Catastrophic Forgetting

Train a neural network on task A. Then train it on task B. Go back and test it on task A. It's forgotten everything. Not gradually, not partially — catastrophically. The weights that encoded task A have been overwritten by the weights needed for task B, and the information is gone. This is the catastrophic forgetting problem, and it's one of the most fundamental differences between how neural networks learn and how biological brains learn.

The problem has been known since the late 1980s, and the simplest solution is obvious: rehearsal. When you learn new things, periodically re-train on old things too. This works, but it requires keeping the old training data around, which defeats much of the point of having a network that's supposed to learn the information rather than just store it. If you need an external database of everything the network has ever seen, what is the network actually doing for you?¹

Pseudorehearsal

Anthony Robins's 1995 solution is clever enough to be worth understanding in detail. Instead of rehearsing the actual old training examples, generate pseudo-examples: random input vectors passed through the current network to produce whatever output the network currently generates for them. These input-output pairs aren't real data — they're samples from the function the network currently implements. Then, when learning new material, interleave these pseudo-examples with the new training data.¹

Why does this work? The core insight is geometric. Catastrophic forgetting happens because learning a new input-output mapping changes the network's function globally — including at points far from the new training example. Rehearsal fixes this by "pinning down" the function at the original training points. Pseudorehearsal does the same thing but pins the function at random points sampled from it. Since the random points are drawn from the function itself, they collectively approximate the function's shape, and re-training on them during new learning prevents the function from drifting too far from its current form in regions away from the new data.¹

The beautiful thing is that this is purely internal. No external memory required. The network generates its own rehearsal material from its own current state. The old data doesn't need to exist anywhere — the network has already compressed it into its weights, and pseudorehearsal samples from that compressed representation.

Sleep and Consolidation

Robins explicitly connected pseudorehearsal to sleep-based memory consolidation in biological brains. The hippocampus rapidly encodes new episodic memories during waking, and then during sleep — particularly slow-wave sleep — replays these memories to the neocortex for gradual integration into long-term storage. The replay isn't exact; it's noisy and recombined. And crucially, the neocortex is simultaneously generating its own spontaneous activity, which functions like pseudorehearsal: it prevents the new hippocampal input from catastrophically disrupting existing cortical representations.¹

This connection is more than analogical. Hopfield networks (an older type of neural network used for associative memory) face a related problem: learning new patterns can destroy stored patterns by creating spurious attractors. The unlearning solution proposed by Crick and Mitchison in 1983 — also inspired by sleep — involves running the network in a noise-driven mode and anti-training on whatever spurious patterns emerge. Robins showed that this unlearning approach and his relearning-based pseudorehearsal approach are related: both use the network's own internally generated activity to maintain stability during learning.¹

The broader point is that biological brains have solved catastrophic forgetting through architectural separation (hippocampus for fast learning, neocortex for slow consolidation) combined with a replay mechanism that bears a family resemblance to pseudorehearsal. The Sleep And Dreams article explores the experiential side of this process — the lost biphasic sleep architecture, the cognitive spectrum from waking to dreaming, and the evidence that dream replay serves exactly the consolidation function Robins describes. Evolution discovered this solution over hundreds of millions of years. Neural network researchers independently discovered a related solution from the computational constraints alone.

The Modern Landscape

In the era of large language models, catastrophic forgetting manifests differently but no less seriously. Fine-tuning a pre-trained model on a specific task can degrade its general capabilities — this is sometimes called "alignment tax" when the fine-tuning is RLHF. The LLM training pipeline is designed around this constraint: pre-training on broad data first, then careful fine-tuning that tries to add capabilities without destroying existing ones.

Parameter-efficient fine-tuning methods like LoRA (Low-Rank Adaptation) can be understood partly as a response to catastrophic forgetting. By freezing most of the network's weights and only training small adapter matrices, LoRA restricts the space of possible weight changes, preventing the global function from shifting too far from its pre-trained state. It's not pseudorehearsal — it's more like putting most of the network in a cast and only allowing a few joints to move — but it addresses the same underlying problem.

The stability-plasticity dilemma at the heart of catastrophic forgetting isn't just a technical nuisance. It points at something fundamental about what it means for a system to learn sequentially. A system that perfectly retains everything is rigid — it can't adapt. A system that freely overwrites old knowledge is amnesiac — it can't accumulate. Biological brains achieve a balance through sleep, through architectural specialisation, through the fact that the hippocampus and neocortex have different learning rates tuned to different temporal horizons. We're still searching for the artificial equivalent, and the search has become more urgent as we try to build systems that learn continuously from deployment rather than being retrained from scratch.