A new open source project seeks to simplify the use of a machine learning framework called “transfer learning” in which the ability to accomplish one task can be applied to subsequent tasks.
Transfer learning specifically involves the reuse of training data and other properties used in a prior model. The Python-based project dubbed Enso and launched this week on GitHub by enterprise AI startup Indico Data Solutions is intended to streamline the benchmarking of transfer learning methods for a variety of natural language processing tasks.
The Boston-based company said Tuesday (June 26) Enso seeks to fill a number of gaps in AI development as technologists look for ways to expand machine learning capabilities beyond the performance of individual tasks.
Transfer learning has been widely used for computer vision and image classification. The use of pre-trained models demonstrated that machines could be trained with dozens rather than hundreds of thousands of images. Indico researchers note that transfer learning has so far made less of an impact in the field of natural language processing.
Described as a standard interface for benchmarking transfer learning methods for natural language processing, Enso addresses a “core set of interrelated problems.” The first is the current inability to reproduce research results. The startup notes that gauging the effectiveness of a new methodology has proven difficult given the use of custom data sets and the lack of standard coding practices.
Another challenge addressed by the Enso project is improving baseline benchmarks used to evaluate new approaches based on a range of data sets. Only then can researchers determine if a new method represents an improvement over current natural language processing frameworks, the company said.
The open source project also seeks to overcome the current inability to reuse training models. “Many of the models used for benchmarking are tied to specific data sets, making it too difficult to take a model trained for one domain and train it on another,” Indico noted.
Enso also would promote the development of “more general data sets” and improve baselines for measuring progress in the development of natural language processing frameworks based on transfer learning. The ultimate goal is expanding machine learning applications to a broader set of tasks.
“Measuring how well methods perform as the amount of training data increases is critical,” said Madison May, Indico’s co-founder and machine learning architect.
“In real life examples, we often need to select for methods that perform well with only a few hundred labeled training examples,” May added. “By providing a standard interface for benchmarking, we believe Enso can facilitate the development of more generalized models that have greater value to a broader base of users.”
Indico, which specializes in AI development based on unstructured content, said the Enso project is compatible with the 3.4+ version of the Python programming language.
Transfer learning is seen as an up-and-coming AI training capability, according to James Kobielus, an analyst with SiliconAngle Wikibon. “Typically, a developer relies on transfer learning to tap into statistical knowledge that was gained on prior projects through supervised, semi-supervised, unsupervised, or reinforcement learning,” Kobielus noted in a Datanami assessment of model training approaches.