Data. What goes in comes out.

Data

Data. What goes in comes out.

The loop from last time learns whatever is in the crates. Tonight we change what goes in, train the same model the same way, and see what comes out.

What is in the crates

The guides' sentences come in three kinds of episode. In copy episodes, a thing found earlier has to come back later. In who episodes, a question has to be answered with the guide who did the deed. Plain episodes need no looking back at all.

In the usual mix, 80 percent of episodes need a look back. Trained on that, the model gets look-back words right 99 percent of the time.

How much is enough

Now cut the look-back episodes down, and keep everything else the same.

Cut the look-back episodes from most of the crates to a twentieth. How much of the skill survives? (the film pauses here)

At 40 percent, it still gets 99 percent right. At 20 percent, 56. At 5 percent, 19.

And with none at all, the model is right 3 percent of the time, less than the untrained model's 7. Its loss on unseen rows is 8.25, worse than before it was trained. It learned to be sure, and wrong.

A skill in one frame only

Take out only the copy episodes, and keep the who episodes, which also need a look back.

On who questions, the model is still right 100 percent of the time. On copy episodes, a kind it never saw in training, 5 percent. The looking back did not carry over to a kind of sentence it had never seen.

Wrong labels

Now swap 10 percent of the training answers for random words.

Ordinary next words barely notice: still right 70 percent of the time. But the look-back skill comes slower: after the same training it gets look-back words right only 76 percent of the time. Copy episodes stay at 100 percent; the who questions take the loss, at 70. At 30 percent noise, look-back words drop to 33.

The same rows, over and over

Last, keep the amount of text the same, but draw it from a small pile of rows, used again and again.

Same amount of text, but only 64 different rows. Better, worse, or the same? (the film pauses here)

With 64 rows, each seen 500 times, the model learns the pile by heart and fails on new rows: a held-out loss of 3.02. Its held-out loss was lowest at the first check, step 100, and only got worse as it learned the pile.

With 1,024 rows: 0.71. With fresh rows every step: 0.69.

Never seen it done

One more crate. Train a model on sentences where Rerun never does the deed in a who episode. Then ask it who did something, when the answer is Rerun.

It names the right guide 0 percent of the time. When the answer is Tallow or Pylon, 100 percent. It has learned its crates faithfully, and its crates never showed Rerun doing it.

Goodnight

A model knows the world through its crates. Whatever is rare in them is weak in the model, and whatever is missing from them is missing from it.

Next in the lighthouse: what happens when the model and the crates get bigger. Goodnight.