I took a Machine Learning course this past winter and this article in particular would have been really helpful since I struggled most with this concept in particular (and gradient descent in general). While most resources show you the mechanics of neural networks, none I found were very good at explaining (to me) the purpose and meaning behind them. Sure, I could follow along and eventually figure out how to write my own neural network, and I did, but I honestly never completely understood what was going on. The problem with most ML texts/resources for people like me without a strong math background is that a lot of high-level math is presented without an explanation of what mathematical concepts are being used. I admit that the onus is on me, the math dummy, to go out and learn the concepts involved, but it's difficult to look at a confusing algorithm chock full of unfamiliar concepts and know where to start. This article explains things nicely and I hope to see more like this in ML.
If you want a nice, visual, explanation of what a neural network is actually doing, I found this blog post to be particularly helpful. [1] Although there is some math, the rest of the post gives a good intuitive understanding of what's going on without it.
This is a very welcome read but it hypes neural networks a bit. I've been working with in JavaScript using IndexedDB and while researching I was disappointed to find that some smart people seem to think they are much more limited than made out to be here and elsewhere https://www.youtube.com/watch?v=AyzOUbkUf3M#t=242
To summarize, people generally abandoned backpropigation trained neural networks for Support Vector Machines because neural nets require labeled and limited datasets, and work slowly and especially so when dealing with multiple layers which is sort of the whole point.
In my work in JavaScript, I was able to pull off only a single layer perceptron and it is neat but limited in what it can model.
Frankly, the reason there's hype around neural nets again is because of the newer ways they can be trained to augment backprop. Neural nets have made a lot of progress in representation learning in recent years.
A major problem that I'm hoping to fix is having the industry catch up to what academia and some of the bigger companies are doing with nets now.
Neural nets are from a silver bullet, and shouldn't be used where feature introspection is a huge requirement (this is why decision tree/random forest is popular), but they are far from being what they were in the 90s.
Note that I have a commercial interest in this so there's going to be inherent bias in my opinions.
To be fair, javascript isn't a scientific computing language. To do most neat training with neural nets, you're going to want to either scale them out, add more layers, and/or use GPUs. That being said, a neat toy example in javascript is convnetjs[1].
It's true that the beginning of the talk is about how people abandoned backprop trained neural networks because they often underperformed SVMs, but the rest of the talk is about deep learning, which introduced a new generation of neural nets (since 2006) that are the current state of the art for a lot of problems.
In fact, deep neural networks are trained in an unsupervised manner at first, but then back propagation is used to "fine tune" and improve the results. Because they require unlabeled data sets, and can perform so well, research into neural networks has experienced a recent resurgence.
By the way, any talk by Geoff Hinton is fantastic. If you are interested in neural networks and their capabilities, and you haven't already seen it, his Coursera course [1] builds from a simple linear perceptron to the current deep learning methods.
I want to add to this, that there's a lot of work that doesn't require pretraining. I recently implemented the more advanced hessian free optimization methods that don't require pretraining in my deep learning framework. The results are amazing. I'm hoping to demonstrate a lot of the tradeoffs of the different methods in a more comprehensive manner here shortly. This was an updated extension by some of hinton's students. The paper I implemented was:
Thanks for the extra resources. Any more are very welcome. Neural nets sometimes feel like quite the black box, despite their ease of implementation and apparent power.
You are linking to a talk by Geoffrey Hinton. A strong advocate for Neural Networks. A single layer perceptron is limited, but a Neural Network with hidden layer(s) is a very powerful tool, and we have now figured out how to deal with multiple hidden layers (given enough data).
The best text I've read about backpropagation is in "Neural Networks: A Systematic Introduction" by Raul Rojas [1]
He uses a nice graphical approach that is easily understandable yet formal. It's been many years since I've read it to learn for an exam at university but I remember it was an enjoyable read and I wished I had more time to spend on the book.
Not having read Rojas' book, I think the best text I've ever read on the subject is Michael Nielsens (in progress) "Neural Networks and Deep Learning." [1] Chapter 2 covers back propagation.
It has a similarly approachable yet formal style, and I have recommended it to people with no ML experience in the past who have found it very intuitive.
Optimization in the context of machine learning is a search function that searches a param space by repeated calculations of gradients (derivatives: eg, calculate the derivative and test the objective function) to arrive at some optimal function.
Learning in this case generally means the creation of models. Optimization is the process by which this happens.
A few examples:
Unsupervised learning - typically clustering and grouping things
Supervised Learning - Example based learning where you're trying to label something. This could be sentiment, spam classification, object recognition over pixels, even extending to sequence labeling.
Prediction/Regression - Predicting values based on a learned function.
Learning is more about the goal trying to be achieved. Optimization is more of numerically solving relative to an objective function (also called an error function)
Learning is the ML-equivalent of what statisticians would call fitting a model. It usually involves optimizing a criterium such as the likelihood function of the data, the expected MSE under the posterior (Bayesian) or the margin of a linear classifier (SVM).
Frankly machine learning is a little bit of a buzz term. Not because it does not have a well defined meaning, but it's a way to make what you are doing sound more interesting.
But I would say that learning is a higher level concept. So you would use an optimization algorithm to learn the solution for an SVM from data.
