Skip to main content

What is machine learning (ML)

What is machine learning (ML)

Machine learning (ML) is a subset of artificial intelligence where systems learn patterns from data instead of following hand-written rules. Think of it as teaching by example: instead of writing a step-by-step recipe, you show the system thousands of finished dishes and let it figure out how to cook. Rather than a programmer coding every decision, the system is shown thousands or millions of examples and discovers the rules on its own.

Most practical AI today — from spam filters and voice assistants to medical image analysis and large language models — is built on machine learning. When a crypto project claims to use AI, the underlying technology is almost always some form of ML.

In short: ML means learning from data. Instead of telling the system what to do, you show it examples and it figures out the patterns. In crypto AI, "machine learning" in a project card means the project claims data-driven learning, not hard-coded logic.

How machine learning works

A machine learning system follows a general cycle. Understanding each step is essential for evaluating whether a project's ML claims are real — every step produces verifiable artifacts that can serve as evidence.

The ML lifecycle

  1. Define the problem — determine what the model needs to predict or classify. For example: "given a blockchain transaction, determine whether it is suspicious." A clear problem statement is the first checkpoint when evaluating ML claims.

  2. Collect data — gather a dataset relevant to the problem. This might include transaction histories, wallet activity logs, text documents, or user behavior records. Data can come from public sources, proprietary databases, or on-chain records.

  3. Prepare and clean the data — raw data is rarely usable as-is. This step involves removing errors, handling missing values, normalizing formats, and splitting the data into training, validation, and test sets. Data preparation typically consumes the majority of an ML project's time.

  4. Choose a model architecture — select the type of AI model suited to the task: a neural network, decision tree, support vector machine, or ensemble method. The choice depends on the data type, problem complexity, and available compute. For text, transformer architectures (the basis for GPT, BERT, and modern LLMs like DeepSeek) dominate. For tabular data, gradient-boosted trees (XGBoost, LightGBM) often outperform deep learning.

  5. Train the model — run a training algorithm that adjusts the model's internal parameters (weights) so that its predictions match the known outcomes in the training data. Training iterates over the data multiple times (called epochs), gradually reducing the difference between predictions and reality (measured by a loss function).

  6. Evaluate — measure the trained model against held-out test data or benchmarks to check if it learned genuine patterns or merely memorized the training set (a problem called overfitting). Common metrics include accuracy, precision, recall, and F1 score for classification; mean squared error and R-squared for regression; perplexity for language models.

  7. Deploy — put the model into production where it performs inference on new, unseen inputs. Deployment involves setting up serving infrastructure, managing latency, and handling edge cases.

  8. Monitor and retrain — real-world data changes over time (data drift). A model trained on last year's blockchain transaction patterns may perform poorly on current transactions. Responsible ML systems include monitoring pipelines that detect performance degradation and trigger retraining on fresh data.

A practical example

Suppose a project claims its ML model detects suspicious wallet activity. A verifiable implementation would show:

  • Data: what on-chain data is collected and how it is labeled (e.g., addresses flagged by known exploit databases)
  • Model: what architecture is used (e.g., a gradient-boosted classifier on transaction features)
  • Metrics: measured detection rate, false positive rate, and methodology (e.g., "evaluated on 10,000 held-out transactions from a defined period")
  • Deployment: how the model serves real-time predictions (API endpoint, latency characteristics)
  • Updates: how often the model is retrained and what triggers retraining

If a project describes none of these, the "ML" claim lacks verifiable substance.

Key ML approaches relevant to crypto AI

Machine learning is not a single technique — it is a family of approaches, each suited to different types of problems.

Approach How it learns Data requirement Crypto example
Supervised learning From labeled examples Labeled dataset Fraud detection, scam classification
Unsupervised learning Finds structure without labels Unlabeled dataset Wallet clustering, anomaly detection
Reinforcement learning Through trial and reward Environment + reward signal Automated agents, protocol optimization
Deep learning Multi-layer neural networks Large dataset + compute LLMs, image generation, NLP
Transfer learning Adapts pre-trained model to new task Small task-specific dataset Fine-tuning LLMs for crypto domain

Supervised learning

The model learns from labeled data — each example comes with a known correct answer. The model learns to map inputs to outputs and is evaluated on how well it predicts the correct answer for new, unseen examples.

