Artificial intelligence (AI) and machine learning are undergoing strong development, mainly due to the technological steps taken with deep learning and neural networks. However, AI’s automatic decision making process remains unfathomable for users. IBM and Harvard University have now developed the Seq2Seq-Vis tool, which gives developers insight into this decision-making process and possibilities for adjustment and correction.
The tool now being developed by Big Blue and the renowned American university enables developers, as Venturebeat writes, to visualise the decision-making process of artificial intelligence when it translates a sequence of words from one language to another.
The tool thus provides insight into the so-called black box problem that developers have because they cannot see which decisions, for example, make deep neural networks. This is one of the main problems that the AI industry is currently facing. Obtaining this insight is important because artificial intelligence is increasingly being used for critical purposes.
Seq2Seq-Vis focuses on so-called sequence-to-sequence models. This is the AI architecture currently used in most automated translation systems. This architecture is capable of converting an input series of any length into an output of any length as well.
The source string of, in this case, words is passed through different neural networks to divide it into the target language and refine it to give a grammatically and semantically correct answer.
Neural networks improve this process but also make it more complex. Developers find it difficult to define where in the decision making process errors have occurred and cannot simply correct them. The tool has now been developed for this purpose.
Specifically, Seq2Seq-Vis provides a visual representation of the different steps in a sequence-to-sequence translation process. Users can then examine the model’s decision making process and see where any errors occur.
By linking to training data for the model, end users can also see whether the errors detected are due to poor training examples, the neural networks that classify sentences in the source and destination languages, a misconfiguration of the source-destination connection or in the beam search, the artificial intelligence model that refines the output of the translation model and ensures grammatical and semantic correctness.
In addition, end users can use the tool to actively correct errors. This allows them to select and correct individual words in the output process. They can also reset the way in which the connection between input and output assigns the output positions.
No holy grail
The researchers do indicate that the tool is not yet the holy grail for really understanding the decision-making processes of artificial intelligence. Much human knowledge of the models used is still required and the tool should have access to the training methods and other technical details of the artificial intelligence it wants to help improve. Seq2Seq-Vis is particularly suitable for architects and trainers of artificial intelligence.This news article was automatically translated from Dutch to give Techzine.eu a head start. All news articles after September 1, 2019 are written in native English and NOT translated. All our background stories are written in native English as well. For more information read our launch article.