AI System Development (Deep Learning & Machine Learning)

■Contract Development / Research & Development

At Souya, we view AI as a tool to help solve real-world challenges—not a goal in itself.
Our aim is to deliver practical solutions and create new value through the smart application of technology.

We provide end-to-end support—from identifying business issues through consulting, to conducting research, analysis, and actual system development.
We work closely with clients to uncover the core challenges and deliver accurate, tailored proposals.
In some cases, the best solution may not involve AI. Instead, traditional statistical methods or image processing techniques may offer a more suitable and effective approach.

We also engage in joint research to explore cutting-edge fields and help turn our clients’ ideas into reality. In addition, we undertake a wide range of contract development projects.
If you’re struggling with issues like:
“I want to try a different technique with this deep learning library,”
“Can we do something with this data?”
“We built a model in-house, but it lacks sufficient accuracy,”
We welcome inquiries at any stage of your AI journey—whether you’re just getting started or facing specific technical challenges.

If you’re unsure about immediately implementing an AI system, we also offer AI Implementation Support Programs with proof-of-concept (PoC) testing to help you evaluate the feasibility and value of AI before committing to full-scale development.
Please don’t hesitate to contact us.

　

■About AI Technology

□ What is Machine Learning?

Machine learning is a technology that repeatedly learns from data to recognize patterns and rules derived from past experiences.
And then, when presented with new and unknown data, the machine autonomously applies the patterns and rules it has learned to generate appropriate answers.

The mechanism that recognizes patterns within data is generally called a model.
The process of feeding data into the model to help it recognize patterns and become more intelligent is called training.
In order to train a model, two types of data are required: input data and the corresponding correct (or labeled) data.
For example, if you want a model to identify which animal appears in a photograph, you would provide image data of animals as the input, and the type of animal shown in each photo as the correct answer.
By using these data pairs, the model can learn to make accurate predictions.

At the beginning of the machine learning process, it is necessary to extract numerical values called features from the input data.
Features are numerical representations that capture key characteristics of the input data, enabling the model to recognize patterns during training.

In machine learning, feature extraction—known as feature engineering—plays a key role in building efficient and highly accurate models. Identifying and designing effective features allows us to develop models with superior performance. However, feature engineering is often a complex and iterative process that requires specialized expertise.
That is why this critical and complex process of feature engineering is handled by experienced data scientists like us, with careful attention and persistence. At Souya, we leverage our many years of hands-on experience in machine learning to design high-quality features that drive better model performance.

□What is Deep Learning?

In traditional machine learning, humans are required to manually design features.
In contrast, deep learning automatically extracts features from the data and learns patterns on its own.
However, the learning process itself remains the same as in traditional machine learning—models are trained using pairs of input data and corresponding correct (labeled) data.
Deep learning is also well-suited for handling unstructured data, such as images and natural language.
This makes it possible to apply AI even to datasets that are large in volume but difficult to work with due to the challenges of feature engineering—such as image data.

Because deep learning automatically extracts features and learns patterns from data, it requires a large volume of data to function effectively.
With only a small amount of data, it cannot extract meaningful features or learn patterns accurately.

In other words, deep learning requires processing a large amount of data—which involves intensive computation.
This is where GPUs (Graphics Processing Units) come into play.

The advancement of GPUs has played a key role in the development of deep learning.
Thanks to GPU-based computation, we are now able to perform calculations that would have been practically impossible just a decade ago.

Inference refers to using a trained model to make predictions on new, unseen data.
Since inference is less computationally intensive than training, a GPU is often not required.

While deep learning requires large amounts of data, many industries operate with defect rates in the high 90% yield range, where defects are rare.
In such cases, it can be difficult to implement AI-based inspection systems, simply because there is not enough labeled data for defective products to train the model effectively.

At Souya, we have the expertise to build effective models even in cases where only limited data is available, or where no labeled data exists.
For example, in scenarios where there is no data for defective products but a large amount of data for non-defective products, we can still develop AI solutions.

Our approaches include methods for enhancing and emphasizing key features within the data to enable efficient learning from small datasets, as well as applying techniques such as unsupervised learning, where models are trained without labeled data.

Of course, we also have extensive experience with supervised learning when labeled data is available.

□Types of Learning

There are several different approaches to learning in machine learning.
Among them, the most widely used approach is called supervised learning.
Other types of learning include unsupervised learning, semi-supervised learning, and reinforcement learning.

