Developing neural networks is not simple– it includes huge amounts of unstructured data and many hypotheses and experimentation which are all expensive and time consuming. A new Israeli startup which came out of stealth on Thursday claims that it can test and repair your networks before any chaos begins.
Instead of trial and error, use a real test
The Israeli startup Tensorleap developed a platform capable of debugging and testing artificial neural networks (ANN) and deep neural networks (DNN) all with the help of algorithms. "The platform adds mathematical operations at each layer in the computational graph of the neural network in order to extract indicators from all the feature maps and evaluate their contribution to the decisions it makes," explained David Ben David, CEO of Tensorleap, in a conversation with Geektime.
"Next, Tensorleap's algorithms build the most informative space to explain the way the model interpreted the information, find clusters of data with similar characteristics, and more." According to him, the accessibility of these analyses to data scientists is a "game changer": "Other tools mainly encourage the perpetuation of the existing paradigm of trial and error and are not deep analysis tools for the models themselves. Therefore, their value is limited, and they do not make a fundamental change in the development process or the quality of the developed model... A company like Tensorleap had to be established at some point or another to allow the advancement of the field to the next stages, which is why we founded it.”
New seed funding in their pocket

According to Ben David, a client from the field of genetic analysis estimated that a project he was developing would take about 9-10 months, but by Tensorleap, it was completed in less than 3 months, as it found bugs in the company's information processing processes. The product can identify areas of weakness in the model and understand why they fail. In addition, the system can identify balance problems in the databases used to train the model and correct them. Based on the transition to production, the system performs dozens and hundreds of in-depth tests that make sure that the model functions in any scenario it may encounter. "This is a fundamental difference from what exists today – a number of basic tests, which must pass a certain percentage in order to bring a model to production, which of course is not enough and leads to many failures later on," said Ben David.
Tensorleap, founded in 2020, announced the completion of a $5.2 million seed round from Angular, Sozo and Industry Ventures, and currently employs approximately 15 people in its offices in Ramat Gan.