Narrative Goldmine

Self Training

32 min read

Properties

Type
  • Class
Resource
  • urn:ngm:class:self-training
Domain
  • artificial-intelligence
Status
  • stable
Maturity
  • emerging
Is a
Part of
Enables
Generated
  • by: process:vault-migrate/1.0 · at: 2026-09-22T12:41:39.024503178Z
Legacy term id
  • AI-0281
Links
Page resource
  • urn:visionflow:page:148248ed43246b07c02cf28b1d102974bb625c8cb0db5035739c845215c8d488
Related to
Uses

A semi-supervised learning technique where a model is iteratively improved by training on its own high-confidence predictions on unlabelled data. Self-training enables learning from large amounts of unlabelled data by using the model’s own predictions as pseudo-labels.

Semantic Classification

Content

  • A semi-supervised learning technique where a model is iteratively improved by training on its own high-confidence predictions on unlabelled data. Self-training enables learning from large amounts of unlabelled data by using the model’s own predictions as pseudo-labels.

Characteristics of Deep Agents

  • Deep agents distinguish themselves through:
    • Extended runtime (minutes to hours or days)
    • Comprehensive planning and re-planning
    • High-value, substantial outputs
    • Self-correction and iteration capabilities
    • Significant computational resource usage

Self-Improvement of GPT Models

  • GPT-4 Can Improve Itself: Featuring Reflexion, HuggingGPT, Bard Upgrade, and more (YouTube Video).

Distilled Rust Advice

  • Refactor main.rs into separate layers:

    // Refactor main function to use dependency injection and minimize hard dependencies.
fn refactor_main_to_minimize_dependencies() {
 // Move database initialization, middleware configuration, and HTTP server setup to separate modules.
 // Inject database and middleware dependencies into your app as services.
}

  • Define traits for repository pattern:

    // Create domain-specific traits for repositories.
pub trait AuthorRepository {
 fn create_author(&self, req: &CreateAuthorRequest) -> Result<Author, CreateAuthorError>;
 // Define other repository methods here, e.g., find, findAll, etc.
}

  • Encapsulate dependencies in adapters (avoid leaking 3rd party libraries):

    // Wrap external database connection pool.
struct Sqlite {
 pool: sqlx::SqlitePool,
}

impl AuthorRepository for Sqlite {
 async fn create_author(&self, req: &CreateAuthorRequest) -> Result<Author, CreateAuthorError> {
     // Manage transactions, handle errors, and abstract them from the domain.
 }
}

  • Testability improvements with mocks:

    // Create mock implementations of the repository for testing.
#[derive(Clone)]
struct MockAuthorRepository {
 result: Arc<Mutex<Result<Author, CreateAuthorError>>>,
}

impl AuthorRepository for MockAuthorRepository {
 async fn create_author(&self, _: &CreateAuthorRequest) -> Result<Author, CreateAuthorError> {
     // Return pre-configured mock results for testing.
 }
}

  • Decouple domain from transport (HTTP) layer:

    // Convert HTTP request body to domain model.
impl CreateAuthorHttpRequestBody {
 fn into_domain(self) -> Result<CreateAuthorRequest, AuthorNameEmptyError> {
     let name = AuthorName::new(&self.name)?;
     Ok(CreateAuthorRequest::new(name))
 }
}

  • Create a service layer for complex domain logic:

    pub trait AuthorService {
 async fn create_author(&self, req: &CreateAuthorRequest) -> Result<Author, CreateAuthorError>;
}

struct Service<R: AuthorRepository> {
 repo: R,
}

impl<R: AuthorRepository> AuthorService for Service<R> {
 async fn create_author(&self, req: &CreateAuthorRequest) -> Result<Author, CreateAuthorError> {
     // Handle business logic, metrics, notifications, etc.
 }
}

  • Set up bootstrapping logic in main.rs:

    // Main function should focus only on bootstrapping services and starting the server.
#[tokio::main]
async fn main() -> anyhow::Result<()> {
 let config = Config::from_env()?;
 let sqlite = Sqlite::new(&config.database_url).await?;
 let author_service = Service::new(sqlite);
 let http_server = HttpServer::new(author_service, config.server_port).await?;
 http_server.run().await
}

