For the past some time Microsoft has made increasing forays into the field of AI and their application in software assisted applications. In the beginning their implementations were very basic in nature and narrow in scope. But with time, their solutions powered by machine learning have become very diverse and advanced. This has happened without added complexity which makes most AI applications inaccessible to first time users.
The specific fields covered by Microsoft have broadened to the areas of language recognition, translation, and image and video recognition. Microsoft has striven to ensure their methods are easily accessible to developers of all skill levels.
This year Microsoft has invested their resources in two broad categories
- Deployment
- Customisation
Machine learning is made available to its users by following two steps, model building and model use. In model building a set of test data is supplied to the machine learning algorithm, which is used to train neural networks to make the model. In model use new data is fed to the neural network which processes it according to to the training it received during the model building phase.
Microsoft is now providing its services with prebuilt models. These services can be enhanced by training them further on specific data sets. This kind of enhancement allows developers to train the models on their own custom data sets, built on top of existing models.
The training process is highly compute intensive, arrays of data processing units or GPUs are required to train the models to fairly good accuracy levels. Using prebuilt models however reduces the computational requirements by considerable amount. This allows customers to modify existing models in a short time in such a way that they fit within their requirements.
To facilitate this functionality Microsoft’s new deployment tools come into focus. Training models can be deployed to Azure IOT or Windows Machine Learning runtime allowing clients to push their services to various connected devices. This allows the models to be run locally allowing considerable savings in bandwidth and computing costs.
Vision will be the first service to support this kind of deployment. The models are deployed in the ONNX format developed and supported by Microsoft, Facebook, AWS, Nvidia, Intel and AMD.
Microsoft is however not resting on its laurels. It is also deploying new services, such as a custom search, a visual search service powered by Bing and forecasting service which make future predictions for weather, seismic activity, sports and so on.
It has also continuing investments in chat bots. While their initial efforts faltered, they have made considerable improvements in their latest versions. For example, Microsoft powers the Flo bot which deployed on Facebook Messenger can handle the entire process of selling insurance policies. A new kind of service called conversation learner allows the chat bot to learn conversational patterns from existing data sets. These are the kind of features which will make future chat bots more natural and more human like.
Services galore
AI has very usual applications in image recognition, sound analysis and so forth. Many people still associate machine learning (ML) with the internet. However it has many obvious applications which were not obvious before. For example retinopathy diagnosis is complex endeavor, it requires a ophthalmologist. However with the help of AI thousands of retina photographs can be analyzed. This can be done using a image system which works in three ways:
- Left eye, right eye detection
- Photo quality analysis
- Detection of retinopathy
Using this system a physician can tell with high accuracy if a patient needs to be further referred to ophthalmologist. AI can potentially help sifting through the large amount of images which are required for this kind of service.
We as a society are at an inflexion point when it comes AI and Machine Learning. On one hand AI has immense potential to help us as a society to solve our most challenging problems. On the other hand as many economists point out AI has the potential to wipe out our entire job market. In the future it might be possible that articles like these could be written by a computer. Whether we get a Ash or a Bishop is up to debate.
I’d like both.
