GPU with Tensorflow+Keras, etc. The development is done by calling the inference program (no GPU required) using only the pre-trained parameters from existing C++/C#/Java systems.
The following development achievements exist (including ongoing development).
  1. Based on a conventional OCR library (our product) that is overwhelmingly fast but does not have the accuracy of deep learning OCR
    2. Realizing high-accuracy OCR while maintaining high speed by re-recognizing only characters with low confidence using deep learning
  2. Based on the above hybrid OCR library
    3. Automatic classification of recognized string results using deep learning-supported natural language processing AI (corporate name, personal name, address, postal code, phone/fax number, email address, URL, date, etc.)
  3. Based on a super-fast conventional RV conversion library
    4. Identifying dashed lines, characters in contact with lines, interruptions in line drawings, arrow tips, and types of intersections using deep learning
    Expanding the functionality of RV conversion to deep learning support while maintaining super speed by narrowing down the candidates for search locations to endpoints and intersections based on the results of conventional RV conversion
  4. Outputting Prolog language programs using the API of a large-scale language model. Achieving a hybrid system with first-order predicate logic programming.
    5. Using a large-scale language model for text analysis and generation. Developing a chat program that enhances the accuracy of the inference part by using Prolog.
    Specifically, we are developing a chat program that presents in text the process of solving advanced essay questions for elementary school students without using equations and leading to the answer. The accuracy is higher than ChatGPT+GPT-4. We are also developing a customized model by fine-tuning the ChatGPT model for comparative verification.

* Denoising Autoencoder

The pioneer of the so-called deep learning boom. A general-purpose deep learning model that can be used for various purposes.
It has been used for classifying the text of business card recognition results (such as company name, full name, affiliation, title, postal code, address, phone number, email, etc.).
  1. Small-scale deep learning suitable for data up to 1GB
  2. Can be used for multiple purposes such as class classification, image recognition, language processing, etc.
  3. Training time is long and the inference accuracy is slightly inferior compared to dedicated algorithms.
  4. Both training and inference can be done in C++, no GPU is required, and it supports 32-bit and 64-bit execution.
  5. Inference can be performed at the same or higher speed as GPU utilization through multi-thread execution.
  6. Source code and library are available.

* CNN (Convolutional Neural Network)

A model for image recognition. It achieves faster learning and higher inference accuracy compared to Denoising Autoencoder in the context of image recognition.
It has been used for symbol recognition on maps.
It has been used for deep learning support of character recognition libraries.
  1. Can be used for learning from small to large-scale data up to 1TB.
  2. Suitable for image recognition, character recognition, shape recognition, etc.
  3. Training is provided in Python/TensorFlow/Keras, and for large-scale learning it supports TFRecord format.
  4. Inference is provided in C++, no GPU is required, and it supports 32-bit and 64-bit execution with multi-thread execution capability.
  5. Source code and library are available.

* word2vec

Model for natural language processing
By obtaining distributed representations of words, it is possible to represent sentences as arrays of multidimensional vectors. The multidimensional vectors can be directly used as input data for deep learning.
This model has been used for language processing in character recognition libraries.
  1. Optimal for small to medium-scale training data of about 30GB
  2. Dedicated to natural language processing
  3. Includes morphological analysis processing to divide sentences into words as preprocessing. No need to install other libraries such as MeCab. Also, it uses its own morphological analysis dictionary and no separate installation is required.
    The dictionary is a modified version of ipadic neologd. The morphological analysis program is built using an original algorithm (parallel exhaustive search + combination optimization using Viterbi trellis)
  4. Training/inference available in C++, GPU not required, can be executed in 32-bit & 64-bit, and supports multi-thread execution
  5. Library will be available for download in February 2023

* Fine-tuning of large-scale language models

API-ready application for fine-tuning large-scale language models
This system is designed for cloud applications and does not support local execution.
  1. Fine-tuning using ChatGPT API
    2. Application: Creating teaching data that generates explanatory texts of the solution process for math word problems at an elementary school level without using equations
  2. Combining ChatGPT with existing systems
    3. Building a chat system using fine-tuning with the following two tasks in combination with existing systems:
    • Converting input text into input data for existing systems
    • Generating text from the output data of existing systems
    Application: Building a chat system that combines text analysis and text generation using ChatGPT for a specialized math word problem-solving system at an elementary school level

Product Introduction

Topics