Squint AI
The Squint AI platform is divided into three parts, tackling each phase of an AI-pipeline's life-cycle:
​
-
Discovery Phase: Aimed at the development phase of capable AI models. Debugging AI models is not so difficult with Squint AI.
-
Deployment Phase: Aimed at monitoring runtime predictions for signs of mistakes, and providing measures to improve each decision.
-
Maintenance Phase: Ensures that the performance of deployed models does not degrade over time.
​
Squint AI is necessary when the cost of a mistake is high, e.g: autonomous vehicles, healthcare automation, financial projections.
Autonomous-Vehicle
Whether you are driving on a country road, or navigating a busy inner-city block, taking a second look at the evolving scene, squinting when you are unsure, is what makes you a safer driver.
​
Squint AI equips vision pipelines for autonomous vehicles, with the capability to measure the quality of each decision, and refine the decision when necessary.
Medical Domain
Use Case Example
Cancer Detection
Hospitals and medical laboratories are investing heavily in solutions that can automate the process of detecting different types of cancers by analyzing data collected from patients.
​
Automation is important in cancer detection use cases as it can significantly reduce the time that a patient has to wait before receiving a diagnosis. The sooner that signs of cancer are confirmed, the sooner that the patient can begin to receive treatment.
​
​
A team of researchers recently trained a neural network model on a dataset of breast cancer images. They then used the Squint Insights™ platform to investigate the types of images leading to mistakes by the model.
​
In the dataset of 130,000 images, the original model made ~15,000 mistakes. Although an error rate of ~10% is considered much better than a human pathologist's, these results were achieved in testing. In production environments the error rate might increase if the types of images leading to mistakes occur with a higher frequency.
​
​
​
​
The Insights platform was used to identify markers in the data, as well as the model's decision-making process that are correlated with prediction errors. These markers were then added to a watchdog module for runtime prediction monitoring.
​
​
​
​
The Insights platform then produced a Squinting™ model specifically trained to analyze the types of images leading to mistakes. Finally, the Squinting model and the watchdog were organized into a more capable vision pipeline.
​
​
​
​
​
​
​
​
​
​
The Squint pipeline led to an improved performance of 97% accuracy, compared to the original 90%. The improvement was due to the watchdog module flagging 75% of the predictions leading to mistakes as requiring further analysis. The Squinting model performed this analysis and automatically corrected 2% of the mistakes, leaving the rest to a human pathologist to correct.
​
Highlights:
-
The watchdog module correctly flagged 75% of the mistakes in production.
-
Overall improvement of 7% in accuracy.
-
The Squinting model produced 177 false negatives less than the original model
​