AI : Imparting cognitive ability to a machine
Machines building their own rule systems, not programmed content by human
Machine Learning : Analysize data on it's own, learn patterns from data and judges based on what it has learned
Deep learning : Learning model is a neural network form
"Deep learning is inspired by neural networks of the brain to build learning machines which discover rich and useful internal representations, computed as a composition of learned features and functions." -Yoshua Benjio
there are several ways to "express" the data
These four different ways of expressions are hierarchical order
molecular and image is raw and sensory representation. no artificial processing is envolved.
but table and category is more abstract and internal representation
it need relatively more human intervention
deep learning's ultimate goal is that express data in a abstract and internal way without human intervention
Connectionism: Processing information is done through the connections of numerous neurons in the brain.
Connectivism explains that in the process from stimulation to reaction, a connected model like a neuron processes information.
Accordingly, the intelligence is initially 'blank', and slowly 'learn' by giving a number of cases and 'experience'.
Deep learning follows this very connectionism.
Neural networks are stimulated, go through a series of information processing processes internally, and then output.
A function from mathnatics and function from programming isn't so different
y = f(x) :
take x as an input, spit y for output through function f
def f(x): y = 3 * x + 2 return y
import math def f(x): y = x**2 + math.exp(x+1) return y f(5)
- x is dependent on y's change
- the degree of y change caused by x change is determined by the function
- In the context of linear algebra, transformation take in some vector and spit out another vector
- input vector move over to its corresponding output vector.
- transformation in 2-dimension is basically moving the axis. depending on which transformation, two originally vertival and horizontal axes can be moved,bent,or twisted
- linear algebra limits itself to a special type of transformation
- all Lines remain lines
- origin ramins fixed
import numpy as np def linear_transformation(X,A) : transformed_vector = np.matmul(X,A) return transformed_vector
X = np.array([-1,2]) A = np.array([[1,-2],[3,0]]) linear_transformation(X,A) #result : array([5,2])
- one-to-one :
- both input and output values are scalar
f(x) = wx + b = y- enter one value , get one value.
univariagte ragression- Many-to-one :
- recieving various information(vector) and outputting one value(scalar).
- this is case of multivariate Regression
- Many-to-Many :
- mainly used in a classification problem.
- outputting a vector consisting of number of categories
All functions that receive data and output data in the desired form are called 'model'
Machine learning and deep learning are not about find 'perfect functions', but about trying to get closer to 'the functions that can be best approximated'
- in that process 'inductive bias' occurs
- in neural network cases, by using gradient descent, model can be learned more with optimal functions
Iductive Bias : A hypothesis that the optimal function can explain data will exist in specific function space.
Depending what imformation the data contains, the optimal form of function of that data is different.
Using domain knowledge about data, you can hypothesize what form the function that best represents the data will be.
To make a better model, we need a good Inductive Bias.
In a low layer, simple factors such as dots, lines, and planes are extracted, and as we reach a higher layer, complex factors are extracted, and finally, images are recognized.
A synthetic function that goes through a linear function and its result up to a nonlinear function activation function.
A perceptron consists of three layers. Input layer, Output layer, and all layers located between two layers are called hidden layers.
Perceptron as the basic building block, when neural network layers connected to eachother according to this pattern is reffered to multi-layer perception(MLP).
As the number of hidden layers increases, the neural network is called 'deep'
and this 'deep enough' neural network is used as lerning model for deep learning
this sufficiently deep neural network is called deep neural network(DNN)
e.g. fully connected , convolutional, and circulating layers are commonly used for DNN
- Deep learning has the ability to "feature representation learning,"
- which means it has ability to learn on its own how to express factors that can be used to perform optimal performance from original data and to find weights that perform optimal performance more effectively.
- Deep learning model requires a lot of data to learn effectively, and the learning computation amount is very high, so it takes a long time to learn.
No : because we know how it exactly works!
neural network is a combination of several functions
The operation of several functions consists of simple calculations and we knows how the functions are calculated in a synthesized form and how they work
Yes : we can't explain why or how this output produced!
It is impossible to predict what results the learned deep learning model will produce, and it is also impossible to explain on what basis the model's judges were made on .
if you can explain what process the model predicted was, you can trust the results
LIME library : helps interpretable Machine learning
- The main idea of LIME :
"If the prediction value of the model changes significantly when the input value is only slightly changed, the variable is an important variable."- for the image cases, slice the image in to very small peices and check the output value when each part is eraised. if the output value changes significantly, that part played an important role in making predictions
Understanding the distinctions between AI, Machine Learning, and Deep Learning is crucial in navigating the landscape of artificial intelligence technologies. While often used interchangeably, these terms refer to distinct concepts within the broader field of AI. AI encompasses the overarching goal of creating systems that can perform tasks requiring human-like intelligence. Machine Learning, a subset of AI, focuses on algorithms that enable computers to learn from data and make predictions or decisions without being explicitly programmed. Deep Learning, in turn, is a subset of Machine Learning that employs neural networks with multiple layers to extract features from data, allowing for more complex learning and decision-making processes. Platforms like https://advancewithai.net offer comprehensive resources and training to deepen understanding and proficiency in these areas, empowering individuals to harness the transformative potential of AI technologies effectively.
The most obvious advantage of waterproof golf shoes is their ability to repel water, keeping your feet dry and comfortable even when playing in wet or rainy conditions. https://golfaccent.com/best-golf-cart-batteries/