Google is one of the pioneers in ML/AI based solutions. Google is constantly trying to find ways to reduce the efforts of the application developers and enterprises to quickly plug and play from many of their solutions. These solutions work seamlessly from mobile to web to IoT devices. Also, Google has products for everyone – from beginner to expert, from developer to business analyst, from startups to enterprises, from owning the infrastructure to pay-as-you-go.
Google solutions and offerings can be broadly classified into 3 main categories. I have also grouped this based on the difficulty level of developers to adopt and start using –
1. Infrastructure support for AI/ML workloads (Developer difficulty level – high) –
Often there is confusion between what type of virtual machines should be used for what type of workloads. Virtual machines with vCPU’s are sufficient in most cases for general-purpose computing. However, for high-performance machine learning, GPU’s and TPU’s can be used.
A) Cloud GPU’s – GPU’s are google proprietary and they run with 2000+ ALU’s. This is most suitable for machine learning, scientific computing, and 3D visualization. GPU’s can be enabled in Compute Engines depending on the zones and region.
B) Cloud TPU’s – TPU’s are designed by Google like a matrix processor that is faster than GPU’s and can be used for deep learning and neural networks. One of the main advantages of TPU is speed and the cheaper cost. TPU’s can be enabled from Compute Engine or GKE or AI Platform depending on the zones and region.
Read more at – What makes TPUs fine-tuned for deep learning?
2. Cloud provisioned AI/ML solutions (Developer difficulty level – medium) –
These are AI tools that run on Google cloud infrastructure and do not need much infrastructure provisioning and management. These tools can be accessed from anywhere right through google cloud console in a browser.
A) AI platform – AI platform provides end to end machine learning lifecycle. It lets developers prepare, build, validate, and deploy the models at scale with the convenience of MLOps by using the pipelines.
B) DialogFlow – It helps in creating a conversational AI including voice bots, chatbots, and IVR bots. Dialog flow can also be called through an API using google private key. Dialog flow provides one-click integration with the majority of the popular social media platforms and chat apps.
C) Contact center AI – Enables the virtual agent to provide intuitive customer experience based on the actions taken from Dialogflow. Also helps the human agent with the agent assist tool. Insights provide tools to understand the metrics for better outcomes of the call.
D) Deep learning containers – Helps in doing the prototypes faster by providing a consistent environment based on the pre-configured containers from the Google container registry. It supports popular frameworks like TensorFlow, PyTorch, and Scikit learn.
E) TensorFlow enterprise – It comes with all the benefits of standard TensorFlow features. It also includes 3-year long term support, priority support for bug fixes, enterprise-ready TensorFlow through containers, and virtual machines. Above all, this comes at no additional cost.
F) Document AI – Used in conjunction with other ML API’s to get the complete insights and action items. First thing is to create a processor and choose between general and specialized processors, then predict and use the structured data based on the business needs.
G) Recommendations AI – Ingest the product catalog and customer event data through integration channels to be used by recommendations AI. Choose recommendation type, objective, and business rules to show recommendations to the customer.
H) AutoML tables – Create a dataset and then import the data from BigQuery, CSV from cloud storage, or local storage. Once the data is imported then it can be trained, evaluated, and deployed to get predictions in the real world.
3. API based AI/ML solutions (Developer difficulty level – low) –
ML API’s from google are just like any other API’s that the developers would call which involves a simple request and response pattern over a REST-style. The response usually will be in a JSON format and can be secured using an API key.
A) Text-to-speech – As the name suggests the API converts the text into natural-sounding speech. It supports more than 220+ voices across 40 languages.
B) Speech-to-text – The API can listen to speech and convert it to text. The API supports more than 125 languages and variants. It can also be deployed at on-prem data centers using Speech-To-Text On-Prem.
C) Vision AI – Pass the image over the REST call and get the response back with object identification, labels, safe search annotations, etc. AutoML Vision helps in training custom machine learning models by simply uploading the images.
D) Video AI – Google’s pre-trained models can detect objects, places, and actions by annotating the video using pre-defined labels. Use AutoML video intelligence to annotate videos using custom labels.
E) Translation – AutoML trains custom model by simply uploading the translated language pairs. Translation API instantly translates into more than 100 languages that can be directly embedded into websites and apps.
F) Natural language – This can be used to derive insights from unstructured data. This can be done in 2 different ways – 1) AutoML natural language, and 2) Natural language API.
G) Healthcare natural language – This early access API helps to find, assess, and link the medical knowledge in the text data.
H) Media translation – Another early access ML API that helps in real-time audio translation of the audio data. This provides an experience similar to how auto closed captioning in YouTube works.
Share your experience with google machine learning tools.