Supervised Learning	In supervised learning, the model learns by being provided with both input data and the corresponding correct output in advance. Through this, it learns patterns such as “when the input has these features, the output should be this.” Supervised learning is used to make predictions about future events based on past data. Examples: Spam detection Image recognition Weather forecasting
Unsupervised Learning	In unsupervised learning, unlike supervised learning, the model is trained using only input data, without any correct (labeled) outputs. It learns the underlying structure of the data by analyzing the features extracted from it. Examples: Product categorization Recommendation systems Anomaly detection
Semi-Supervised Learning	Semi-supervised learning trains models in a similar way to supervised learning. What makes it different is that it can handle a mix of labeled data (with correct answers) and unlabeled data (without correct answers) during training. It is often said that semi-supervised learning can achieve higher accuracy than supervised learning alone, by effectively leveraging both types of data.
Reinforcement Learning	Reinforcement learning is fundamentally different from other learning methods. It observes and recognizes the current state, then selects actions that maximize the reward. In other words, it learns how to make the best possible decisions. Examples: Robotics, quality control, gaming (such as shogi, Go, and chess).

□Use Cases

AI can be applied to a wide range of real-world scenarios.
Choosing the right algorithm depends on the type and structure of the data involved.

・Image Dataタ

What It Can Do	Description	Expected Algorithms:
Recognition / Classification	The system identifies and classifies what is shown in an image. This type of AI is used in various applications such as: Visual inspection for product quality control Sorting vegetables and fruits based on visual features Supporting medical diagnosis by interpreting medical images.	FCN、EfficientNet、ResNet、NF-Nets、ViT
Object Detection	While image classification identifies what is present in an image, object detection also determines where it is located. This technology is widely applicable, including in building inspections using drones, security, and medical imaging.	YOLO、SSD、DERT
Anomaly Detection	This detects whether an image shows a deviation from a normal pattern, identifying abnormal conditions. By using unsupervised learning, anomaly detection can be performed even with only images of normal conditions. For example, in visual inspection tasks where it is difficult to gather images of defective products, models can be trained solely on images of non-defective (normal) products.	FCN、AutoEncoder、GAN、MetricLearning
Similarity Detection	We compare images to identify visually similar ones and perform clustering. This technology can be used for applications such as facial recognition or finding similar products (e.g., industrial parts or clothing). It is also utilized in our AI Similar Drawing Search system (link available).	AutoEncoder、GAN、MetricLearning
Image Generation / Enhancement	Given image data or related information, the system can generate or enhance images or associated content. It can be used for illustration generation, image captioning (translation), sharpening image clarity, or highlighting/editing parts of an image. This is a highly versatile technology with broad applicability.	GAN、U-Net

・Natural Language Data

What It Can Do	Description	Expected Algorithms:
Classification	Text classification determines the category of a given document or sentence. For example, it can classify news articles by topic, or determine sentiment polarity (positive/negative) in survey responses.	BERT、Word2Vec、fastText
Tagging / Key Phrase Extraction	This involves extracting important terms (key phrases) or named entities from text. It also enables tagging of documents, helping to identify their attributes or content type.	BERT
Similarity & Clustering	We compare texts and words to identify similarities and perform clustering. This technology is also used in our AI Similar Document Search system (link available).	BERT、Word2Vec、fastText

・Numerical Data (e.g., Sensor Data)

What It Can Do	Description	Expected Algorithms:
Classification	Numerical data from sensors can be classified and analyzed. For example, acceleration data can be used to determine the type of movement being performed, and vibration data can be analyzed to assess equipment degradation.	DNN、FCN、SVM、LightGBM
Anomaly Detection	This detects whether numerical sensor data deviates from normal patterns, indicating potential abnormalities or malfunctions.	Autoencoders, GANs (Generative Adversarial Networks), Metric Learning, and k-Nearest Neighbors (k-NN)
Prediction / Optimization	Numerical data can be used to predict future values or perform optimization. Applications include demand forecasting, parameter tuning for machinery, and automated quotation systems—a widely applicable set of technologies.	DNN (Deep Neural Networks), FCN (Fully Connected Networks), Linear, Nonlinear, and Logistic Regression, and LightGBM

■AI Case Studies

Please feel free to contact us.
We will introduce relevant case studies tailored to your specific challenges.

■What is Generative AI?

Generative AI refers to AI that can automatically create content such as text, images, or music.
Just like a human writes a document, AI can answer questions, generate images, or compose music.
RAG (Retrieval-Augmented Generation) can further enhance AI responses by incorporating timely, external knowledge—particularly useful when dealing with dynamic or specialized domains.

□What Generative AI + RAG Can Achieve

• Operational Efficiency
  ・Automation of administrative tasks
  ・Reduced time required for document creation
• Knowledge Discovery
  ・Enhanced search functionality
  ・Expanded scope of searchable information
  ・Shortened search times
• Creativity Enhancement
  ・Idea generation for product development and content creation

■Contact Us

If you have any questions or concerns, please don’t hesitate to contact us.