Why "Brute Force" Sucks for Training Models

Jeff Northrup

---

There is a saying when learning to read music: “The long way is the short way.” This implies that there is no short way–that the ONLY way is to use “brute force” to learn hard things.

But what if the long way sucks? What if there is a short way to learn something equally well?

In Machine Learning, the brute force method involves ever-increasing amounts of data, time, money and energy. The scenario sometimes looks like this: You are training the 5th model to accomplish your business objective. But the accuracy of your model plateaus…again. At that point, your choices generally point to a combination of MORE data, MORE time, MORE money and MORE energy usage. You may even see MORE layers added to your model and an INCREASED demand for new hardware ($$$$). You may see INCREASED latency and delays in delivering to your customers.

You step back and think about the other 4 models you have tinkered with. You even look out over the landscape of new models to try. Is it worth resurrecting one of the prior models? Should you venture out to a new model? Should you reassess your dataset? Do you even have the time and money to consider all these options? 

The possibilities are almost endless. And no choice guarantees the successful outcome you need.

THIS is the brute force method of training models.

Our Approach

GiiLD rejects the notion that “The long way is the short way.” We reject brute force training…because it totally sucks.

GiiLD has developed a method to analyze and prune weights. We do CAT scans and brain surgery on your models. We cut out diseased weights and promote healthy ones. Our “short way” will trim down your time to market. 

We help you fail faster in transfer learning. Our weights analysis will tell you whether you should scrap your model or if there is a way to fix it. 

We can help you stabilize your model. We can repair NaN errors and help smooth your accuracy and loss curves.

We can prune your model. We will find the weights that are preventing accuracy gains and repair them. We will remove the weights that do nothing other than increase overhead.

And we can give it amnesia. We can identify the weights associated with specific data–and remove it.

Prune Your Models with GiiLD

With GiiLD you can deliver smaller models faster using less data, less money and less energy.  

Contact us to harness the power of pruning to make your models faster, smarter, and cheaper.