C2 · 50 words
Machine learning in Spanish
Training, evaluation and the limits of learning systems — the terms the real work hides behind.
Here el lote is a batch: el lote de datos, the handful of examples a model sees at once. The word does three other jobs that English hands to different nouns. At an auction it is a lot, el lote número siete. At Christmas it is a hamper, un lote de Navidad, the box of food a company sends round. And across much of Latin America it is a plot of land, un lote urbanizable. What holds them together is the idea of a portion set aside, which is exactly what a batch of data is.
The words
View
Phrase (44)
el modelo estadístico entrenadothe model that has worked the rule out from examplesPhraseC2
El modelo estadístico entrenado deduce la regla de los ejemplos.The trained statistical model works the rule out from the examples.
el aprendizaje supervisadothe learning that needs somebody to have labelled the answersPhraseC2
El aprendizaje supervisado necesita etiquetas.Supervised learning needs labels.
el aprendizaje no supervisadothe learning that looks for structure with no labels at allPhraseC2
El aprendizaje no supervisado busca estructura sin etiquetas.Unsupervised learning looks for structure without labels.
el aprendizaje por refuerzothe learning that goes by reward rather than by examplePhraseC2
El aprendizaje por refuerzo aprende de la recompensa.Reinforcement learning learns from the reward.
el conjunto de entrenamientothe data the model learns from, never used to judge itPhraseC2
El conjunto de entrenamiento nunca se usa para juzgar.The training set is never used to judge.
el conjunto de validaciónthe data kept back to choose the settings withPhraseC2
El conjunto de validación sirve para elegir los ajustes.The validation set serves to choose the settings.
el conjunto de pruebathe data opened once, at the very endPhraseC2
El conjunto de prueba se abre una sola vez.The test set is opened only once.
la validación cruzadathe splitting into parts so that each takes a turn as the testPhraseC2
La validación cruzada divide los datos en cinco partes.Cross-validation divides the data into five parts.
la fuga de datosthe seeping of the answer into the training data unnoticedPhraseC2
La fuga de datos infla el resultado sin que se note.Data leakage inflates the result without anybody noticing.
el compromiso entre sesgo y varianzathe trade nobody escapes: too rigid against too eagerPhraseC2
El compromiso entre sesgo y varianza no se elimina, se elige.The bias-variance trade-off is not removed, it is chosen.
la penalización de los pesosthe pressure that keeps the coefficients from growingPhraseC2
La penalización de los pesos frena el crecimiento de los coeficientes.Penalising the weights slows the growth of the coefficients.
la función de pérdidathe definition of what counts as being wrongPhraseC2
La función de pérdida define qué significa equivocarse.The loss function defines what being wrong means.
el descenso de gradientethe method that walks downhill in the steepest directionPhraseC2
El descenso de gradiente baja por la pendiente más pronunciada.Gradient descent goes down the steepest slope.
la tasa de aprendizajethe size of each step, which overshoots if it is too bigPhraseC2
La tasa de aprendizaje demasiado alta hace saltar el mínimo.A learning rate that is too high jumps over the minimum.
el mínimo localthe dip that is not the deepest one, and often does not matterPhraseC2
El mínimo local no siempre es un problema.The local minimum is not always a problem.
la época de entrenamientothe one pass through every examplePhraseC2
La época de entrenamiento recorre todos los ejemplos una vez.The training epoch goes through all the examples once.
el lote de datosthe handful of examples looked at together, here thirty-twoPhraseC2
El lote de datos tiene treinta y dos ejemplos.The batch of data has thirty-two examples.
la red neuronal profundathe network with many layers stacked on each otherPhraseC2
La red neuronal profunda apila muchas capas.The deep neural network stacks many layers.
la capa ocultathe layer that nothing outside the model ever seesPhraseC2
La capa oculta no se observa desde fuera.The hidden layer is not observed from outside.
la función de activaciónthe function that stops the whole thing being one straight linePhraseC2
La función de activación introduce la no linealidad.The activation function introduces the non-linearity.
el gradiente que se desvanecethe signal that dies away before it reaches the early layersPhraseC2
El gradiente que se desvanece detiene el aprendizaje en las capas bajas.The vanishing gradient stops the learning in the lower layers.
la red convolucionalthe network built to exploit what is next to whatPhraseC2
La red convolucional aprovecha la vecindad espacial.The convolutional network exploits spatial neighbourhood.
el mecanismo de atenciónthe mechanism that weighs every position of the inputPhraseC2
El mecanismo de atención pesa cada posición de la entrada.The attention mechanism weighs every position of the input.
la representación vectorialthe turning of a word or an item into a point in spacePhraseC2
La representación vectorial acerca lo que significa parecido.The vector representation brings similar meanings close together.
