How to name a machine learning algorithm
Real ML literature names follow detectable patterns. ResNet (Residual Network), BERT (Bidirectional Encoder Representations from Transformers), GPT (Generative Pre-trained Transformer), AlphaGo. Note: many are acronyms that also work as pronounceable words. For your algorithm, try the same: a descriptive acronym sounding like a real word.
The 'method-foundation-task' technique. RoBERTa = Robustly Optimized BERT Approach. ELECTRA = Efficiently Learning an Encoder that Classifies Token Replacements Accurately. Each letter of the acronym justifies something from the paper. Exercise: first write what your model does in one sentence, then extract initials forming a word.
Another school uses mythological or astronomical names: AlphaGo, AlphaZero, MuZero (DeepMind), Gemini, Claude, GPT-4, Llama. This tradition seeks to evoke intelligence or universality without describing technique. If your algorithm is commercial product more than academic paper, this route sells better.
For academic papers vs. commercial products
NeurIPS, ICML, ACL papers: the name should be memorable and describe technical contribution. ResNet stuck because the central method is residual connection. Transformer stuck because of attention. If your paper introduces a new mechanism, the name should encode it. Attention Is All You Need is title; the algorithm is Transformer, evoking 'representation transformation'.
Commercial products: the name must be pronounceable, registerable and memorable. Watson, Cortana, Alexa, Siri, Bard, ChatGPT, Gemini, Copilot. Note the trend toward short human names. For your product invent a name pronounceable in any language without stumble. Watson works in English and Spanish; a name with 'ñ' or special characters doesn't export well.
Open-source models: the culture is more playful. LLaMA, Vicuna, Alpaca, Falcon, Mistral. Animals, mythology, wordplay. If your model is derived from Llama, you can call it Vicuna, Guanaco, Alpaca to signal lineage. This visual naming convention creates community belonging.
Common mistakes naming algorithms
Mistake 1: too long or unpronounceable name. 'BiLSTM-CRF-Attention-Multi-Task-Network' is technical paper name nobody cites. The rule is 1-2 pronounceable words or a 3-5 letter acronym. If your paper is named that, it will appear awkwardly abbreviated in literature and the brand dilutes.
Mistake 2: collision with existing names. Before fixing your algorithm in a paper, google it. 'OpenAI launches GPT-X' collides with OpenAI family. 'Bard' clashed with Google's brand. The 30-second search avoids months of legal problems and lost citations.
Mistake 3: generic name. 'DeepNet' or 'NeuralNet' is like calling your car 'Car'. You need specificity. Going Deeper with Convolutions gave GoogLeNet/Inception, better than 'DeeperConvNet'. Think what visual metaphor or distinguishable concept you can capture in 5-8 letters.
Applications in sci-fi narrative and tech thriller
In near sci-fi, the algorithm name is world element. Person of Interest has The Machine and Samaritan. Westworld has Rehoboam. Devs has Devs (deus). Each name has resonance: biblical, mythical, religious. Your fictional algorithm can be Tiresias, Cassandra, Oracle, Lighthouse. The choice communicates what it does before explaining.
In contemporary thrillers, the algorithm is subtle villain. The Social Dilemma, Coded Bias name 'the algorithm' generically, producing diffuse horror. For specific narrative, give it a name: 'the FACE-IV system', 'the Recidivism-3 model', 'the SocialScore-7 ranker'. Technical nomenclature concretizes the danger.
For corporate satire, the algorithm name can be absurd-aspirational: 'TalentSync v2.0', 'SmileScore Pro', 'EmpathyMetric Enterprise'. Severance and Mr. Robot use this technique. The mix of empty marketing tech words with specific dehumanizing function is its own subgenre.