  • How it works: given pairs of (input, correct output), the algorithm adjusts model parameters to minimize prediction errors.
  • Example: a classifier trained on labeled transactions (fraudulent / legitimate) to flag scams. Each training example is a transaction with features (amount, sender history, contract interaction) and a label.
  • Common tools: scikit-learn, XGBoost, LightGBM, and deep learning frameworks like PyTorch and TensorFlow.
  • Strength: high accuracy when good labels are available.
  • Limitation: requires labeled data, which can be expensive and time-consuming to create.

Unsupervised learning

The model finds patterns without labels. No one tells it what to look for; it discovers structure, clusters, or anomalies in the data on its own.

  • How it works: the algorithm groups similar data points together or identifies outliers based on statistical properties.
  • Example: clustering wallet addresses by behavioral patterns — transaction frequency, amounts, interaction with known protocols — to identify groups of related wallets without prior labels.
  • Common tools: k-means clustering, DBSCAN, principal component analysis (PCA), autoencoders.
  • Strength: works when labeling is impractical or too expensive.
  • Limitation: results require human interpretation; the algorithm finds structure but does not explain what the clusters mean.

Reinforcement learning (RL)

The model learns through trial and reward. It takes actions in an environment and receives feedback (positive or negative). Over time, it learns strategies that maximize cumulative reward.

  • How it works: an agent observes the current state, takes an action, receives a reward or penalty, and updates its strategy. This loop repeats thousands or millions of times.
  • Example: an automated agent that learns to optimize gas fees by timing transactions based on network congestion patterns.
  • Milestones: RL powers game-playing systems like AlphaGo. In crypto, RL appears in protocol parameter tuning and agent-based strategies.
  • Strength: suited for sequential decision-making problems where the optimal action depends on long-term consequences.
  • Limitation: requires a well-defined reward signal and many iterations; can be unstable during training.

Deep learning

A subset of ML that uses neural networks with many layers (hence "deep"). Deep learning powers most modern AI breakthroughs: image recognition, natural language processing, generative models, and large language models.

  • How it works: data flows through successive layers of artificial neurons, each extracting progressively more abstract features. A convolutional layer might detect edges; deeper layers combine edges into shapes, then objects.
  • Example: a transformer model (the architecture behind GPT, BERT, and DeepSeek) fine-tuned on crypto documentation to extract structured information from unstructured text.
  • Key architectures: transformers (text/sequences), convolutional neural networks / CNNs (images), recurrent neural networks / RNNs (time-series).
  • Strength: excels at complex, high-dimensional data (text, images, audio, sequences).
  • Limitation: requires large datasets, significant compute power, and careful tuning. Performance is harder to interpret than simpler models.

Transfer learning

A technique where a model pre-trained on a large general dataset is adapted to a specific task with a smaller dataset. Instead of training from scratch, the model starts with knowledge already learned.

  • How it works: take a model like BERT (pre-trained on general text) and fine-tune it on a smaller dataset of crypto whitepapers to classify project claims.
  • Why it matters: transfer learning reduces the data and compute required for new tasks. Many crypto AI projects use fine-tuned versions of existing models rather than training from zero.

From classical ML to modern deep learning

Understanding the evolution of ML helps evaluate project claims:

  • Classical ML (1990s-2010s): decision trees, support vector machines, logistic regression. These work well on structured (tabular) data and remain common in crypto analytics for transaction classification and risk scoring.

  • Deep learning era (2012-present): neural networks with many layers began outperforming classical methods on complex tasks. The 2017 transformer architecture ("Attention Is All You Need") reshaped the field. BERT (2018) and GPT (2018-present) brought transformers to language understanding and generation.

  • Foundation models (2020s-present): very large models pre-trained on broad data, fine-tuned or prompted for specific tasks. GPT, Claude, Gemini, and Mistral are well-known examples. These models make powerful NLP accessible but also make it easy to build thin wrappers that claim "AI" with minimal original work.

When a project says "machine learning," it is worth asking: is this classical ML on structured data, fine-tuning of existing foundation models, or training a new model from scratch?