  • Define async repository methods and handle error scenarios:

    pub trait AuthorRepository: Send + Sync {
 async fn find(&self, id: &Uuid) -> Result<Option<Author>, CreateAuthorError>;
 async fn find_all(&self) -> Result<Vec<Author>, CreateAuthorError>;
}

Human flourishing and expression

  • Celebrating Human Diversity: Now to Later
    • AI understanding and adapting to human experiences and identities.
  • AI supported creativity: Soon to Next
    • AI assisting and leading in art, music, and literature creation.
  • Equity: Now to Next
  • Self Guided Learning for Children: Next to Later
    • Playful, memorable AI education experiences for children, globally.
  • The age of the productive tinker: Later
    • AI revolutionizing industries with specialized gadgets and applications.

Cloud AI versus On-Premise Models

  • Cloud services offer simplicity and scale, whereas self-hosted models (e.g. running Stable Diffusion locally or deploying your own LLM) give you full control over data and costs.

Characteristics of Deep Agents

  • Deep agents distinguish themselves through:
    • Extended runtime (minutes to hours or days)
    • Comprehensive planning and re-planning
    • High-value, substantial outputs
    • Self-correction and iteration capabilities
    • Significant computational resource usage

Self-Improvement of GPT Models

  • GPT-4 Can Improve Itself: Featuring Reflexion, HuggingGPT, Bard Upgrade, and more (YouTube Video).

Distilled Rust Advice

  • Refactor main.rs into separate layers:

    // Refactor main function to use dependency injection and minimize hard dependencies.
fn refactor_main_to_minimize_dependencies() {
 // Move database initialization, middleware configuration, and HTTP server setup to separate modules.
 // Inject database and middleware dependencies into your app as services.
}

  • Define traits for repository pattern:

    // Create domain-specific traits for repositories.
pub trait AuthorRepository {
 fn create_author(&self, req: &CreateAuthorRequest) -> Result<Author, CreateAuthorError>;
 // Define other repository methods here, e.g., find, findAll, etc.
}

  • Encapsulate dependencies in adapters (avoid leaking 3rd party libraries):

    // Wrap external database connection pool.
struct Sqlite {
 pool: sqlx::SqlitePool,
}

impl AuthorRepository for Sqlite {
 async fn create_author(&self, req: &CreateAuthorRequest) -> Result<Author, CreateAuthorError> {
     // Manage transactions, handle errors, and abstract them from the domain.
 }
}

  • Testability improvements with mocks:

    // Create mock implementations of the repository for testing.
#[derive(Clone)]
struct MockAuthorRepository {
 result: Arc<Mutex<Result<Author, CreateAuthorError>>>,
}

impl AuthorRepository for MockAuthorRepository {
 async fn create_author(&self, _: &CreateAuthorRequest) -> Result<Author, CreateAuthorError> {
     // Return pre-configured mock results for testing.
 }
}

  • Decouple domain from transport (HTTP) layer:

    // Convert HTTP request body to domain model.
impl CreateAuthorHttpRequestBody {
 fn into_domain(self) -> Result<CreateAuthorRequest, AuthorNameEmptyError> {
     let name = AuthorName::new(&self.name)?;
     Ok(CreateAuthorRequest::new(name))
 }
}

  • Create a service layer for complex domain logic:

    pub trait AuthorService {
 async fn create_author(&self, req: &CreateAuthorRequest) -> Result<Author, CreateAuthorError>;
}

struct Service<R: AuthorRepository> {
 repo: R,
}

impl<R: AuthorRepository> AuthorService for Service<R> {
 async fn create_author(&self, req: &CreateAuthorRequest) -> Result<Author, CreateAuthorError> {
     // Handle business logic, metrics, notifications, etc.
 }
}

  • Set up bootstrapping logic in main.rs:

    // Main function should focus only on bootstrapping services and starting the server.
#[tokio::main]
async fn main() -> anyhow::Result<()> {
 let config = Config::from_env()?;
 let sqlite = Sqlite::new(&config.database_url).await?;
 let author_service = Service::new(sqlite);
 let http_server = HttpServer::new(author_service, config.server_port).await?;
 http_server.run().await
}

