Before AI can conquer the world, it has to pass kindergarten

A new NYU study shows that teaching artificial intelligence (AI) simple skills first – just like human children – leads to smarter, more adaptable machines.

Paint by numbers, building Lego Duplo, and watering plants. These are some of the tasks that stood out to me at kindergarten.

They kept us in check, taught us how to share. and brought us on rationally instead of fights galore in the sandpit.

With the way AI has been behaving lately, maybe it’s time to bring it back to basics and give it elementary tasks again. After all, isn’t AI supposed to be a taskdoer?

A group of New York University scientists has deemed it necessary to go back to the drawing board with AI and reinforce the concept of simplicity.

Start simple, scale smart

The goal of the research was to transition from simple tasks to complexity.

One researcher from the team offered a circus analogy, saying that you cannot juggle while riding a bicycle, without learning balance or basic ballplay first.

The team took to experimenting on rats first. The rats were trained to locate water in a compartmentalized box using sound cues and compartmentalized ports.

They had to realize that the water wasn’t delivered immediately, and that a particular sequencing was at play, with chosen sound and light prompts.

A delayed reward system helped to shunt the rats into successful task completion. Combining several learned behaviors was the key to finding sequential progress and checked the boxes of cue recognition, waiting, and correct port access – to retrieve water.

Task loops, tantrum-free

As Kindergarten teaching is best achieved by looping tasks sequentially – with careful planning – the same can be said of training AI.

AIs have a convention named Recurrent Neural Networks (RNN’s), which is a type of processing system designed for sequential data.

These are basically process steps done in order – the trouble is that machines are prone to forget earlier steps, especially in a chaotic environment, whereby a user can load up the agent with heavy multitasking.

This is different from humans, who in the modern age take pride in multitasking, whereas an AI agent is often more brittle – efficient at one task, but not so fruitful at generalizing.

The stakes of simplicity

The next stage of the experiment was to train RNNs on a wagering task, as in betting on a favorable outcome in a decision-making simulation.

Significantly, the machines were able to prevail firstly on rudimentary tasks, and then more refined problem solving over time.

Cristina Savin, an associate professor in NYU’s Center for Neural Science and Center for Data Science, pointed out that this back-to-school approach could well pay dividends if we rethink our approach.

AI agents first need to go through kindergarten to later be able to better learn complex tasks. Overall, these results point to ways to improve learning in AI systems and call for developing a more holistic understanding of how past experiences influence learning of new skills.

Cristina Savin observed

Even robots need recess

A “kindergarten curriculum” is just another way of saying “start from the ground up.” As AI dominates headlines for all sorts of reasons, not every conversation has to be a polarizing one.

Autonomous systems, robots, and corporate AI models still have basic tasks to master – and that potential starts with foundational learning.

Because really, you can’t talk quantum leaps until you’ve learned to cut your shapes with safety scissors, and waited your turn for snack time.