el espacio latentethe compressed space in which the essentials livePhraseC2
El espacio latente comprime lo esencial.The latent space compresses the essentials.
la reducción de dimensionalidadthe squeezing of many dimensions into a fewPhraseC2
La reducción de dimensionalidad facilita la visualización.Dimensionality reduction makes visualisation easier.
el árbol de decisiónthe model that reads as a sequence of questionsPhraseC2
El árbol de decisión se lee como una serie de preguntas.The decision tree reads like a series of questions.
el conjunto de modelosthe many models voting together, usually better than the best onePhraseC2
El conjunto de modelos suele superar al mejor solo.The ensemble of models usually beats the best one alone.
la matriz de confusiónthe table showing what gets mistaken for whatPhraseC2
La matriz de confusión muestra qué se confunde con qué.The confusion matrix shows what is mistaken for what.
la precisión del modelothe share of the flagged cases that were rightPhraseC2
La precisión del modelo engaña con clases desiguales.The precision of the model misleads with unequal classes.
el área bajo la curvathe single number that sums up the whole trade-offPhraseC2
El área bajo la curva resume el compromiso entero.The area under the curve sums up the whole trade-off.
la calibración del modelothe tuning that makes a stated probability worth believingPhraseC2
La calibración del modelo hace creíble la probabilidad.Calibration of the model makes the probability believable.
el desplazamiento de la distribuciónthe drifting of the world away from the training dataPhraseC2
El desplazamiento de la distribución degrada el modelo con el tiempo.Distribution shift degrades the model over time.
el sesgo del conjunto de datosthe tilt the data carried in, which the model inheritsPhraseC2
El sesgo del conjunto de datos se hereda sin remedio.The bias of the dataset is inherited without remedy.
la equidad algorítmicathe fairness whose definitions cannot all hold at oncePhraseC2
La equidad algorítmica admite definiciones incompatibles entre sí.Algorithmic fairness admits definitions incompatible with one another.
la explicabilidad del modelothe ability to say why a model answered as it didPhraseC2
La explicabilidad del modelo se exige en decisiones sensibles.Explainability is demanded in sensitive decisions.
la atribución de importanciathe pointing out of which feature weighed mostPhraseC2
La atribución de importancia señala qué rasgo pesó.Importance attribution points out which feature weighed.
el ejemplo adversariothe input changed by noise nobody can see, and answered wronglyPhraseC2
El ejemplo adversario cambia la respuesta con un ruido invisible.The adversarial example changes the answer with invisible noise.
el ajuste finothe adapting of a model that was already trainedPhraseC2
El ajuste fino adapta un modelo ya entrenado.Fine-tuning adapts an already trained model.
el aprendizaje por transferenciathe reusing of what was learnt elsewhere, which saves dataPhraseC2
El aprendizaje por transferencia ahorra datos y tiempo.Transfer learning saves data and time.
la anotación de datosthe labelling by hand, which is the work nobody seesPhraseC2
La anotación de datos es el trabajo que nadie ve.Data annotation is the work nobody sees.
el coste de inferenciathe price paid every time the model is asked somethingPhraseC2
El coste de inferencia se paga en cada consulta.The inference cost is paid on every query.
la trazabilidad del entrenamientothe record of which data and which version produced a modelPhraseC2
La trazabilidad del entrenamiento documenta datos y versiones.Training traceability documents data and versions.
la supervisión humanathe person who must be able to overturn the decisionPhraseC2
La supervisión humana debe poder cambiar la decisión.Human supervision must be able to change the decision.
Noun (m) (4)
el sobreajustethe learning of the noise along with the patternNoun (m)C2
El sobreajuste memoriza en lugar de generalizar.Overfitting memorises instead of generalising.
el subajustethe model too simple to fit even the data it was givenNoun (m)C2
El subajuste falla ya en los propios datos.Underfitting fails on its own data already.
el hiperparámetrothe setting chosen by hand rather than learnedNoun (m)C2
El hiperparámetro no se aprende, se fija.The hyperparameter is not learned, it is set.
el agrupamientothe finding of natural groups with nobody naming themNoun (m)C2
El agrupamiento encontró cuatro conjuntos naturales.Clustering found four natural groups.
Noun (f) (2)
la retropropagaciónthe sharing out of the error backwards through the layersNoun (f)C2
La retropropagación reparte el error hacia atrás.Backpropagation shares the error out backwards.
la exhaustividadthe share of the real cases that were foundNoun (f)C2
La exhaustividad importa cuando fallar sale caro.Recall matters when missing a case is expensive.
This list belongs to Technology and industry — 26 lists and 1296 words that can also be practised as a whole.
Practise this list
From these 50 words, as many per round as you like — up to all of them: first what is due today, then new ones. First you choose, then you fill the gap, then you write it yourself. Mistakes come back again shortly.