At IBM, we took a tangential approach to empower subject matter experts with tools to train models for AI solutions. Imagine an radiologist who understands anomalies from MRI and xRays with the ability to train models to integrate AI into their practice. AI in radiology is transforming health care with MRI machines can study the images with analysis for radiologists. Saving time and ensuring faster personalized care to their patients.

IBM PowerAI Vision abstracts the deep learning techniques required to train models for computer vision. Its a clicker tool where subject matter experts bring their datasets to label, train and deploy models of great accuracies. In this blog, we introduce features to accelerate transformation of your businesses with AI. Complete overview of the product can be found on IBM Marketplace.

Data augmentation

IBM PowerAI Vision is based on Deep Learning techniques where the machine automatically learns features from the training dataset. The caveat to this approach is providing a large amount of data with variations. Unfortunately, there are domains where large amounts of datasets are simply not available. Image a drone flying around a transmission tower, capturing video and images of assets like conductors, insulators, bird cages etc. One would not find a large number of conductors which might be broken, rusted or torn. And sending the drone back does not guarantee discovery of more defective components. In IBM PowerAI Vision, we introduce a Data Augmentation capability, where user cans select algorithms to rotates, crop, flip, sharpen or blur images. These features help create variations from the initial datasets and augment the data required for training. This AI on AI is resulting in savings on costs & time associated with flying drones.

Metrics to measure accuracy and detect issues with datasets

Once models are trained, a detailed dashboard helps business users identify a range of statistics that infer to the accuracy of the models and gaps in datasets. Identifying gaps in distribution of datasets that represents desired categories are crucial to ensure high accuracy of resulting models. Statistics like confusion matrix, total recall and precision makes it easy for business users to zero in on images that are either miss categorized or need augmentation. The dashboard displays configurable-hyperparameters for a data scientist to tweak on the next iteration.

Multiple categories on inference

Once the models are trained and deployed, it is important to ensure the models are relavent when inferred against real-time data. Our ability to show sorted listed of multiple categories with a score helps business users to monitor the accuracy of model and identify datasets required to retrain. This simple provides a feedback mechanism required to keep models abreast with things that change around us.

Data scientists get to Bring your own Models

IBM PowerAI Vision abstracts all the deep learning techniques and lowers the technical requirements to train and deploy models. In addition, the product also supports custom models created by Data Scientists (limited to TensorFlow) to be trained and deployed on IBM PowerAI Vision. With the activities related to training and deployment now owned by IBM PowerAI Vision, data scientists can now use the saved time and energy to innovate their models.

Train on datacenter, infer on remote end points

IBM PowerAI Vision engages hardware that is at the core of Super Computer (Summit, IBM Power Systems) with GPUs for acceleration. As training jobs are resource and time consuming, enterprises would dedicate a server in their data centers for training and validating models. Once trained, IBM PowerAI Vision provides abilities to deploy the models on remote end points with atleast one accelerator (GPU or FPGA) or even with no accelerator (CPU only). In addition, the models trained on IBM PowerAI Vision can be deployed on x86 and Power systems. IBM PowerAI Vision supports deployments on a kubernettes cluster (limited IBM Cloud Private) where accelerators across the cluster are engaged for training and deployment of models.

Optimize and deploy models onto servers with FPGA

Models trained on larger resource units like GPUs are not compatible to be deployed onto FPGAs. Traditionally FPGAs consume less power, less expensive, easily programmable for a mission and can achieve best latencies (case by case basis). FPGAs are used for inference. IBM PowerAI Vision has eased the journey of training models on GPUs but deploying them onto FPGA. A pre-processing job that is required to optimize models (parse, prune, compress, precision) and flashing the resutling IP Code now has become a click away. IBM PowerAI Vision streamlines the required processes from FPGAs (limited to Xilink Alveo U200).


IBM PowerAI Vision intends to revolutionize the way models are trained, deployed and managed. In addition to lowering the barriers on technicalities, the core intent is to save time for business users and data scientists. We welcome you to take IBM PowerAI Vision for a spin on our cloud. Please feel free to review our user experience for training models for classification and object detection. You can find more indepth details of our product on the IBM marketplace. For developers, we got code samples to build the next revolutionizing solution. Early adopters can download the SW free for 30 days. Looking forward to hearing your valuable feedback.

Join The Discussion

Your email address will not be published. Required fields are marked *