In our research reports, we cover “the recently possible,” and what makes “the recently possible” possible. In addition to a detailed how-to guide of new machine learning capabilities, each of our reports contains a section on open source projects, commercial offerings, and vendors that help implement this new machine learning capability to realize the opportunities opened up by technological innovation. We like to keep an eye on the the technologies we research, of course. Our report on deep learning (for image classification) was published a few years ago, and we have seen noteworthy new developments.
Training neural networks requires a set of “hyperparameters” that guide the training process, and sometimes aspects of the network itself (such as the number of layers of the network or the number of nodes per layer). Hyperparameter tuning is important because hyperparameters have substantial effects on the quality of the trained model.
Tuning hyperparameters is important to a neural net’s performance (Image Credit: Torsten Maue, Wikimedia Creative Commons (Compressorhead))
Hyperparameter tuning has generally been at least as much of an art as a science. Machine learning specialists select preliminary hyperparameters, train the model with those hyperparameters, and evaluate the performance and characteristics of the trained model. Then they adjust the hyperparameters (applying some educated guesses), retrain the model with those hyperparameters, and evaluate again. This process is repeated until the specialist is satisfied with the model.
Several new options promise to automate at least a part of the work of hyperparameter tuning, freeing up the personnel building the model, and possibly finding better hyperparameter settings than a human might. They generally work by automating the hyperparameter selection, model training and evaluation, and hyperparameter adjustment, searching for better model performance. These automated hyperparameter tuning packages include offerings from major cloud services providers (e.g., SageMaker from Amazon and Azure ML Studio from Microsoft) as well as smaller independent vendors (e.g., Comet.ml, Weights & Biases, and SigOpt). Afficionados of open source projects may enjoy Hyperopt, a python library for “serial and parallel optimization over awkward search spaces.” As anyone who has ever embarked on a quest for optimal hyperparameters can tell you, “awkward” is a good description of the process of searching for the optimal hyperparameters of neural networks.
For a thorough overview, Jesus Rodriguez rounds up automated hyperparameter tuners in a recent Hackernoon article. Hyperparameter tuning is one of the more easily automated aspects of machine learning (in contrast to, for example, feature engineering). In most practical settings, careful feature engineering requires an understanding of the organization and business, data collection protocols, as well as data science tools and algorithms; it is hard to automate with out-of-the-box solutions. We welcome appropriate automation; automated hyperparameter tuning promises to cut down development time of neural networks and lead to better results.
More from the Blog
Jun 26 2018
by — Today’s machines can identify objects in photographs, predict loan repayments or defaults, write short summaries of long articles, or recommend movies you may like. Up until now, machines have achieved mastery through laser-like focus; most machine learning algorithms today train models to master one task, and one task only. We are excited to introduce multi-task learning in our upcoming webina...
Jun 29 2018
Sequence labeling tasks attempt to assign a categorical label to each member in the sequence. In natural language processing, where a sequence generally refers to a sentence, examples of sequence labeling include named entity recognition (NER), part-of-speech tagging (POS) and error detection. NER, as the name implies, tries to recognize names in a sentence and classify them into pre-defined la...
Apr 3 2019
by — Many interesting learning problems exist in places where labeled data is limited. As such, much thought has been spent on how best to learn from limited labeled data. One obvious answer is simply to collect more data. That is valid, but for some applications, data is difficult or expensive to collect. If we will collect more data, we ought at least be smart about the data we collect. This motiv...