By Shridhan Rokade
Artificial Intelligence (AI) has been evolving at a higher pace. It is poised to become one of the most impactful technologies humanity has ever created, boosting us and giving the ability to solve the problems once thought to be unsolvable.
Intelligent AI systems are being built having innovative algorithms & models that predict and prescribe the outcomes accurately, which are powered by essential a) underlying HPC compute platform, b) computes ability to interface, and self-tune at device level, c) associate & bind necessary built-in multimatrix operations, and d) effectively tweaking the runtime environments on the fly, for the desired outcome. GPU’s and its closely coupled platform plays a key role in building the smartest AI systems and successful AI Startups!
Typical journey of an AI model to evolve from Base model to Personalized model for an AI Startup and GPU Usage –
Stage -I
During stage-I journey for an AI models – typically a generalized model for multiple industries would be considered e.g. for Image Recognition; Sentiment Analysis; Speech to text across industries, wherein the advantage is to have a tested AI model as a jumpstart model to begin with. These are typically applicable on publically available datasets providing reasonable accuracy to the tune of 70% to 80% (depending on the pain area/model). Its gives head-start for the startup for @40% of the execution to build the final AI model. The GPU Usage during this stage remains occasional and recommended GPU consumption would be on GPU on Cloud environments.
Stage- II
Stage-II in AI model development, typically has been to target a specific industry segment & have smaller relevant datasets for the model to be built upon. Typical example includes Speech to text model trained for Indian accent and its ability to decode Indian Automobile Sector terms, wherein the advantage is targeting specific industry pain point. Maximum proportion of the Industry Specific Data in the base data to be trained upon e.g. Indian Automobile Sector in this case with accuracy between 80%-85% (depending on the pain area/model). Its gives head-start for the startup for @60% of the execution to build the final AI model. The GPU Usage during this stage is typically continuous for training and recommended GPU consumption would be on GPU on Cloud environments.
Stage-III
During this stage the focus is on a specific target customer on customer specific dataset. Example Speech to text model trained on customer specific automobile components, policies & offers. The accuracy % goes higher to the tune of 95% (depending on the pain area/model) as the model is trained on more than 65% of the customer specific dataset. It leads to @80% of the execution to build the final AI model. The GPU Usage during this stage is continuous for training and recommended GPU consumption would be on GPU on Cloud environments.
Stage-IV
Stage-IV focus remains on having personalized model for the end users. wherein loopback methodology is implemented after inferencing in order to retrain the model, so as to capture personalized preferences. Typical example in the retail Industry, where its expected to have personalized preferences being generated while making purchasing decision and would be interested in capturing it for awesome customer experience. Accuracy is expected to be much higher here close to 100%, hitting the business objectives. This takes it to >95% of the execution to build the final AI model. The GPU Usage during this stage is continuous for training & inferencing and recommended GPU consumption would be on in-house/on-prem or GPU on Cloud environments.
While GPUs are equally important, its also essential to have the GPU platform associated with the underlying key components fulfilling all the dependencies in order to expedite AI Startup Journey towards success. Specifically if the AI Startups are empowered with the GPU’s and the platform with below features –
Key GPU & associated platform components for a successful AI startup
- Instant access to the industry datasets from various repositories such as UCI, Kaggle & Research Institutes
- Access to set of well-known ML algorithms to jumpstart the AI model development
- Easy access to auto detect the relevant model applicable to the dataset of your interest
- Platform to link the trained model to the end-customer through globally accepted protocols such as REST & RPC
- Customized & AI based DevOps to manage the AI model lifecycle
- And finally collaboration options to share the cost & simplified pricing through the AI model lifecycle
Selecting appropriate mix of hardware platform, software mix, frameworks, access to the datasets, models, pipe-line, along with simplified pricing for base models, industry models, customer model, and end-user models is the key for a successful AI startup.
(The author is the Founder and CEO, GPUONCLOUD)