  • Define async repository methods and handle error scenarios:

    pub trait AuthorRepository: Send + Sync {
 async fn find(&self, id: &Uuid) -> Result<Option<Author>, CreateAuthorError>;
 async fn find_all(&self) -> Result<Vec<Author>, CreateAuthorError>;
}

Human flourishing and expression

  • Celebrating Human Diversity: Now to Later
    • AI understanding and adapting to human experiences and identities.
  • AI supported creativity: Soon to Next
    • AI assisting and leading in art, music, and literature creation.
  • Equity: Now to Next
  • Self Guided Learning for Children: Next to Later
    • Playful, memorable AI education experiences for children, globally.
  • The age of the productive tinker: Later
    • AI revolutionizing industries with specialized gadgets and applications.

Cloud AI versus On-Premise Models

  • Cloud services offer simplicity and scale, whereas self-hosted models (e.g. running Stable Diffusion locally or deploying your own LLM) give you full control over data and costs.

Self-Improvement of GPT Models

  • GPT-4 Can Improve Itself: Featuring Reflexion, HuggingGPT, Bard Upgrade, and more (YouTube Video).

Human flourishing and expression

  • Self Guided Learning for Children: Next to Later
    • Playful, memorable AI education experiences for children, globally.
  • The age of the productive tinker: Later
    • AI revolutionizing industries with specialized gadgets and applications.

Cloud AI versus On-Premise Models

  • Cloud services offer simplicity and scale, whereas self-hosted models (e.g. running Stable Diffusion locally or deploying your own LLM) give you full control over data and costs.

Planning and Execution

  • Extended runtime (minutes to hours or days)
  • Comprehensive planning and re-planning
  • High-value, substantial outputs
  • Self-correction and iteration capabilities
  • Significant computational resource usage

The future of AI in call centers is likely to be shaped by the following trends:

  • Increased automation: AI will continue to automate more tasks, freeing up human agents to focus on more complex and strategic activities3. This includes the increasing use of process automation, which is expected to be the most prominent application of AI in contact centers 24.
  • Hyper-personalization: AI will enable even more personalized customer interactions, tailoring support and offers to individual needs and preferences25.
  • Enhanced self-service options: AI-powered chatbots and virtual assistants will become more sophisticated, providing customers with more comprehensive self-service options25.
  • Advanced analytics: AI will provide deeper insights into customer behavior and call center performance, enabling data-driven decision-making3. This includes the use of voice authentication and speech analytics to enhance security and improve customer understanding 24.
  • Generative AI integration: Generative AI will be used to create more human-like interactions and automate tasks like summarizing conversations and generating reports26.
  • Addressing data scarcity: AI-powered contact center solutions will help overcome the challenge of limited data analysis by enabling the examination of a larger proportion of customer interactions26.
  • Robotic Process Automation (RPA): RPA will continue to play a significant role in automating routine tasks and streamlining workflows in call centers27.
  • AI Predictive Auto Dialing: AI will be used to optimize dialing strategies and improve contact rates in outbound call centers27.

August 2024

The future of AI in call centers is likely to be shaped by the following trends:

  • Increased automation: AI will continue to automate more tasks, freeing up human agents to focus on more complex and strategic activities3. This includes the increasing use of process automation, which is expected to be the most prominent application of AI in contact centers 24.
  • Hyper-personalization: AI will enable even more personalized customer interactions, tailoring support and offers to individual needs and preferences25.
  • Enhanced self-service options: AI-powered chatbots and virtual assistants will become more sophisticated, providing customers with more comprehensive self-service options25.
  • Advanced analytics: AI will provide deeper insights into customer behavior and call center performance, enabling data-driven decision-making3. This includes the use of voice authentication and speech analytics to enhance security and improve customer understanding 24.
  • Generative AI integration: Generative AI will be used to create more human-like interactions and automate tasks like summarizing conversations and generating reports26.
  • Addressing data scarcity: AI-powered contact center solutions will help overcome the challenge of limited data analysis by enabling the examination of a larger proportion of customer interactions26.
  • Robotic Process Automation (RPA): RPA will continue to play a significant role in automating routine tasks and streamlining workflows in call centers27.
  • AI Predictive Auto Dialing: AI will be used to optimize dialing strategies and improve contact rates in outbound call centers27.

