Why Do We Need Machine Learning?
We need machine learning for tasks that are too complex for humans to be coded directly, eg Tasks are so complex that it is impractical, if not impossible, for us to do all the nuances and codes for them explicitly. So instead, we provide machine learning algorithms with a large amount of data and let it explore and search for models that will determine what programmers want to achieve.
Let’s look at these two examples:
It is very difficult to write programs that solve problems such as recognizing 3D objects, from a new perspective, in new lighting conditions, in a messy landscape. Even if we have a good idea to do it, the program might be very complicated.
It is difficult to write a program to calculate the probability that credit card transactions are fraudulent. Maybe there are no simple and reliable rules.Fraud is a moving target, but the program must continue to change.
Then comes the Machine Learning Approach: instead of writing programs by hand for each specific task, we collect many examples that determine the correct output for the input given. Machine learning algorithms then take these examples and produce programs that do work. Programs produced by learning algorithms may look very different from typical handwriting programs – they may contain millions of numbers. If we do it right, this program works for new cases, as well as those we train. If data changes, the program can change also with training from new data. You should note that doing large amounts of calculations is now cheaper than paying someone to write a special assignment program.
Some examples of tasks that are best completed with machine learning include:
The pattern of recognition: objects in real scenes, facial identities or facial expressions, and/or spoken words
Recognize anomalies: unusual order of credit card transactions, unusual sensor reading patterns in nuclear power plants
Predictions: future stock prices or currency exchange rates, which films someone likes
What is an Artificial Neural Network?:
Neural Networks are a class of models in the general machine learning literature. Neural networks are a set of specific algorithms that have revolutionized machine learning. They are inspired by biological neural networks and so-called deep neural networks have proven to work well. Neural Networks themselves are estimates of general functions, which is why they can be applied to almost all machine learning problems about studying complex mapping from input to output space.
Here are three reasons you should learn nerve calculations:
To understand how the brain actually works: it’s very large and very complicated and made of dead objects when you pierce it, so we need to use computer simulations.
To understand parallel computational forces inspired by neurons and their adaptive connections: this style is very different from sequential computing.
To solve practical problems using new brain-inspired learning algorithms: learning algorithms can be very useful even if they are not how the brain actually works.
Top 3 Nervous Network Architecture You Need To Know
1 – Perceptrons:
Regarded as the first generation of neural networks, the perception is only a computational model of one neuron. The Perceptron was originally created by Frank Rosenblatt (“The perceptron: a probabilistic model for storing information and organization in the brain”). Also called an advanced feed network, the perceptron feeds information from front to back. Perceptron training usually requires back propagation, which provides a network of pairs of input and output data. Inputs are sent to neurons, processed, and produce output. Repeated errors are usually the difference between input and output data. If the network has enough hidden neurons, it can always model the relationship between input and output. Practically, their use is far more limited, but they are combined popularly with other networks to form new networks.
2 – Convolutional Neural Networks:
In 1998, Yann LeCun and his colleagues developed a very good identifier for a handwritten digit called LeNet. This uses back-propagation on feedforward net with many hidden layers, many map units replicated in each layer, output pooling from nearby replicated units, wide webs that can handle multiple characters at once, even if they overlap, and are good at training complete system, not just identifiers. It was then formalized with the name *** neural convolutional neural networks (CNN).
Convolutional neural networks are very different from most other networks. They are mainly used for image processing, but can also be used for other types of inputs, such as audio. A common use case for CNN is where you feed network images and classify data. CNN tends to start with an input “scanner”, which is not intended to describe all training data at once. For example, to insert an image measuring 100 x 100 pixels, you don’t want a layer with 10,000 points. Instead, you make the scanning input layer say, 10 x 10, and you feed the first 10 x 10 pixels of the image. After you pass that input, you give it the next 10 x 10 pixels by moving the scanner one pixel to the right.
3 – Repeated Nervous Tissue:
To understand the RNN, we need to have a brief description of sequence modelling. When applying machine learning to sequences, we often want to change the order of inputs into output sequences that live in different domains. For example, changing the sequence of sound pressure into a sequence of identity words. When there are no separate target sequences, we can get the teaching signal by trying to predict the next term in the input sequence. The target output sequence is a sequence of inputs with a one-step forward. This seems far more natural than trying to predict one pixel in an image from another pixel, or a patch of the image from the rest of the image. Predicting the next term in sequence obscures the difference between supervised and unsupervised learning. It uses methods designed for supervised learning but does not require separate teaching signals.
Specifically, the auto regressive model can predict the next term sequentially from a fixed number of previous terms using “delayed taps.” The advanced feed neural net is a general auto regressive model that uses one or more layers of hidden non-linear units. However, if we give our generative model some hidden circumstances, and if we give this hidden state internal dynamics, we get a much more interesting type of model that can keep information hidden for a long time. If the dynamics and the way to produce output from its hidden state are noisy, we will never know exactly the hidden state. The best we can do is deduce the probability distribution in the vector space of the hidden state. This conclusion can only be traced to two types of hidden state models