Why ML matters in DecAIHub

When a project card lists "machine learning" as part of its AI component, it means the project claims to use a system that learns from data rather than relying on static rules. This distinction matters because:

  • ML implies a training pipeline — there must be data, a training process, and a resulting model. These are all verifiable artifacts.
  • ML implies measurable performance — a learned model can be benchmarked and compared. A hand-coded rule set cannot be evaluated the same way.
  • ML implies ongoing maintenance — models degrade over time as data distributions shift. A project using ML needs to retrain and re-evaluate periodically.
  • ML implies compute requirements — training and inference require computational resources. Decentralized ML projects must explain how compute is provisioned and who bears the cost.

DecAIHub checks for Tier-1 evidence of these artifacts when evaluating whether "machine learning" is real or just a marketing label. The AI reality flags (AI_0 through AI_4) map directly to evidence checkpoints: does the project describe its ML component (AI_1)? Are there working demos (AI_2)? Published metrics (AI_3)? Technical architecture documentation (AI_4)?

A project that mentions "machine learning" without providing any of these artifacts receives lower AI reality flags, which limits its DecAI Fit score through the ai_score ceiling mechanism.

What ML alone does not prove

  • Model quality — claiming ML does not mean the model is accurate, robust, or well-tuned. Quality requires benchmark evidence.
  • Data quality — the model is only as good as its training data. Poor, biased, or insufficient data produces poor results regardless of the algorithm.
  • Genuine learning — some projects label rule-based systems or simple heuristics as "ML." See Automation vs real AI component.
  • Proprietary advantage — using an off-the-shelf ML library with default settings on public data does not constitute meaningful AI depth.
  • Decentralization — ML models can run on fully centralized infrastructure. The fact that a project uses ML says nothing about whether its AI is decentralized or its blockchain integration is meaningful.

Do not confuse with

  • Artificial intelligence — the broader field. ML is a subset of AI, not a synonym for it.
  • AI model — the trained artifact. ML is the process of creating that artifact through data-driven learning.
  • Deep learning — a subset of ML that uses multi-layer neural networks. All deep learning is ML, but not all ML is deep learning.
  • Automation — executing predefined rules. ML learns its own rules from data. See Automation vs real AI component.
  • Statistics — ML builds on statistical foundations but goes beyond traditional statistics by focusing on predictive performance on new data rather than on explaining historical data.

Frequently asked questions

What is the difference between machine learning and traditional programming?

In traditional programming, a developer writes explicit rules: "if transaction amount exceeds a threshold and sender has no history, flag it." In machine learning, the developer provides labeled examples, and the algorithm discovers the rules itself. The result is a model that can detect patterns too subtle for hand-written rules — but it requires quality training data and ongoing evaluation.

Does machine learning require big data?

Not always. Deep learning models (especially LLMs) require large datasets, but many classical ML techniques work effectively with modest amounts of data. Transfer learning further reduces data requirements by starting from a pre-trained model. The key factor is data quality and relevance, not just quantity. A small, well-curated, properly labeled dataset can outperform a large, noisy one.

What is the difference between supervised and unsupervised learning?

Supervised learning uses labeled data — each example includes the correct answer, and the model learns to predict it. Unsupervised learning works with unlabeled data, discovering patterns, clusters, or anomalies without being told what to look for. A third category, semi-supervised learning, combines a small amount of labeled data with a larger amount of unlabeled data. The choice depends on whether labels are available and affordable to produce.

How can you tell if a crypto project uses real ML?

Look for verifiable artifacts: a described dataset, a documented model architecture, published benchmarks with methodology, a public demo, and evidence of retraining. DecAIHub evaluates these through the AI reality flags (AI_1 through AI_4) and Tier-1 evidence checks. A project that says "ML-powered" but provides none of these is likely using the term as marketing.

Is deep learning the same as machine learning?

Deep learning is a subset of machine learning, not a synonym. ML includes many techniques beyond deep learning: decision trees, support vector machines, Bayesian methods, and ensemble methods. Deep learning uses neural networks with multiple layers to learn hierarchical representations. It excels at complex data (text, images, audio) but simpler ML methods often outperform it on structured, tabular data.

See also

External references

No comments to display

Back to top