August 2024

Jul 2024

June 2024

The future of AI in call centers is likely to be shaped by the following trends:

  • Increased automation: AI will continue to automate more tasks, freeing up human agents to focus on more complex and strategic activities3. This includes the increasing use of process automation, which is expected to be the most prominent application of AI in contact centers 24.
  • Hyper-personalization: AI will enable even more personalized customer interactions, tailoring support and offers to individual needs and preferences25.
  • Enhanced self-service options: AI-powered chatbots and virtual assistants will become more sophisticated, providing customers with more comprehensive self-service options25.
  • Advanced analytics: AI will provide deeper insights into customer behavior and call center performance, enabling data-driven decision-making3. This includes the use of voice authentication and speech analytics to enhance security and improve customer understanding 24.
  • Generative AI integration: Generative AI will be used to create more human-like interactions and automate tasks like summarizing conversations and generating reports26.
  • Addressing data scarcity: AI-powered contact center solutions will help overcome the challenge of limited data analysis by enabling the examination of a larger proportion of customer interactions26.
  • Robotic Process Automation (RPA): RPA will continue to play a significant role in automating routine tasks and streamlining workflows in call centers27.
  • AI Predictive Auto Dialing: AI will be used to optimize dialing strategies and improve contact rates in outbound call centers27.

August 2024

Jul 2024

June 2024

The future of AI in call centers is likely to be shaped by the following trends:

  • Increased automation: AI will continue to automate more tasks, freeing up human agents to focus on more complex and strategic activities3. This includes the increasing use of process automation, which is expected to be the most prominent application of AI in contact centers 24.
  • Hyper-personalization: AI will enable even more personalized customer interactions, tailoring support and offers to individual needs and preferences25.
  • Enhanced self-service options: AI-powered chatbots and virtual assistants will become more sophisticated, providing customers with more comprehensive self-service options25.
  • Advanced analytics: AI will provide deeper insights into customer behavior and call center performance, enabling data-driven decision-making3. This includes the use of voice authentication and speech analytics to enhance security and improve customer understanding 24.
  • Generative AI integration: Generative AI will be used to create more human-like interactions and automate tasks like summarizing conversations and generating reports26.
  • Addressing data scarcity: AI-powered contact center solutions will help overcome the challenge of limited data analysis by enabling the examination of a larger proportion of customer interactions26.
  • Robotic Process Automation (RPA): RPA will continue to play a significant role in automating routine tasks and streamlining workflows in call centers27.
  • AI Predictive Auto Dialing: AI will be used to optimize dialing strategies and improve contact rates in outbound call centers27.

Key Characteristics

  • Uses model’s own predictions

    • Iterative improvement

    • Semi-supervised learning

    • Leverages unlabelled data

    • Requires confidence thresholding

    • Can amplify biases if not careful

      Academic Context

      Self-training demonstrates that models can bootstrap their own performance by leveraging confident predictions on unlabelled data, reducing reliance on expensive labelling.

  • Semi-Supervised Learning: Broader paradigm

    • Pseudo-Labelling: Related technique

    • Co-Training: Multi-view variant

      UK English Notes

    • “Labelled/unlabelled” (not “labeled/unlabeled”)

      Last Updated: 2025-10-27 Verification Status: Verified against semi-supervised learning literature

      Academic Context

  • Self-training is a semi-supervised learning technique where a model iteratively improves by training on its own high-confidence predictions on unlabelled data.

  • This approach leverages the model’s own predictions as pseudo-labels, enabling learning from large volumes of unlabelled data.

  • It builds on foundational concepts in machine learning, combining elements of supervised and unsupervised learning to reduce reliance on costly labelled datasets.

  • The academic foundation of self-training lies in iterative refinement, where the model alternates between predicting labels for unlabelled data and retraining on these predictions to improve accuracy.

  • This method dates back decades but has seen renewed interest with advances in deep learning architectures and increased availability of unlabelled data.

  • It is closely related to, but distinct from, self-supervised learning, which generates supervisory signals from the data itself without explicit pseudo-labels[1][2][6].

    Current Landscape (2025)

  • Self-training is widely adopted in industry for tasks where labelled data is scarce but unlabelled data is abundant, such as natural language processing, computer vision, and speech recognition.

