IBM to Launch Neural Network Learning Control Service

Developers believe that new service will bring greater transparency to the reasons for the decisions made by AI and can eliminate the "black box problem"
20 September 2018   472

IBM has developed a service for monitoring the processes that occur during the training of neural networks. The system identifies emerging misconceptions and gives greater transparency to the reasons for the decisions made by AI.

The new tool works with popular AI-frameworks, such as Watson, Tensorflow, SparkML, AWS SageMaker and AzureML. The service is implemented on the IBM Cloud platform and will help monitor the learning process by making the necessary adjustments. According to the representatives of the company, the software is easy to adapt to any architecture of the neural network. Moreover, the system is able to automatically offer correction of input data to eliminate delusions.

The service shows the parameters of the learning process using visual diagrams, which makes the user's work easier. Among the data displayed is a combination of factors accepted for consideration, confidence in the decision made and the foundation of this confidence. In addition, changes to the parameters are stored in the log, which will allow you to study the actions of AI more closely.

The monitoring service is not free, but at the same time IBM said it plans to release an open source version of the product. The company declares this as a contribution to international cooperation in eliminating AI's misconceptions.

The reasons for the decisions made by artificial intelligence are in most cases hidden from the end user. At the same time, studies have shown that neural networks are able to assimilate inherent misconceptions and stereotypes, for example, gender or racial. This gave rise to some mistrust of AI and the fear of losing control over the technology. According to an IBM poll, 82% of entrepreneurs consider the introduction of neural networks. However, while 60% are afraid of possible problems, and 63% are not sure that they will be able to confidently manage new tools.

The so-called "black box problem", consisting in the non-transparency of AI decisions, is taken seriously by the world community. Work to increase transparency and trust is being carried out quite actively. In September 2018, MIT scientists published their development, illustrating the decision-making process by the neural network.

Facebook to Release PyTorch 1.0

This release added support for large cloud platforms, a C ++ interface, a set of JIT compilers
10 December 2018   125

Facebook has released a stable version of the library for machine learning PyTorch 1.0. This iteration added support for large cloud platforms, a C ++ interface, a set of JIT compilers, and various improvements.

The stable version received a set of JIT compilers that eliminate the dependence of the code on the Python interpreter. The model code is transformed into Torch Script - a superstructure over Python. Keeping the opportunity to work with the model in the Python environment, the user can download it to other projects not related to this language. So, the PyTorch developers state that the code processed in this way can be used in the C ++ API.

The torch.distributed package and the torch.nn.parallel.DistributedDataParallel module are completely redesigned. torch.distributed now has better performance and works asynchronously with the Gloo, NCCL and MPI libraries.

The developers added a C ++ wrapper to PyTorch 1.0. It contains analogs of Python interface components, such astorch.nn,torch.optim, torch.data. According to the creators, the new interface should provide high performance for C ++ applications. True, the C ++ API is still experimental, but it can be used in projects now.

To improve the efficiency of working with PyTorch 1.0, a Torch Hub repository has been created, which stores pre-trained models of neural networks. You can publish your own development using the hubconf.py file, after which the model will be available for download by any user via the torch.hub.load API.

Support for C extensions and the module torch.utils.trainer were removed from the library.

Facebook released the preliminary version of PyTorch 1.0 at the beginning of October 2018, and in two months the developers brought the framework to a stable state.