https://tomroth.dev/tags/neural-network/Recent content in neural network on Tom RothHugo -- gohugo.ioenFri, 07 Oct 2022 00:00:00 +0000https://tomroth.dev/weight_tying/Fri, 07 Oct 2022 00:00:00 +0000https://tomroth.dev/weight_tying/Weight-tying is where you have a language model and use the same weight matrix for the input-to-embedding layer (the input embedding) and the hidden-to-softmax layer (the output embedding). The idea is that these two matrices contain essentially the same information, each having a row per word in the vocabulary. The concept seems to be first mentioned by Press and Wolf in 2016. Below we illustrate this with a RNN. This post assumes that you are somewhat familiar already with how a RNN works.https://tomroth.dev/pytorch-cnn/Wed, 14 Apr 2021 00:00:00 +0000https://tomroth.dev/pytorch-cnn/Find the code for this blog post here: https://github.com/puzzler10/simple_pytorch_cnn We will build a classifier on CIFAR10 to predict the class of each image, using PyTorch along the way. This is basically following along with the official Pytorch tutorial except I add rough notes to explain things as I go. There were a lot of things I didn’t find straightforward, so hopefully this piece can help someone else out there. If you’re reading this, I recommend having both this article and the Pytorch tutorial open.https://tomroth.dev/invariance/Fri, 06 Mar 2020 00:00:00 +0000https://tomroth.dev/invariance/Sensitivity-based adversarial attacks Link to heading Neural networks are vulnerable to adversarial attack, and there’s a lot of talk about this. There are many definitions of an adversarial attack, but let’s go with this one: an input to a model designed to produce an incorrect output. Models can run on many different types of data (text, audio, tabular) but let’s use images as a working example. The most well-known kind of adversarial attack is to add some kind of perturbation to an image to make a classifier mislabel it.https://tomroth.dev/deepfool/Thu, 27 Feb 2020 00:00:00 +0000https://tomroth.dev/deepfool/Paper: Deepfool: a simple and accurate method to fool deep neural networks. (Moosavi-Dezfooli et al, 2016). Idea: If you want to find an adversarial example to an image, look for the closest decision boundary, and then move the image towards a linear approximation of the decision boundary by orthogonally projecting it onto the boundary. Once it crosses the boundary it will be an adversarial image. Background Link to heading Affine functions.https://tomroth.dev/exploding_vanishing_gradients/Wed, 19 Feb 2020 00:00:00 +0000https://tomroth.dev/exploding_vanishing_gradients/The following code explores how gradients explode or vanish when backpropagating through a neural network. It depends on the type of activation function you are using and how many layers you have in the network. Observations for some activation functions: sigmoid tends to lead to vanishing gradient sometimes with relu you get a lot of zeros in the gradient. I didn’t see any vanishing gradient with relu, but there was exploding gradient.https://tomroth.dev/intdim/Tue, 07 Jan 2020 00:00:00 +0000https://tomroth.dev/intdim/The paper in one sentence. The Pareto principle for neural networks: what’s the least number of parameters needed for to achieve most of the results? Structure of this article. I give an introduction to objective landscapes and talk about how to define solutions. I use these concepts to describe the subspace training method. Finally, I summarise the key results of the paper. Objective landscapes Link to heading For every combination of neural network architecture and dataset, the shape of the objective/loss function is fixed.https://tomroth.dev/dqn-nnet/Thu, 01 Nov 2018 00:00:00 +0000https://tomroth.dev/dqn-nnet/After building the basic DQN, I started trying some different neural network architectures. I found vastly different results between them. Spoiler: more complex doesn’t equal better. Almost the opposite. Here are some results for the rewards obtained in each episode. I trained these models on cartpole-v1. Each was trained for 1000 episodes. Disclaimer: I did only one run with each architectures. I kept other hyperparameters the same as the previous post, but there are likely some interaction effects unaccounted for.https://tomroth.dev/dqn-simple/Wed, 31 Oct 2018 00:00:00 +0000https://tomroth.dev/dqn-simple/Click the image above to see the source code. Last time, we established some theory for our DQN. Now it’s time to implement it. We won’t build a full DQN from the DeepMind paper, not yet. Instead we’ll build a simplified version. I call it the basic DQN. The basic DQN is the same as the full DQN, but missing a target network and reward clipping. We’ll get to that in the next post.https://tomroth.dev/dqn-basics/Tue, 23 Oct 2018 00:00:00 +0000https://tomroth.dev/dqn-basics/In February 2015, Google Brain published a paper combining deep neural networks and reinforcement learning for the first time - called DQN (deep q-network). It was a landmark moment. DQN was the first algorithm that could successfully play a wide range of Atari games. Other algorithms at the time could perform well at a single game but couldn’t generalise across games. Impressively, DQN was able to perform above human level at a range of Atari games using only the screen pixels as input.