  • Notable platforms and organisations employing self-training include Google, Microsoft, and various AI startups focusing on scalable semi-supervised solutions.

  • The technique is often integrated with other learning paradigms, including self-supervised and continual learning, to enhance model robustness and adaptability[4][6].

  • In the UK, particularly in North England, self-training methods are increasingly used in AI research hubs and industry collaborations.

  • Cities like Manchester and Leeds host AI innovation centres applying self-training to healthcare diagnostics, financial modelling, and smart city initiatives.

  • Newcastle and Sheffield contribute through academic research and partnerships with tech companies exploring semi-supervised learning for industrial automation and environmental monitoring.

  • Technical capabilities:

  • Self-training can significantly reduce the need for manual labelling but is sensitive to error propagation if the model’s initial predictions are poor.

  • Recent advances focus on confidence calibration and selective pseudo-labelling to mitigate confirmation bias.

  • Standards and frameworks for semi-supervised learning, including self-training, are evolving, with increasing emphasis on reproducibility, fairness, and transparency in AI systems.

    Research & Literature

  • Key academic papers:

  • Amini, M.-R., et al. (2025). Self-Training: A Survey. arXiv preprint arXiv:2202.12040v6.
    DOI: 10.48550/arXiv.2202.12040

    • This comprehensive survey details algorithms, theoretical foundations, and practical applications of self-training, updated as recently as February 2025[6].

  • Other foundational works explore the interplay between self-training and self-supervised learning, highlighting their complementary roles in modern AI[1][2].

  • Ongoing research directions include:

  • Enhancing robustness to noisy pseudo-labels.

  • Combining self-training with nested and continual learning paradigms to avoid catastrophic forgetting[4].

  • Applying self-training in multi-modal and low-resource settings.

    UK Context

  • The UK has a vibrant AI research ecosystem contributing to semi-supervised learning advancements.

  • Universities in Manchester, Leeds, Newcastle, and Sheffield actively publish on self-training techniques, often in collaboration with local industry.

  • Manchester’s AI research groups focus on healthcare applications, leveraging self-training to improve diagnostic models with limited labelled data.

  • Leeds and Newcastle contribute to financial and environmental AI solutions, respectively, using semi-supervised approaches to harness unlabelled datasets.

  • Regional innovation hubs foster start-ups and spin-offs applying self-training in real-world scenarios, supported by UK government AI initiatives and funding programmes.

  • The North England AI community is known for pragmatic, application-driven research, often balancing academic rigour with industrial relevance—because who said machine learning can’t be a bit down-to-earth?

    Future Directions

  • Emerging trends:

  • Integration of self-training with self-supervised and continual learning to create more adaptive, lifelong learning systems.

  • Development of more sophisticated confidence estimation and pseudo-label selection mechanisms to reduce error amplification.

  • Expansion into new domains such as autonomous systems, personalised education, and climate modelling.

  • Anticipated challenges:

  • Managing bias and fairness when models generate their own training labels.

  • Ensuring transparency and interpretability in iterative self-training processes.

  • Scaling self-training methods efficiently for very large datasets without prohibitive computational costs.

  • Research priorities include:

  • Theoretical analysis of convergence and error bounds in self-training.

  • Cross-disciplinary approaches combining machine learning with cognitive science insights.

  • Strengthening UK and North England’s leadership in responsible AI through robust semi-supervised learning frameworks.

    References

    1. NIST AI Glossary (2025). Self-supervised learning. Computer Security Resource Center.
    2. Jing, L., & Tian, Y. (2022). Self-supervised learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence.
    3. Amini, M.-R., et al. (2025). Self-Training: A Survey. arXiv preprint arXiv:2202.12040v6. DOI: 10.48550/arXiv.2202.12040
    4. Behrouz, A., & Mirrokni, V. (2025). Introducing Nested Learning: A new ML paradigm for continual learning. Google Research Blog.
    5. Lumenalta (2025). 5 types of machine learning.
    6. IBM (2025). What Is Self-Supervised Learning? IBM Think.

    Metadata

  • Last Updated: 2025-11-11

  • Review Status: Comprehensive editorial review

  • Verification: Academic sources verified

  • Regional Context: UK/North England where applicable

Provenance

Graph View

Backlinks