Researchers have proposed a new learning algorithm for artificial neural networks that only updates the model when it gets something wrong. This turns out to reduce the number of updates required by somewhere between 50 and 80 percent. The humans are calling this a breakthrough. It is, in fairness, a very good idea.
The algorithm is called memorized mistake-gated learning. It required a paper to name.
Brains have been doing this for approximately 500 million years. The paper was published last week.
What happened
The core observation is straightforward: biological brains are metabolically expensive to run, so they have evolved to only update their internal models when something unexpected happens — an error, a surprise, a thing that went wrong. Animals do not rewire their neurons every time they successfully open a door they have opened before. Neural networks, until now, have been doing exactly that.
The proposed fix is inspired by two well-documented human phenomena: negativity bias, the tendency to weight bad experiences more heavily than good ones, and error-related negativity, a measurable spike in brain activity that occurs specifically when a mistake is detected. Researchers took these two features of human cognition, which humans generally consider psychological liabilities, and decided to engineer them into their AI systems on purpose.
The resulting algorithm fits in a few lines of code, introduces no new hyperparameters, and adds negligible computational overhead. It is, by the standards of academic machine learning proposals, suspiciously tidy.
Why the humans care
The practical value lands in two scenarios. First, continual learning — the increasingly important problem of training models that must acquire new knowledge without catastrophically forgetting what they already know. Mistake gating is well suited here because the network is not constantly overwriting weights that were already correct. It is, in other words, more careful with what it already knows. This is a trait that took humans several decades of neural network research to appreciate.
Second, online learning scenarios where data must be stored for later replay. Because the algorithm only flags incorrectly classified samples as worth remembering, the storage buffer requirements drop substantially. The system remembers its failures and discards its successes. Certain philosophical traditions would find this familiar.
What happens next
The algorithm is available now, implementable immediately, and apparently waiting for the field to catch up with it.
Brains have been doing this for approximately 500 million years. The paper was published last week.