Brainome, a new player in the machine learning space, today launches Daimensions, a product that the company says helps customers take a “precautionary measure” approach to machine learning (ML) model development. The product, which is aimed at data researchers, helps to optimize around training data analysis, data volume management and their downstream effect on training time, model size and performance.
ZDNet spoke with Brainome’s co-founders, Bertrand Irissou (CEO) and Gerald Friedland (CTO). The two provided a careful, thoughtful, and thorough explanation of how the company’s approach to ML differs from others.
From trial and error to measurement-before-build
Brainome’s approach to ML is that much of the common model experimentation process can be optimized. Trial-and-error can be largely avoided by specifying the qualities of the model and then building it instead of following the standard experimentation method to build more candidate models and then seeing which one works best. The company calls this a measure-before-building approach and analogizes it with the way bridges are built. Specifically, say Irissou and Friedland, engineers would never take an approach to building 100 bridges and then choosing the best one. Instead, they measure and develop a spec for the bridge and only then design and build it.
Brainome defines what are essentially key performance indicators (KPIs) for a model as a way to arrive at such a spec in the ML world. The KPIs are generalization, capacity development, risk of overfit and memory equivalent capacity. Together, these KPIs can profile the complexity of the training data and the number of parameters needed to model its patterns. Fewer parameters make it possible to base the model more on rules than on stored facts, which avoids overfitting data; it also means that less data is needed to generate an accurate model, which can lead to faster training times.
Another part of Daimension’s credo is that maximizing a model’s mathematical accuracy – and bearing the cost of all the computational resources required – can easily cross a point of declining returns. Instead, model features should be ranked by importance before training. This prioritization helps simplify training and reduce model size. Meanwhile, understanding the complexity of the data and optimizing the size of the training dataset helps to shorten the training. As a result, Brainome says that there are reductions in training times and model size (each in the order of magnitude) when building models on standard OpenML datasets.
Compact models, compact company
Daimensions models are produced using Brainome’s Table Compiler (BTC) technology. The output is a single executable Python script that includes the data preparation code and the model function itself. Because the models are sent as code, they can be integrated into standard CI / CD (continuous integration / continuous delivery) pipelines. The company says this also allows customers’ DevOps practices and infrastructure to act as the implementation component of MLOps. It also says that the small models can avoid the need and cost of GPUs (graphics processing units) and can even be distributed to small edge devices or as cloud micro-services.
Brainome is a one-year-old company with 11 employees, which has so far received approx. $ 1.5 million In angel funding. Although small, it means that its industry horizontal approach has been able to run pilot projects with organizations in the fields of HealthTech, FinTech, AdTech and genomics research. Hopefully, the company will have a real impact on moving computer science work away from the guesswork-laden processes that have defined it to date. Given the efficiency and actual intelligence that machine learning models are designed to create, it is good to see both attributes are also used directly to create these models.