Posted by Erika Greelish on September 4, 2020 | No Comments
Despite the incredibly complex math and analytics behind the process, training a neural net has often been compared to teaching a toddler. Some commonalities:
- A steep learning curve. Even if you have some prior experience with children, or you’ve read every childcare book out there, there’s just no substitute for first-hand experience. Until you raise little Liam day in and day out, there is no way to know that he’ll eat a scrambled egg but not a hard-boiled one. In the same way, data scientists new to a field need to master the subtle nuances of the domain knowledge. For example, you can’t train your neural net to diagnose a specific pathology if you yourself don’t clearly understand the difference between the appearance of a normal anatomic structure versus an abnormal one.
- Patience and persistence. Parents must tell toddlers the same thing again and again. And again. And again. It takes a monumental amount of patience to repeatedly inform a small person that “food is for eating, not for finger painting.” Similarly, long training cycles present an endurance challenge for data scientists. Training a neural net to act in the ways you want it to can take anywhere between several hours to several days, depending of the complexity of the data set and the task. And just like a multi-tasking parent, a data scientist must be able to focus on other work in addition to the training session.
- So. Much. Cleaning. Whether you’re wiping their noses, sweeping up crushed Cheerios they’ve tracked through the house, or scrubbing mashed peas out of their clothes, toddlers are a cleaning long haul, especially if you want to make them presentable to the outside world. Interestingly enough, so are data sets. Your outcome is only as good as your data, so data scientists spend a lot of time “cleansing” data. Data cleansing is the process of detecting and correcting or removing corrupt, incomplete, inaccurate, or irrelevant records from a data set. The acquisition, cleansing, and preparation of data needs to be done carefully and thoroughly in order to produce a high-quality result.
- Uncertainty. Is it acceptable for your toddler to watch TV if it’s educational? Is it okay if the grandparents feed him candy? Sometimes parenting decisions are black and white; other times you’re faced with shades of gray. In the same vein, blurry or unclear MRI studies (“noisy labels”) confuse the neural network and decrease the learning rate.
- Learning happens—supervised or unsupervised. Any parent who’s ever left a young child alone for even a few minutes will be surprised to learn just how much of a wall can be scribbled on in such a short time, or how many rolls of toilet paper can be unraveled. Toddlers can come to some pretty unconventional conclusions about what should be done with the available stimulus, and so can an unsupervised AI algorithm. Usually, an AI algorithm is trained to sort a data set with a particular desired output. In this supervised learning situation, the neural network optimizes its predictions based on continuous feedback from the humans in charge. Sometimes, however, we learn more by letting the AI algorithm decide for itself how to sort the data and what conclusions to draw. From inventing whimsical approaches to evolving more efficient ways of moving, to discovering new cancer cell types, unsupervised machine learning sometimes leads to creative and unexpected outcomes that can help advance our understanding of the human body.
- Evolution over time. Children learn to do many things as they grow, but linguistic development is particularly noteworthy (especially when your toddler, with perfect clarity, announces some highly embarrassing observation about a random stranger in the grocery store). At 18 months, a typical child knows between five and forty words; with regular attention and modeling, by the age of three, a typical child’s vocabulary could be as large as 1,000 words. In the world of AI, the accuracy of the output also evolves over time, either because the quantity of available input data grows, or because the label quality improves, or because new training iterations are performed, all of which refine the learning capabilities of the algorithm.