Neural networks talk

Tech meeting NN talk

Transcript

1. NEURAL NETWORKS IN PHP a.k.a how to make skynet in four easy steps
2. WHY NEURAL NETWORKS? How we normally program: step by step algorithm.
3. WHY NEURAL NETWORKS? What if it’s hard to write a step-by-step recipe to solve our problem? Sometimes the underlying pattern is too complicated!
4. WHY NEURAL NETWORKS? Solution: MACHINE LEARNING Machine learning systems can learn from and make predictions based on data.
5. WHY NEURAL NETWORKS? Instead of an algorithm, we use data to make a ‘magic box’ that can solve our problem.
6. WHY NEURAL NETWORKS? Neural networks can help us solve problems for which it is hard to write a step-by-step algorithm. Examples: fraud detection, facial recognition, stock market prediction All you need is data!
7. HOW DO THEY WORK? Neural networks emulate the structure of the brain. Input data is transformed by layers of ‘neurons’ to create output data
8. HOW DO THEY WORK? Image by Colin M.L. Burnett.
9. HOW DO THEY WORK? The value of a neuron is the sum of all connected neurons, multiplied by a weight value. By changing the weight values of our network, we can make it do anything. The right weight values are determined by training.
10. HOW TO TRAIN YOUR NEURAL NETWORK 1. Collect a lot of examples. Example: PHP files grouped by ‘infected' and ‘clean’. 2. Write a function that converts your examples to something the Neural Network understands: a list of numbers (vector) per example. This is your dataset.
11. HOW TO TRAIN YOUR NEURAL NETWORK 3. The vector goes into the input neurons, a result is calculated in the output neurons. 4. Find the weight values that result in the smallest difference between calculated and actual result on your dataset. Example: the most PHP files classified correctly.
12. USING YOUR NEURAL NETWORK 1. Use the function you made earlier to convert new data to a vector. 2. Fill the input neurons with this data. 3. After calculating, a value appears in your output neurons. This is your result! 
 Example: Chance that a PHP file is infected.
13. NEURAL NETWORK CHECKLIST 1. A function from any data to a vector. 2. A meaningful result that can be calculated by your network. 3. As many examples as possible. 4. A training algorithm.
14. TESTING YOUR NEURAL NETWORK 1. Divide your data in two parts: training and test. 2. Use the first part (80%) to train your network. 3. Calculate results using the rest as input (20%). 4. Use success rate on test set to determine how good your network is.
15. IN PHP: FANN FANN is a PHP extension that does all of the heavy lifting for you A training algorithm is included (cascade2) It’s very fast, probably faster than something you make yourself.
16. EXAMPLE: XOR Create a simple dataset with examples of the ‘XOR' function.
17. EXAMPLE: XOR
18. EXAMPLE: XOR
19. DISADVANTAGES Not always right: no 100% success rate. Success rate depends heavily on your data. Hard to tell what’s going on exactly.
20. LEARN MORE PHP architect has a great article: https:// www.phparch.com/magazine/2015-2/october/ FANN: http://leenissen.dk/fann/wp/