From c324c0bff4b34042af5843bf95004a463de8aa82 Mon Sep 17 00:00:00 2001 From: YaBoiSkinnyP Date: Mon, 18 Sep 2023 15:32:47 -0400 Subject: [PATCH] updated supervised --- README.md | 2 +- word2vec_supervised_guesser.ipynb | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index 2c8a1f5..34b9f9f 100644 --- a/README.md +++ b/README.md @@ -21,5 +21,5 @@ Demonstrate original approach to interpretable Decrypto Guessing agent Use word2vec embeddings to create heuristics that make our unsupervised guesser prototype competent # word2vec supervised guessing -Explore more traditional approach to Guesser using neural nets. May lead to proof that game-ification may inspire attention-based/Transformer architecture. +Explore more traditional approach to Guesser using neural nets. May lead to proof that game-ification may inspire attention-based architecture. diff --git a/word2vec_supervised_guesser.ipynb b/word2vec_supervised_guesser.ipynb index f559823..9d53e5b 100644 --- a/word2vec_supervised_guesser.ipynb +++ b/word2vec_supervised_guesser.ipynb @@ -233,7 +233,7 @@ "\n", "```\n", "\n", - "This encodes the input as keyword-clue pairs that the classifier may have an easier time learning associations for. The matrix is also relatively small (although we have many channels for embedding features), so it may make sense to opt for a more fully-connected architecture." + "This encodes the input as keyword-clue pairs that the classifier may have an easier time learning associations for, because it implicitly establishes the keywords and clues as orthogonal. If we can build features for each keyword-clue pair, then we can take each row of the matrix above and similarly set them as orthogonal axes in a larger tensor, just like we took the keywords and clues as orthogonal axes to make the matrix. This would allow us to build features representing 2 keyword-clue pairings, and if we repeated this operation one more time we would have features representing 3 keyword-clue pairings, which is exactly what our model should be learning to classify." ] } ],