Synthetic Gradients with Tensorflow
I stumbled upon Max Jaderberg's Synthetic Gradients paper while thinking about different forms of communication between neural modules. It's a simple idea: rather than compute gradients through backpropagation, we can train a model to predict what those gradients will be, and use our prediction to update our weights. I wanted to try using this in my own work and didn't find a Tensorflow implementation to my liking, so here is mine. I also take this opportunity to (attempt to) answer one of the questions I had while reading the paper: why not use synthetic loss instead of synthetic gradients?
Deconstruction with Discrete Embeddings
In my post Beyond Binary, I showed how easy it is to create trainable "one-hot" neurons with the straight-through estimator. My motivation for this is made clear in this post, in which I demonstrate the potential of discrete embeddings. In short, discrete embeddings allow for explicit deconstruction of inherently fuzzy data, which allows us to apply explicit reasoning and algorithms over the data, and communicate fuzzy ideas with concrete symbols. Using discrete embeddings, we can (1) create a language model over the embeddings, which immediately gives us access to RNN-based generation of internal embeddings (and sequences thereof), and (2) index sub-parts of the embeddings, instead of entire embedding vectors, which gives us (i.e., our agents) access to search techniques that go beyond cosine similarity, such as phrase search and search using lightweight structure.
Beyond Binary: Ternary and One-hot Neurons
While playing with some applications of binary neurons, I found myself wanting to use explicit activations that go beyond a simple yes/no decision. For example, we might want our neural network to make a choice between several categories (in the form of a one-hot vector) or we might want it to make a choice between ordered categories (e.g., a scale of 1 to 10). It's rather easy to extend the straight-through estimator to work well on both of these cases, and I thought I would share my work in this post. I share code for implementing ternary and one-hot neurons in Tensorflow, and show that they can learn to solve MNIST.
Non-Zero Initial States for Recurrent Neural Networks
The default approach to initializing the state of an RNN is to use a zero state. This often works well, particularly for sequence-to-sequence tasks like language modeling where the proportion of outputs that are significantly impacted by the initial state is small. In some cases, however, it makes sense to (1) train the initial state as a model parameter, (2) use a noisy initial state, or (3) both. This post examines the rationale behind trained and noisy intial states briefly, and presents drop-in Tensorflow implementations.
Recurrent Neural Networks in Tensorflow III - Variable Length Sequences
This is the third in a series of posts about recurrent neural networks in Tensorflow. In this post, we'll use Tensorflow to construct an RNN that operates on input sequences of variable lengths.
Binary Stochastic Neurons in Tensorflow
In this post, I introduce and discuss binary stochastic neurons, implement trainable binary stochastic neurons in Tensorflow, and conduct several simple experiments on the MNIST dataset to get a feel for their behavior. Binary stochastic neurons offer two advantages over real-valued neurons: they can act as a regularizer and they enable conditional computation by enabling a network to make yes/no decisions. Conditional computation opens the door to new and exciting neural network architectures, such as the choice of experts architecture and heirarchical multiscale neural networks.
Preliminary Note on the Complexity of a Neural Network
This post is a preliminary note on the "complexity" of neural networks. It's a topic that has not gotten much attention in the literature, yet is of central importance to the general understanding of neural networks. In this post I discuss complexity and generalization in broad terms, and make the argument that network structure (including parameter counts), the training methodology, and the regularizers used, though each different in concept, all contribute to this notion of neural network "complexity".
Written Memories: Understanding, Deriving and Extending the LSTM
When I was first introduced to Long Short-Term Memory networks (LSTMs), it was hard to look past their complexity. I didn't understand why they were designed the way they were designed, just that they worked. It turns out that LSTMs can be understood, and that, despite their superficial complexity, LSTMs are actually based on a couple incredibly simple, even beautiful, insights into neural networks. This post is what I wish I had when first learning about recurrent neural networks (RNNs).
Recurrent Neural Networks in Tensorflow II
This is the second in a series of posts about recurrent neural networks in Tensorflow. In this post, we will build upon our vanilla RNN by learning how to use Tensorflow's scan and dynamic_rnn models, upgrading the RNN cell and stacking multiple RNNs, and adding dropout and layer normalization. We will then use our upgraded RNN to generate some text, character by character.
Styles of Truncated Backpropagation
In my post on Recurrent Neural Networks in Tensorflow, I observed that Tensorflow's approach to truncated backpropagation (feeding in truncated subsequences of length n) is qualitatively different than "backpropagating errors a maximum of n steps". In this post, I explore the differences, and ask whether one approach is better than the other.