What is the most important step or practice for building effective domain-specific AI models that outperform general-purpose models?

Most answers emphasize the importance of high-quality, domain-specific data and processes for handling it, while a few highlight incorporating domain knowledge or fine-tuning as key practices.

9 answers

Answers

  • Naveen Rao

    [un]CEO · Unconventional Ai

    Curating Data

    Naveen Rao says the practice is curating data and choosing the right datasets and curriculum, because most capability gains come from data rather than advances on the model itself.

    20VC with Harry Stebbings · Watch on YouTube
  • Pushmeet Kohli

    Chief Scientist, Google Cloud · Google

    Incorporate Domain Knowledge

    Pushmeet Kohli says to incorporate domain knowledge into the model, arguing that scientific intuition can be baked in to give it an unfair advantage and make it much more data-efficient because it doesn't have to relearn everything.

    Latent Space · Latent Space · Read
  • Bharat Patel

    Managing Director, AI & Data, Defense Portfolio · Accenture Federal Services

    Fast Data Pipeline

    Bharat Patel says the more important focus is building a true pipeline that rapidly collects data, trains, tests, and pushes models back out, rather than building models from internal exercises, since those exercises may not represent the next fight.

    ChinaTalk · Watch on YouTube
  • Ali Ansari

    Founder & CEO · Micro1

    Filtering High-Quality Data

    Ali Ansari says the key step is filtering extracted data for quality, since large-scale data from any institution contains a lot of low-quality data that someone has to evaluate.

    Antler Global · Watch on YouTube
  • Michael Andrew

    Chief Data Officer · Salesforce

    Provide Domain-Specific Data

    Michael Andrew says you have to give agents data and information about your business, since they don't arrive knowing your business and won't learn unless you provide it.

    Super Data Science: ML & AI Podcast with Jon Krohn · Watch on YouTube
  • Rob Key

    CEO · Converseon

    Advanced Enrichment

    Rob Key says the key step is advanced enrichment—structuring data to specific brand and industry definitions with advanced NLP—because language differs by domain, and it reached 92% accuracy versus 65% for native enrichment.

    CEO Series · CEO Series · Listen at 15:57
  • Francois Chaubard

    Visiting Group Partner · Y Combinator

    Analyze the data

    Francois Chaubard says the correct next step is to look at the data, examining false positives and negatives and sorting them into buckets to find the biggest issue, because model tweaks like more layers won't fix problems such as fog or a person blocking the view.

    Y Combinator · Watch on YouTube
  • Kevin W. Tung

    Senior Partner · Interplay

    Logging Mistakes

    Kevin W. Tung says companies and people able to use AI to fix problems are those diligent about logging mistakes and problems along the way, because logging them and having AI provide solutions gets closer to a perfect outcome.

    Mark Peter Davis · Watch on YouTube
  • Ramin Hasani

    Co-founder and CEO · Liquid Ai

    Fine-Tuning

    Ramin Hasani says domain-specific models are built by fine-tuning: small models are bad at generalization but can be customized at low cost to solve a dedicated problem well.

    SiliconANGLE theCUBE · Watch on YouTube

More Data Science & Machine Learning questions

View all →