openboundlabs.com

Goal

Simplest embedding problem using Tensorflow Inspiried from this: http://matpalm.com/blog/2015/03/28/theano_word_embeddings/ https://colah.github.io/posts/2015-01-Visualizing-Representations/ https://colah.github.io/posts/2014-07-NLP-RNNs-Representations/ I first worked with embedding vectors in machine learning / neural networks many years ago where we called them context vectors and used them to do things like automatic text summarization, named entity recognition, and semantic machine translation. Embedding is powerful: word2vec, glove, neural image captioning, ... Hinton started calling them thought vectors. The basic idea is to embed objects that might not have much direct relation to each other into a space where higher semantic meaning might be more easily seen using simple operations like distance. An example is English words. Looking in the dictionary the words "happen" and "happy" occur right next to each other. However their meanings are quite different; alphabetic distance isn't too useful when dealing with word meanings. Wouldn't it be nice to embed those words in a space where words with similar meanings are close together (viz: synonyms happen==occur / happy==cheerful).... ================================

The Setup

- Predict whether two objects (o and p) are "compatible" / should occur together or not. - Embed N arbitrary objects identified by integers into 2d (x,y) (so I can plot them) - Try learning a simple linear ordering "compatibility" function: - Make adjacent pairs compatible Prob(o,p)=1.0 : (o,p) = (0,1), (1,2), (4,5), ... - Make all others incompatible Prob(o,p)=0.0 : (o,p) = (0,2), (3,5), ... - Assume pairs are ordered so (o,p) -> p>o - Try a very simple learning network: Input: (o, p) (object integers) Embed: embedding(o) = 2d Computation: f(o,p) = single linear layer. input = [embedding(x), embedding(y)] output= logit for [1-P, P] where P is probability o,p are compatible Output: softmax of network output The key here is that when learning, gradients can be pushed back to the embedding vectors so they can be updated. It is _automatic_ in frameworks like Tensorflow. Source Code: https://github.com/michaelbrownid/simplest-embedding or simply simplest-embedding.python.txt The relevant parts are quite simple: # input batchSize by two integer object numbers inputData = tf.placeholder(tf.int32, [None,2]) # target is batchSize probability of compatible: [1-Prob(compat), Prob(compat)] target = tf.placeholder(tf.float32, [None,2]) # this is the matrix of embedding vectors. rows=objects. cols=embedding. random [-1,1] embedding = tf.Variable(tf.random_uniform([numObjects, embeddingSize], -1.0, 1.0)) # pull out the embedding vectors from integer ids emap = tf.nn.embedding_lookup(embedding, inputData) # reshape to feed pair of embedding points as single vector per datapoint inputs = tf.reshape(emap , [-1, 2*embeddingSize]) # Set model to linear function to [1-p,p] 2d prediction vector softmaxW = tf.Variable(tf.random_normal([2*embeddingSize,2],stddev=0.1), name="softmaxW") softmaxB = tf.Variable(tf.random_normal([1,2], stddev=0.1), name="softmaxB") logits = tf.matmul(inputs, softmaxW ) + softmaxB probs = tf.nn.softmax(logits) # cross entropy loss = \sum( -log(P prediction) of correct label)/N. Jensen's E(logX) < = logE(X). cross = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits( logits, target)) # optimize optimizer = tf.train.AdamOptimizer(learningRate).minimize(cross) Note this is interesting because we will estimate both the embedding and the linear discriminant function on the embedding simultaneously. TODO: eliminate the linear function and use dot product with no learned parameters. ================================

RESULT

Animate the embeddings and the decision boundries for the first 4 points (colors=black, red, green, blue) every 3 epochs of optimization for 1000 epochs total.

What experiment do you want to see?

Coverged1
NotQuite
Converged2
current / 333
frame step
frame delay Smaller for faster animation!

Animate!


================================ The network can mostly learn the simple linear ordering! (A couple of errors when not converged fully) Here is a plot of the cross entropy objective by epoch and the bound relation mean cross entropy to mean accuracy: Jensen's Inequality relates the mean accuracy to mean cross-entropy. Here $P$ is the probability assigned to the correct label and the brackets are the expectation operator: $$cross = < -\log P>$$ $$ -cross = (< \log P>) \leq (\log < P>)$$ $$ (< P>) \geq (\exp(-cross)) $$ This isn't too exciting but it is interesting to see the nice linear chain that is learned from a number of pairwise examples. It reminds me of the Hammersley-Clifford theorem where a joint density can be factorized over cliques (Hammersley-Clifford theorem) It seems the ordering happens fairly quickly then the decision boundry tightens up. It is simple but interesting. Next it might be interesting to take ZIP codes and embed them. First simply try to "learn" ZIP->(latitude,longitude). Then try to pull out additional covariates like median income, total population, ..., from the embedded vectors. ================================